Final Government Distribution                                                  Chapter 3                                                             IPCC AR6 WGI

 1   Table of Contents

 2   Executive Summary .................................................................................................................................................. 4

 3   3.1       Scope and Overview ....................................................................................................................................... 9

 4   3.2       Methods ........................................................................................................................................................ 10
 5         3.2.1      Methods Based on Regression .................................................................................................. 11
 6         3.2.2      Other Probabilistic Approaches ................................................................................................ 12

 7   3.3       Human Influence on the Atmosphere and Surface ...................................................................................... 12
 8         3.3.1      Temperature                          .............................................................................................................. 12
 9 Surface Temperature .............................................................................................................. 12
10 Upper-Air Temperature ............................................................................................................ 23

11   Cross-Chapter Box 3.1: Global Surface Warming over the Early 21st Century ................................................... 26
12         3.3.2      Precipitation, Humidity and Streamflow ................................................................................... 29
13 Atmospheric Water Vapour ...................................................................................................... 31
14 Precipitation                           .............................................................................................................. 32
15 Streamflow                              .............................................................................................................. 36

16   Cross-Chapter Box 3.2:Human Influence on Large-scale Changes in Temperature and Precipitation Extremes 37
17         3.3.3      Atmospheric Circulation ........................................................................................................... 39
18 The Hadley and Walker Circulations ........................................................................................ 39
19 Global Monsoon                          .............................................................................................................. 42
20 Extratropical Jets, Storm Tracks and Blocking .......................................................................... 44
21 Sudden Stratospheric Warming Activity ................................................................................... 47

22   3.4       Human Influence on the Cryosphere ........................................................................................................... 47
23         3.4.1      Sea Ice                              .............................................................................................................. 47
24 Arctic Sea Ice                          .............................................................................................................. 47
25 Antarctic Sea Ice                       .............................................................................................................. 49
26         3.4.2      Snow Cover                           .............................................................................................................. 51
27         3.4.3      Glaciers and Ice Sheets ............................................................................................................. 52
28 Glaciers                                .............................................................................................................. 52
29 Ice Sheets                              .............................................................................................................. 53

30   3.5       Human Influence on the Ocean .................................................................................................................... 55
31         3.5.1      Ocean Temperature                    .............................................................................................................. 55
32 Sea Surface and Zonal Mean Ocean Temperature Evaluation .................................................... 55
33 Tropical Sea Surface Temperature Evaluation........................................................................... 57
34 Ocean Heat Content Change Attribution ................................................................................... 58
     Do Not Cite, Quote or Distribute                                                    3-2                                                             Total pages: 202
     Final Government Distribution                                                   Chapter 3                                                             IPCC AR6 WGI

 1         3.5.2       Ocean Salinity                      .............................................................................................................. 60
 2 Sea Surface and Depth-profile Salinity Evaluation .................................................................... 61
 3 Salinity Change Attribution ...................................................................................................... 62
 4         3.5.3       Sea Level                           .............................................................................................................. 63
 5 Sea Level Evaluation .............................................................................................................. 63
 6 Sea Level Change Attribution ................................................................................................... 64
 7         3.5.4       Ocean Circulation                   .............................................................................................................. 65
 8 Atlantic Meridional Overturning Circulation (AMOC) .............................................................. 65
9 Southern Ocean Circulation ...................................................................................................... 67

10   3.6       Human Influence on the Biosphere .............................................................................................................. 68
11         3.6.1       Terrestrial Carbon Cycle ........................................................................................................... 68
12         3.6.2       Ocean Biogeochemical Variables.............................................................................................. 71

13   3.7       Human Influence on Modes of Climate Variability ..................................................................................... 73
14         3.7.1       North Atlantic Oscillation and Northern Annular Mode ............................................................ 73
15         3.7.2       Southern Annular Mode............................................................................................................ 77
16         3.7.3       El Niño-Southern Oscillation .................................................................................................... 79
17         3.7.4       Indian Ocean Basin and Dipole Modes ..................................................................................... 83
18         3.7.5       Atlantic Meridional and Zonal Modes ....................................................................................... 85
19         3.7.6       Pacific Decadal Variability ....................................................................................................... 86
20         3.7.7       Atlantic Multidecadal Variability .............................................................................................. 88

21   3.8       Synthesis across Earth System Components ................................................................................................ 91
22         3.8.1       Multivariate Attribution of Climate Change .............................................................................. 91
23         3.8.2       Multivariate Model Evaluation ................................................................................................. 92
24 Integrative Measures of Model Performance ............................................................................. 92
25 Process Representation in Different Classes of Models ............................................................. 97

26   Frequently Asked Questions ..................................................................................................................................100
27         FAQ 3.1: How do we Know Humans are Responsible for Climate Change? ...................................... 100
28         FAQ 3.2: What is Natural Variability and How has it Influenced Recent Climate Changes? .............. 102
29         FAQ 3.3: Are Climate Models Improving? ........................................................................................ 104

30   Acknowledgements .................................................................................................................................................106

31   References ..............................................................................................................................................................107

32   Figures ....................................................................................................................................................................153

     Do Not Cite, Quote or Distribute                                                     3-3                                                              Total pages: 202
     Final Government Distribution                                  Chapter 3                    IPCC AR6 WGI

 1   Executive Summary
 3   The evidence for human influence on recent climate change strengthened from the IPCC Second Assessment
 4   Report to the IPCC Fifth Assessment Report, and is now even stronger in this assessment. The IPCC Second
 5   Assessment Report (1995) concluded ‘the balance of evidence suggests that there is a discernible human
 6   influence on global climate’. In subsequent assessments (TAR, 2001; AR4, 2007 and AR5, 2013), the
 7   evidence for human influence on the climate system was found to have progressively strengthened. AR5
 8   concluded that human influence on the climate system is clear, evident from increasing greenhouse gas
 9   concentrations in the atmosphere, positive radiative forcing, observed warming, and physical understanding
10   of the climate system. This chapter updates the assessment of human influence on the climate system for
11   large-scale indicators of climate change, synthesizing information from paleo records, observations and
12   climate models. It also provides the primary evaluation of large-scale indicators of climate change in this
13   report, that is complemented by fitness-for-purpose evaluation in subsequent chapters.
15   Synthesis across the Climate System
17   It is unequivocal that human influence has warmed the global climate system since pre-industrial
18   times. Combining the evidence from across the climate system increases the level of confidence in the
19   attribution of observed climate change to human influence and reduces the uncertainties associated with
20   assessments based on single variables. Large-scale indicators of climate change in the atmosphere, ocean,
21   cryosphere and at the land surface show clear responses to human influence consistent with those expected
22   based on model simulations and physical understanding. {3.8.1}
24   For most large-scale indicators of climate change, the simulated recent mean climate from the latest
25   generation Coupled Model Intercomparison Project Phase 6 (CMIP6) climate models underpinning
26   this assessment has improved compared to the CMIP5 models assessed in the AR5 (high confidence).
27   High-resolution models exhibit reduced biases in some but not all aspects of surface and ocean climate
28   (medium confidence), and most Earth system models, which include biogeochemical feedbacks, perform as
29   well as their lower-complexity counterparts (medium confidence). The multi-model mean captures most
30   aspects of observed climate change well (high confidence). The multi-model mean captures the proxy-
31   reconstructed global-mean surface air temperature (GSAT) change during past high- and low-CO2 climates
32   (high confidence) and the correct sign of temperature and precipitation change in most assessed regions in
33   the mid-Holocene (medium confidence). The simulation of paleoclimates on continental scales has improved
34   compared to AR5 (medium confidence), but models often underestimate large temperature and precipitation
35   differences relative to the present day (high confidence). {3.8.2}
37   Human Influence on the Atmosphere and Surface
39   The likely range of human-induced warming in global-mean surface air temperature (GSAT) in 2010–
40   2019 relative to 1850–1900 is 0.8°C–1.3°C, encompassing the observed warming of 0.9°C–1.2°C, while
41   the change attributable to natural forcings is only −0.1°C–0.1°C. The best estimate of human-induced
42   warming is 1.07°C. Warming can now be attributed since 1850–1900, instead of since 1951 as done in the
43   AR5, thanks to a better understanding of uncertainties and because observed warming is larger. The likely
44   ranges for human-induced GSAT and global mean surface temperature (GMST) warming are equal (medium
45   confidence). Attributing observed warming to specific anthropogenic forcings remains more uncertain. Over
46   the same period, forcing from greenhouse gases 1 likely increased GSAT by 1.0°C–2.0°C, while other
47   anthropogenic forcings including aerosols likely decreased GSAT by 0.0°C–0.8°C. It is very likely that
48   human-induced greenhouse gas increases were the main driver 2 of tropospheric warming since
49   comprehensive satellite observations started in 1979, and extremely likely that human-induced stratospheric
50   ozone depletion was the main driver of cooling in the lower stratosphere between 1979 and the mid-1990s.
51   {3.3.1}

         In this chapter, ‘greenhouse gases’ refers to well-mixed greenhouse gases.
         In this chapter, ‘main driver’ means responsible for more than 50% of the change.
     Do Not Cite, Quote or Distribute                                  3-4                       Total pages: 202
     Final Government Distribution                               Chapter 3                                            IPCC AR6 WGI

 1   The CMIP6 model ensemble reproduces the observed historical global surface temperature trend and
 2   variability with biases small enough to support detection and attribution of human-induced warming
 3   (very high confidence). The CMIP6 multi-model mean GSAT anomaly between 1850–1900 and 2010–2019
 4   is close to the best estimate of observed warming, but some CMIP6 models simulate a warming that is
 5   outside the assessed 5–95% range of observed warming. CMIP6 models broadly reproduce surface
 6   temperature variations over the past millennium, including the cooling that follows periods of intense
 7   volcanism (medium confidence). For upper air temperature, there is medium confidence that most CMIP5 and
 8   CMIP6 models overestimate observed warming in the upper tropical troposphere by at least 0.1°C per
 9   decade over the period 1979 to 2014. The latest updates to satellite-derived estimates of stratospheric
10   temperature have resulted in decreased differences between simulated and observed changes of global mean
11   temperature through the depth of the stratosphere (medium confidence). {3.3.1}
13   The slower rate of GMST increase observed over 1998–2012 compared to 1951–2012 was a temporary
14   event followed by a strong GMST increase (very high confidence). Improved observational data sets since
15   AR5 show a larger GMST trend over 1998–2012 than earlier estimates. All the observed estimates of the
16   1998–2012 GMST trend lie within the 10th–90th percentile range of CMIP6 simulated trends (high
17   confidence). Internal variability, particularly Pacific Decadal Variability, and variations in solar and volcanic
18   forcings partly offset the anthropogenic surface warming trend over the 1998–2012 period (high confidence).
19   Global ocean heat content continued to increase throughout this period, indicating continuous warming of
20   the entire climate system (very high confidence). Since 2012, GMST has warmed strongly, with the past five
21   years (2016–2020) being the warmest five-year period in the instrumental record since at least 1850 (high
22   confidence). {Cross-Chapter Box 3.1, 3.3.1; 3.5.1}
24   It is likely that human influence has contributed to 3 moistening in the upper troposphere since 1979.
25   Also, there is medium confidence that human influence contributed to a global increase in annual surface
26   specific humidity, and medium confidence that it contributed to a decrease in surface relative humidity over
27   mid-latitude Northern Hemisphere continents during summertime. {3.3.2}
29   It is likely that human influence has contributed to observed large-scale precipitation changes since the
30   mid-20th century. New attribution studies strengthen previous findings of a detectable increase in Northern
31   Hemisphere mid- to high-latitude land precipitation (high confidence). Human influence has contributed to
32   strengthening the zonal mean precipitation contrast between the wet tropics and dry subtropics (medium
33   confidence). Yet, anthropogenic aerosols contributed to decreasing global land summer monsoon
34   precipitation from the 1950s to 1980s (medium confidence). There is also medium confidence that human
35   influence has contributed to high-latitude increases and mid-latitude decreases in Southern Hemisphere
36   summertime precipitation since 1979 associated with the trend of the Southern Annular Mode toward its
37   positive phase. Despite improvements, models still have deficiencies in simulating precipitation patterns,
38   particularly over the tropical ocean (high confidence). {3.3.2, 3.3.3, 3.5.2}
40   Human-induced greenhouse gas forcing is the main driver of the observed changes in hot and cold
41   extremes on the global scale (virtually certain) and on most continents (very likely). It is likely that human
42   influence, in particular due to greenhouse gas forcing, is the main driver of the observed intensification of
43   heavy precipitation in global land regions during recent decades. There is high confidence in the ability of
44   models to capture the large-scale spatial distribution of precipitation extremes over land. The magnitude and
45   frequency of extreme precipitation simulated by CMIP6 models are similar to those simulated by CMIP5
46   models (high confidence). {Cross-Chapter Box 3.2}
48   It is likely that human influence has contributed to the poleward expansion of the zonal mean Hadley
49   cell in the Southern Hemisphere since the 1980s. There is medium confidence that the observed poleward
50   expansion of the zonal mean Hadley cell in the Northern Hemisphere is within the range of internal
51   variability. The causes of the observed strengthening of the Pacific Walker circulation since the 1980s are
52   not well understood, and the observed strengthening trend is outside the range of trends simulated in the

      In this chapter the phrase ‘human influence has contributed to’ an observed change means that the response to human influence is
     nonzero and consistent in sign with the observed change.
     Do Not Cite, Quote or Distribute                               3-5                                              Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   coupled models (medium confidence). While CMIP6 models capture the general characteristics of the
 2   tropospheric large-scale circulation (high confidence), systematic biases exist in the mean frequency of
 3   atmospheric blocking events, especially in the Euro-Atlantic sector, some of which reduce with increasing
 4   model resolution (medium confidence). {3.3.3}
 6   Human Influence on the Cryosphere
 8   It is very likely that anthropogenic forcing, mainly due to greenhouse gas increases, was the main
 9   driver of Arctic sea ice loss since the late 1970s. There is new evidence that increases in anthropogenic
10   aerosols have offset part of the greenhouse gas-induced Arctic sea ice loss since the 1950s (medium
11   confidence). In the Arctic, despite large differences in the mean sea ice state, loss of sea ice extent and
12   thickness during recent decades is reproduced in all CMIP5 and CMIP6 models (high confidence). By
13   contrast, global climate models do not generally capture the small observed increase in Antarctic sea ice
14   extent during the satellite era, and there is low confidence in attributing the causes of this change. {3.4.1}
16   It is very likely that human influence contributed to the observed reductions in Northern Hemisphere
17   spring snow cover since 1950. The seasonal cycle in Northern Hemisphere snow cover is better reproduced
18   by CMIP6 than by CMIP5 models (high confidence). Human influence was very likely the main driver of the
19   recent global, near-universal retreat of glaciers. It is very likely that human influence has contributed to the
20   observed surface melting of the Greenland Ice Sheet over the past two decades, and there is medium
21   confidence in an anthropogenic contribution to recent overall mass loss from the Greenland Ice Sheet.
22   However, there is only limited evidence, with medium agreement, of human influence on Antarctic Ice Sheet
23   mass balance through changes in ice discharge. {3.4.2, 3.4.3}
25   Human Influence on the Ocean
27   It is extremely likely that human influence was the main driver of the ocean heat content increase
28   observed since the 1970s, which extends into the deeper ocean (very high confidence). Since AR5, there
29   is improved consistency between recent observed estimates and model simulations of changes in upper
30   (<700 m) ocean heat content, when accounting for both natural and anthropogenic forcings. Updated
31   observations and model simulations show that warming extends throughout the entire water column (high
32   confidence), with CMIP6 models simulating 58% of industrial-era heat uptake (1850 to 2014) in the upper
33   layer (0–700 m), 21% in the intermediate layer (700–2000 m) and 22% in the deep layer (>2000 m). The
34   structure and magnitude of multi-model mean ocean temperature biases have not changed substantially
35   between CMIP5 and CMIP6 (medium confidence). {3.5.1}
37   It is extremely likely that human influence has contributed to observed near-surface and subsurface
38   oceanic salinity changes since the mid-20th century. The associated pattern of change corresponds to fresh
39   regions becoming fresher and salty regions becoming saltier (high confidence). Changes to the coincident
40   atmospheric water cycle and ocean–atmosphere fluxes (evaporation and precipitation) are the primary
41   drivers of the observed basin-scale salinity changes (high confidence). The observed depth-integrated basin-
42   scale salinity changes have been attributed to human influence, with CMIP5 and CMIP6 models able to
43   reproduce these patterns only in simulations that include greenhouse gas increases (medium confidence). The
44   basin-scale changes are consistent across models and intensify through the historical period (high
45   confidence). The structure of the biases in the multi-model mean has not changed substantially between
46   CMIP5 and CMIP6 (medium confidence). {3.5.2}
48   Combining the attributable contributions from glaciers, ice sheet surface mass balance and thermal
49   expansion, it is very likely that human influence was the main driver of the observed global mean sea
50   level rise since at least 1970. Since the AR5, studies have shown that simulations that exclude
51   anthropogenic greenhouse gases are unable to capture the sea level rise due to thermal expansion
52   (thermosteric) during the historical period and that model simulations that include all forcings
53   (anthropogenic and natural) most closely match observed estimates. It is very likely that human influence
54   was the main driver of the observed global mean thermosteric sea level increase since 1970. {3.5.3, 3.5.1,
55   3.4.3}
     Do Not Cite, Quote or Distribute                       3-6                                        Total pages: 202
     Final Government Distribution                     Chapter 3                                    IPCC AR6 WGI

 2   While observations show that the Atlantic Meridional Overturning Circulation (AMOC) has
 3   weakened from the mid-2000s to the mid-2010s (high confidence) and the Southern Ocean upper
 4   overturning cell has strengthened since the 1990s (low confidence), observational records are too short
 5   to determine the relative contributions of internal variability, natural forcing, and anthropogenic
 6   forcing to these changes (high confidence). No changes in Antarctic Circumpolar Current transport or
 7   meridional position have been observed. The mean zonal and overturning circulations of the Southern Ocean
 8   and the mean overturning circulation of the North Atlantic (AMOC) are broadly reproduced by CMIP5 and
 9   CMIP6 models. However, biases are apparent in the modelled circulation strengths (high confidence) and
10   their variability (medium confidence). {3.5.4}
12   Human Influence on the Biosphere
14   The main driver of the observed increase in the amplitude of the seasonal cycle of atmospheric CO2 is
15   enhanced fertilization of plant growth by the increasing concentration of atmospheric CO2 (medium
16   confidence). However, there is only low confidence that this CO2 fertilization has also been the main driver
17   of observed greening because land management is the dominating factor in some regions. Earth system
18   models simulate globally averaged land carbon sinks within the range of observation-based estimates (high
19   confidence), but global-scale agreement masks large regional disagreements. {3.6.1}
21   It is virtually certain that the uptake of anthropogenic CO2 was the main driver of the observed
22   acidification of the global ocean. The observed increase in CO2 concentration in the subtropical and
23   equatorial North Atlantic since 2000 is likely associated in part with an increase in ocean temperature, a
24   response that is consistent with the expected weakening of the ocean carbon sink with warming. Consistent
25   with AR5 there is medium confidence that deoxygenation in the upper ocean is due in part to human
26   influence. There is high confidence that Earth system models simulate a realistic time evolution of the global
27   mean ocean carbon sink. {3.6.2}
29   Human Influence on Modes of Climate Variability
31   It is very likely that human influence has contributed to the observed trend towards the positive phase
32   of the Southern Annular Mode (SAM) since the 1970s and to the associated strengthening and
33   southward shift of the Southern Hemispheric extratropical jet in austral summer. The influence of
34   ozone forcing on the SAM trend has been small since the early 2000s compared to earlier decades,
35   contributing to a weaker SAM trend observed over 2000–2019 (medium confidence). Climate models
36   reproduce the summertime SAM trend well, with CMIP6 models outperforming CMIP5 models (medium
37   confidence). By contrast, the cause of the Northern Annular Mode (NAM) trend towards its positive phase
38   since the 1960s and associated northward shifts of Northern Hemispheric extratropical jet and storm track in
39   boreal winter is not well understood. Models reproduce observed spatial features and variance of the SAM
40   and NAM very well (high confidence). {3.3.3, 3.7.1, 3.7.2}
42   Human influence has not affected the principal tropical modes of interannual climate variability or
43   their associated regional teleconnections beyond the range of internal variability (high confidence).
44   Further assessment since the AR5 confirms that climate and Earth system models are able to reproduce most
45   aspects of the spatial structure and variance of the El Niño–Southern Oscillation and Indian Ocean Basin and
46   Dipole modes (medium confidence). However, despite a slight improvement in CMIP6, some underlying
47   processes are still poorly represented. In the Tropical Atlantic basin, which contains the Atlantic Zonal and
48   Meridional modes, major biases in modelled mean state and variability remain. {3.7.3 to 3.7.5}
50   There is medium confidence that anthropogenic and volcanic aerosols contributed to observed changes
51   in the Atlantic Multi-decadal Variability (AMV) index and associated regional teleconnections since
52   the 1960s, but there is low confidence in the magnitude of this influence There is high confidence that
53   internal variability is the main driver of Pacific Decadal Variability (PDV) observed since pre-industrial
54   times, despite some modelling evidence for potential human influence. Uncertainties remain in quantification
55   of the human influence on AMV and PDV due to brevity of the observational records, limited model
     Do Not Cite, Quote or Distribute                     3-7                                       Total pages: 202
    Final Government Distribution                  Chapter 3                                 IPCC AR6 WGI

1   performance in reproducing related sea surface temperature anomalies despite improvements from CMIP5 to
2   CMIP6 (medium confidence), and limited process understanding of their key drivers. {3.7.6, 3.7.7}

    Do Not Cite, Quote or Distribute                  3-8                                   Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   3.1   Scope and Overview
 3   This chapter assesses the extent to which the climate system has been affected by human influence and to
 4   what extent climate models are able to simulate observed mean climate, changes and variability. This
 5   assessment is the basis for understanding what impacts of anthropogenic climate change may already be
 6   occurring and informs our confidence in climate projections. Moreover, an understanding of the amount of
 7   human-induced global warming to date is key to assessing our status with respect to the Paris Agreement
 8   goals of holding the increase in global average temperature to well below 2°C above pre-industrial levels and
 9   pursuing efforts to limit the temperature increase to 1.5°C (COP21, UNFCCC (2015)).
11   The evidence of human influence on the climate system has strengthened progressively over the course of
12   the previous five IPCC assessments, from the Second Assessment Report that concluded ‘the balance of
13   evidence suggests a discernible human influence on climate’ through to the Fifth Assessment Report (AR5)
14   which concluded that ‘it is extremely likely that human influence caused more than half of the observed
15   increase in global mean surface temperature (GMST) from 1951 to 2010’ (see also Sections 1.3.4 and
16 AR5 concluded that climate models had been developed and improved since the Fourth Assessment
17   Report (AR4) and were able to reproduce many features of observed climate. Nonetheless, several
18   systematic biases were identified (Flato et al., 2013). This chapter additionally builds on the assessment of
19   attribution of global temperatures contained in the IPCC Special Report on Global Warming of 1.5°C
20   (SR1.5; IPCC, 2018), assessments of attribution of changes in the ocean and cryosphere in the IPCC Special
21   Report on the Ocean and Cryosphere in a Changing Climate (SROCC; IPCC, 2019), and assessments of
22   attribution in changes in the terrestrial carbon cycle in the IPCC Special Report on Climate Change and Land
23   (IPCC, 2019a).
25   This chapter assesses the evidence for human influence on observed large-scale indicators of climate change
26   that are described in Cross-Chapter Box 2.2 and assessed in Chapter 2. It takes advantage of the longer
27   period of record now available in many observational datasets. The assessment of the human-induced
28   contribution to observed climate change requires an estimate of the expected response to human influence, as
29   well as an estimate of the expected climate evolution due to natural forcings and an estimate of variability
30   internal to the climate system (internal climate variability). For this we need high quality models, primarily
31   climate and Earth system models. Since the AR5, a new set of coordinated model results from the World
32   Climate Research Programme (WCRP) Coupled Model Intercomparison Project Phase 6 (CMIP6; Eyring et
33   al.,2016a) has become available. Together with updated observations of large-scale indicators of climate
34   change (Chapter 2), CMIP simulations are a key resource for assessing human influences on the climate
35   system. Pre-industrial control and historical simulations are of most relevance for model evaluation and
36   assessment of internal variability, and these simulations are evaluated to assess fitness-for-purpose for
37   attribution, which is the focus of this chapter (see also Section 1.5.4). This chapter provides the primary
38   evaluation of large-scale indicators of climate change in this report, and is complemented by other fitness-
39   for-purpose evaluations in subsequent chapters. CMIP6 also includes an extensive set of idealized and single
40   forcing experiments for attribution (Eyring et al., 2016a; Gillett et al., 2016a). In addition to the assessment
41   of model performance and human influence on the climate system during the instrumental era up to the
42   present-day, this chapter also includes evidence from paleo-observations and simulations over past millennia
43   (Kageyama et al., 2018a).
45   Whereas in previous IPCC Assessment Reports the comparison of simulated and observed climate change
46   was done separately in a model evaluation chapter and a chapter on detection and attribution, in AR6 these
47   comparisons are integrated together. This has the advantage of allowing a single discussion of the full set of
48   explanations for any inconsistency in simulated and observed climate change, including missing forcings,
49   errors in the simulated response to forcings, and observational errors, as well as an assessment of the
50   application of detection and attribution techniques to model evaluation. Where simulated and observed
51   changes are consistent, this can be interpreted both as supporting attribution statements, and as giving
52   confidence in simulated future change in the variable concerned (see also Box 4.1). However, if a model’s
53   simulation of historical climate change has been tuned to agree with observations, or if the models used in an
54   attribution study have been selected or weighted on the basis of the realism of their simulated climate
55   response, this information would need to be considered in the assessment and any attribution results
     Do Not Cite, Quote or Distribute                      3-9                                       Total pages: 202
     Final Government Distribution                     Chapter 3                                   IPCC AR6 WGI

 1   correspondingly tempered. An integrated discussion of evaluation and attribution supports such a robust and
 2   transparent assessment.
 4   This chapter starts with a brief description of methods for detection and attribution of observed changes in
 5   Section 3.2, which builds on the more general introduction to attribution approaches in Cross-Working
 6   Group Box: Attribution (Chapter 1). In this chapter we assess the detection of anthropogenic influence on
 7   climate on large spatial scales and long temporal scales, a concept related to, but distinct from, that of
 8   emergence of anthropogenically-induced climate change from the range of internal variability on local scales
 9   and shorter timescales (Section The following sections address the climate system component-by-
10   component, in each case assessing human influence and evaluating climate models’ simulations of the
11   relevant aspects of climate and climate change. This chapter assesses the evaluation and attribution of
12   continental and ocean basin-scale indicators of climate change in the atmosphere and at the Earth’s surface
13   (Section 3.3), cryosphere (Section 3.4), ocean (Section 3.5), and biosphere (Section 3.6), and the evaluation
14   and attribution of modes of variability (Section 3.7), the period of slower warming in the early 21st century
15   (Cross-Chapter Box 3.1) and large-scale changes in extremes (Cross-Chapter Box 3.2). Model evaluation
16   and attribution on sub-continental scales are not covered here, since these are assessed in the Atlas and in
17   Chapter 10, and extreme event attribution is not covered since it is assessed in Chapter 11. Section 3.8
18   assesses multivariate attribution and integrative measures of model performance based on multiple variables,
19   as well as process representation in different classes of models. The chapter structure is summarised in
20   Figure 3.1.
25   Figure 3.1: Visual Abstract for Chapter 3: Human influence on the climate system.
27   [END FIGURE 3.1 HERE]
30   3.2   Methods
32   New methods for model evaluation that are used in this chapter are described in Section 1.5.4. These include
33   new techniques for process-based evaluation of climate and Earth system models against observations that
34   have rapidly advanced since the publication of AR5 (Eyring et al., 2019) as well as newly developed CMIP
35   evaluation tools that allow a more rapid and comprehensive evaluation of the models with observations
36   (Eyring et al., 2016b, 2016a).
38   In this chapter, we use the Earth System Model Evaluation Tool (ESMValTool, Eyring et al. (2020); Lauer et
39   al. (2020); Righi et al. (2020)) and the NCAR Climate Variability Diagnostic Package (CVDP, Phillips et al.,
40   2014) that is included in the ESMValTool to produce most of the figures. This ensures traceability of the
41   results and provides an additional level of quality control. The ESMValTool code to produce the figures in
42   this chapter is released as open source software at the time of the publication of AR6 (see details in the
43   Chapter Data Table, Table SM.3.1). Figures in this chapter are produced either using one ensemble member
44   from each model, or using all available ensemble members and weighting each simulation by 1/(𝑁𝑁𝑀𝑀𝑖𝑖 ),
45   where N is the number of models and Mi is the ensemble size of the ith model, prior to calculating means and
46   percentiles. Both approaches ensure that each model used is given equal weight in the figures, and details on
47   which approach is used are provided in the figure captions.
49   An introduction to recent developments in detection and attribution methods in the context of this report is
50   provided in Cross-Working Group Box: Attribution (in Chapter 1). Here we discuss new methods and
51   improvements applicable to the attribution of changes in large-scale indicators of climate change which are
52   used in this chapter.

     Do Not Cite, Quote or Distribute                     3-10                                     Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   3.2.1   Methods Based on Regression
 3   Regression-based methods, also known as fingerprinting methods, have been widely used for detection of
 4   climate change and attribution of the change to different external drivers. Initially, these methods were
 5   applied to detect changes in global surface temperature, and were then extended to other climate variables at
 6   different time and spatial scales (e.g., Hegerl et al., 1996; Hasselmann, 1997; Allen and Tett, 1999; Gillett et
 7   al., 2003b; Zhang et al., 2007; Min et al., 2008a, 2011). These approaches are based on multivariate linear
 8   regression and assume that the observed change consists of a linear combination of externally forced signals
 9   plus internal variability, which generally holds for large-scale variables (Hegerl and Zwiers, 2011). The
10   regressors are the expected space-time response patterns to different climate forcings (fingerprints), and the
11   residuals represent internal variability. Fingerprints are usually estimated from climate model simulations
12   following spatial and temporal averaging. A regression coefficient which is significantly different from zero
13   implies that a detectable change is identified in the observations. When the confidence interval of the
14   regression coefficient includes unity and is inconsistent with zero, the magnitude of the model simulated
15   fingerprints is assessed to be consistent with the observations, implying that the observed changes can be
16   attributed in part to a particular forcing. Variants of linear regression have been used to address uncertainty
17   in the fingerprints due to internal variability (Allen and Stott, 2003) as well as the structural model
18   uncertainty (Huntingford et al., 2006).
20   In order to improve the signal-to-noise ratio, observations and model-simulated responses are usually
21   normalized by an estimate of internal variability derived from climate model simulations. This procedure
22   requires an estimate of the inverse covariance matrix of the internal variability, and some approaches have
23   been proposed for more reliable estimation of this (Ribes et al., 2009). A signal can be spuriously detected
24   due to too small noise, and hence simulated internal variability needs to be evaluated with care. Model-
25   simulated variability is typically checked through comparing modelled variance from unforced simulations
26   with the observed residual variance using a standard residual consistency test (Allen and Tett, 1999b), or an
27   improved one (Ribes and Terray, 2013). Imbers et al. (2014) tested the sensitivity of detection and attribution
28   results to the different representations of internal variability associated with short-memory and long-memory
29   processes. Their results supported the robustness of previous detection and attribution statements for the
30   global mean temperature change but they also recommended the use of a wider variety of robustness tests.
32   Some recent studies focused on the improved estimation of the scaling factor (regression coefficient) and its
33   confidence interval. Hannart et al. (2014) describe an inference procedure for scaling factors which avoids
34   making the assumption that model error and internal variability have the same covariance structure. An
35   integrated approach to optimal fingerprinting was further suggested in which all uncertainty sources (i.e.,
36   observational error, model error, and internal variability) are treated in one statistical model without a
37   preliminary dimension reduction step (Hannart, 2016). Katzfuss et al. (2017) introduced a similar integrated
38   approach based on a Bayesian model averaging. On the other hand, DelSole et al. (2019) suggested a
39   bootstrap method to better estimate the confidence intervals of scaling factors even in a weak-signal regime.
40   It is notable that some studies do not optimise fingerprints, as uncertainty in the covariance introduces a
41   further layer of complexity, but results in only a limited improvement in detection (Polson and Hegerl,
42   2017).
44   Another fingerprinting approach uses pattern similarity between observations and fingerprints, in which the
45   leading empirical orthogonal function obtained from the time-evolving multi-model forced simulation is
46   usually defined as a fingerprint (e.g., Santer et al., 2013; Marvel et al., 2019; Bonfils et al., 2020).
47   Observations and model simulations are then projected onto the fingerprint to measure the degree of spatial
48   pattern similarity with the expected physical response to a given forcing. This projection provides the signal
49   time series, which is in turn tested against internal variability, as estimated from long control simulations. As
50   a way to extend this pattern-based approach to a high-dimensional detection variable at daily timescales,
51   Sippel et al. (2019, 2020) proposed using the relationship pattern with a global climate change metric as a
52   fingerprint. To solve the high-dimensional regression problem which makes regression coefficients not well
53   constrained, they incorporated a statistical learning technique based on a regularized linear regression, which
54   optimizes a global warming signal by giving lower weight to regions with large internal variability.
     Do Not Cite, Quote or Distribute                      3-11                                       Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 2   3.2.2    Other Probabilistic Approaches
 4   Considering the difficulty in accounting for climate modelling uncertainties in the regression-based
 5   approaches, Ribes et al. (2017) introduced a new statistical inference framework based on an additivity
 6   assumption and likelihood maximization, which estimates climate model uncertainty based on an ensemble
 7   of opportunity and tests whether observations are inconsistent with internal variability and consistent with
 8   the expected response from climate models. The method was further developed by Ribes et al. (2021), who
 9   applied it to narrow the uncertainty range in the estimated human-induced warming. Hannart and Naveau
10   (2018), on the other hand, extended the application of standard causal theory (Pearl, 2009) to the context of
11   detection and attribution by converting a time series into an event, and calculating the probability of
12   causation, an approach which maximizes the causal evidence associated with the forcing. On the other hand,
13   Schurer et al. (2018) employed a Bayesian framework to explicitly consider climate modelling uncertainty in
14   the optimal regression method. Application of these approaches to attribution of large-scale temperature
15   changes supports the dominant anthropogenic contribution to the observed global warming.
17   Climate change signals can vary with time and discriminant analysis has been used to obtain more accurate
18   estimates of time-varying signals, and has been applied to different variables such as seasonal temperatures
19   (Jia and DelSole, 2012) and the South Asian monsoon (Srivastava and DelSole, 2014). The same approach
20   was applied to separate aerosol forcing responses from other forcings (Yan et al., 2016b) and results using
21   climate model output indicated that detectability of the aerosol response is maximised by using a
22   combination of temperature and precipitation data. Paeth et al. (2017) introduced a detection and attribution
23   method applicable for multiple variables based on a discriminant analysis and a Bayesian classification
24   method. Finally, a systematic approach has been proposed to translating quantitative analysis into a
25   description of confidence in the detection and attribution of a climate response to anthropogenic drivers
26   (Stone and Hansen, 2016).
28   Overall, these new fingerprinting and other probabilistic methods for detection and attribution as well as
29   efforts to better incorporate the associated uncertainties have addressed a number of shortcomings in
30   previously applied detection and attribution techniques. They further strengthen the confidence in attribution
31   of observed large-scale changes to a combination of external forcings as assessed in the following sections.
34   3.3     Human Influence on the Atmosphere and Surface
36   3.3.1    Temperature
38 Surface Temperature
40   Surface temperature change is the aspect of climate in which the climate research community has had most
41   confidence over past IPCC Assessment Reports. This confidence comes from the availability of longer
42   observational records compared to other indicators, a large response to anthropogenic forcing compared to
43   variability in the global mean, and a strong theoretical understanding of the key thermodynamics driving its
44   changes (Collins et al., 2010; Shepherd, 2014). AR5 assessed that it was extremely likely that human
45   activities had caused more than half of the observed increase in global mean surface temperature from 1951
46   to 2010, and virtually certain that internal variability alone could not account for the observed global
47   warming since 1951 (Bindoff et al., 2013). The AR5 also assessed with very high confidence that climate
48   models reproduce the general features of the global-scale annual mean surface temperature increase over
49   1850-2011 and with high confidence that models reproduce global and Northern Hemisphere (NH)
50   temperature variability on a wide range of time scales (Flato et al., 2013). This section assesses the
51   performance of the new generation CMIP6 models (see Table AII.5) in simulating the patterns, trends, and
52   variability of surface temperature, and the evidence from detection and attribution studies of human
53   influence on large-scale changes in surface temperature.
55   Model evaluation
     Do Not Cite, Quote or Distribute                     3-12                                      Total pages: 202
     Final Government Distribution                          Chapter 3                                       IPCC AR6 WGI

 1   To be fit for detecting and attributing human influence on globally-averaged surface temperatures, climate
 2   models need to represent, based on physical principles, both the response of surface temperature to external
 3   forcings and the internal variability in surface temperature over various time scales. This section assesses the
 4   performance of those aspects in the latest generation CMIP6 climate models. See Section 3.8 for evaluation
 5   at continental scales, Chapter 10 for model evaluation in the context of regional climate information, and the
 6   Atlas for region-by-region assessments of model performance.
 8   Reconstructions of past temperature from paleoclimate proxies (Cross-Chapter Box 2.1, Section
 9   have been used to evaluate modelled past climate temperature change patterns. The AR5 found that CMIP5
10   (Taylor et al., 2012) models were able to reproduce the large-scale patterns of temperature during the Last
11   Glacial Maximum (LGM) (Flato et al., 2013) and simulated a polar amplification broadly consistent with
12   reconstructions for warm (Pliocene and Eocene) and cold (LGM) periods (Masson-Delmotte et al., 2013a).
13   Since AR5, a better understanding of temperature proxies and their uncertainties and in some cases the
14   forcing applied to model simulations has led to better agreement between models and reconstructions over a
15   wide range of past climates. For the Pliocene and Eocene warm periods, understanding of uncertainties in
16   temperature proxies (Hollis et al., 2019; McClymont et al., 2020) and the boundary conditions used in
17   climate simulations (Haywood et al., 2016; Lunt et al., 2017) has improved, and some models now agree
18   better with temperature proxies for these time periods compared to models assessed in AR5 (Zhu et al.,
19   2019; Haywood et al., 2020; Lunt et al., 2021) (Sections;; Cross-Chapter Box 2.4). For the
20   Last Interglacial (LIG), improved temporal resolution of temperature proxies (Capron et al., 2017) and better
21   appreciation of the importance of freshwater forcing (Stone et al., 2016) have clarified the reasons behind
22   apparent model-data inconsistencies. Regional LIG temperature responses simulated by CMIP6 are within
23   the uncertainty ranges of reconstructed temperature responses, except in regions where unresolved changes
24   in regional ocean circulation, meltwater, or vegetation changes may cause model mismatches (Otto-Bliesner
25   et al., 2021). For the LGM, the CMIP5 and CMIP6 ensembles compare similarly to new sea surface
26   temperature (SST) and surface air temperature (SAT) proxy reconstructions (Cleator et al., 2020; Tierney et
27   al., 2020b) (Figure 3.2a). The very cold CMIP6 LGM simulation by CESM2.1 is an exception related to the
28   high Equilibrium Climate Sensitivity (ECS) of that model (Section 7.5.6) (Kageyama et al., 2021a; Zhu et
29   al., 2021). Figure 3.2a illustrates the wide range of simulated global LGM temperature responses in both
30   ensembles. CMIP6 models tend to underestimate the cooling over land, but agree better with oceanic
31   reconstructions. For the mid-Holocene, the regional biases found in CMIP5 simulations are similar to those
32   in pre-industrial and historical simulations (Harrison et al., 2015; Ackerley et al., 2017), suggesting common
33   causes. CMIP5 models underestimate Arctic warming in the mid-Holocene (Yoshimori and Suzuki, 2019).
34   CMIP6 models simulate a mid-latitude, subtropical, and tropical cooling compared to pre-industrial, whereas
35   temperature proxies indicate a warming (Brierley et al., 2020; Kaufman et al., 2020; see also Section
36, although accounting for seasonal effects in the proxies may reduce the discrepancy (Bova et al.,
37   2021). Over the past millennium, reconstructed and simulated temperature anomalies, internal variability,
38   and forced response agree well over NH continents, but those statistics disagree strongly in the Southern
39   Hemisphere (SH), where models seem to overestimate the response (PAGES 2k-PMIP3, 2015). That
40   disagreement is partly explained by the lower quality of the reconstructions in the SH, but model and/or
41   forcing errors may also contribute (Neukom et al., 2018). Figure 3.2b shows that land/sea warming contrast
42   behaves coherently in model simulations across multiple periods, with a slight non-linearity in land warming
43   due to a smaller contribution of snow cover to temperature response in warmer climates. A metric-based
44   assessment of paleoclimate model simulations is carried out in Section 3.8.2.
49   Figure 3.2: Changes in surface temperature for different paleoclimates. (a) Comparison of reconstructed and
50                modelled surface temperature anomalies for the Last Glacial Maximum over land and ocean in the
51                Tropics (30°N to 30°S). Land-based reconstructions are from Cleator et al., (2020). Ocean-based
52                reconstructions are from Tierney et al. (2020). Model points are calculated as the difference between Last
53                Glacial Maximum and pre-industrial control simulations of the PMIP3 and PMIP4 ensembles, sampled at
54                the reconstruction data points. (b) Land-sea contrast in global mean surface temperature change for
55                different paleoclimates. Crosses show individual model simulations from the CMIP5 and CMIP6
56                ensembles. Filled symbols show ensemble means and assessed values. Acronyms are LGM Last Glacial
     Do Not Cite, Quote or Distribute                          3-13                                          Total pages: 202
     Final Government Distribution                          Chapter 3                                         IPCC AR6 WGI

 1                Maximum, LIG Last Inter Glacial, MPWP mid-Pliocene Warm Period, EECO Early Eocene Climatic
 2                Optimum. (c) Upper panel shows time series of volcanic radiative forcing, in W m−2, as used in the
 3                CMIP5 (Gao et al., 2008; Crowley and Unterman, 2013; see also Schmidt et al., 2011) and CMIP6 (850
 4                BCE to 1900 CE from Toohey and Sigl (2017), 1850-2015 from Luo (2018)). The forcing has been
 5                calculated from the stratospheric aerosol optical depth at 550 nm shown in Figure 2.2. Lower panel shows
 6                time series of global mean surface temperature anomalies, in K, with respect to 1850-1900 for the CMIP5
 7                and CMIP6 past1000 simulations and their historical continuation simulations. Simulations are coloured
 8                according to the volcanic radiative forcing dataset they used. The temperature median reconstruction by
 9                PAGES 2k Consortium (2019) is shown in black, the 5-95% confidence interval by grey lines and the
10                grey envelopes show the 1, 5, 15, 25, 35, 45, 55, 65, 75, 85, 95, and 99 percentiles. All data in both panels
11                are band-passed, where frequencies longer than 20 years have been retained. Further details on data
12                sources and processing are available in the chapter data table (Table 3.SM.1).
14   [END FIGURE 3.2 HERE]
17   For the historical period, AR5 assessed with very high confidence that CMIP5 models reproduced observed
18   large-scale mean surface temperature patterns, although errors of several degrees appear in elevated regions,
19   like the Himalayas and Antarctica, near the edge of the sea ice in the North Atlantic, and in upwelling
20   regions. This assessment is updated here for the CMIP6 simulations. Figure 3.3 shows the annual-mean
21   surface air temperature at 2 m for the CMIP5 and CMIP6 multi-model means, both compared to the ERA5
22   reanalysis (see Section 1.5.2) for the period 1995-2014. The distribution of biases is similar in CMIP5 and
23   CMIP6 models, as already noted by several studies (Crueger et al., 2018; Găinuşă-Bogdan et al., 2018;
24   Kuhlbrodt et al., 2018; Lauer et al., 2018). Arctic temperature biases seem more widespread in both
25   ensembles than assessed at the time of the AR5. The fundamental causes of temperature biases remain
26   elusive, with errors in clouds (Lauer et al., 2018), oceanic circulation (Kuhlbrodt et al., 2018), winds (Lauer
27   et al., 2018), and surface energy budget (Hourdin et al., 2015; Séférian et al., 2016; Găinuşă-Bogdan et al.,
28   2018) being frequently cited candidates. Increasing horizontal resolution shows promise of decreasing long-
29   standing biases in surface temperature over large regions (Bock et al., 2020). Panels e and f of Figure 3.3
30   show that biases in the mean High-Resolution Model Intercomparison Project (HighResMIP, Haarsma et al.,
31   2016) models (see also Table AII.6) are smaller than those in the mean of the corresponding lower-resolution
32   versions of the same models simulating the same period (see also Section However, the bias
33   reduction is modest (Palmer and Stevens, 2019). In addition, the biases of the limited number of models
34   participating in HighResMIP are not entirely representative of overall CMIP6 biases, especially in the
35   Southern Ocean, as indicated by comparing panels b and f of Figure 3.3.
40    Figure 3.3: Annual-mean surface (2 m) air temperature (°C) for the period 1995–2014. (a) Multi-model
41                (ensemble) mean constructed with one realization of the CMIP6 historical experiment from each model.
42                (b) Multi-model mean bias, defined as the difference between the CMIP6 multi-model mean and the
43                climatology of the Fifth generation of ECMWF atmospheric reanalyses of the global climate (ERA5).
44                (c) Multi-model mean of the root mean square error calculated over all months separately and averaged
45                with respect to the climatology from ERA5. (d) Multi-model-mean bias as the difference between the
46                CMIP6 multi-model mean and the climatology from ERA5. Also shown is the multi-model mean bias
47                as the difference between the multi-model mean of (e) high resolution and (f) low resolution
48                simulations of four HighResMIP models and the climatology from ERA5. Uncertainty is represented
49                using the advanced approach: No overlay indicates regions with robust signal, where ≥66% of models
50                show change greater than variability threshold and ≥80% of all models agree on sign of change;
51                diagonal lines indicate regions with no change or no robust signal, where <66% of models show a
52                change greater than the variability threshold; crossed lines indicate regions with conflicting signal,
53                where ≥66% of models show change greater than variability threshold and <80% of all models agree on
54                sign of change. For more information on the advanced approach, please refer to the Cross-Chapter Box
55                Atlas.1. Dots in panel e) mark areas where the bias in high resolution versions of the HighResMIP
56                models is lower in at least 3 out of 4 models than in the corresponding low resolution versions. Further
57                details on data sources and processing are available in the chapter data table (Table 3.SM.1).
     Do Not Cite, Quote or Distribute                          3-14                                          Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   [END FIGURE 3.3 HERE]
 4   AR5 assessed with very high confidence that models reproduce the general history of the increase in global-
 5   scale annual mean surface temperature since the year 1850, although AR5 also reported that an observed
 6   reduction in the rate of warming over the period 1998-2012 was not reproduced by the models (Flato et al.,
 7   2013) (see Cross-Chapter Box 3.1). Figure 3.2c and Figure 3.4 show time series of anomalies in annually
 8   and globally averaged surface temperature simulated by CMIP5 and CMIP6 models for the past millennium
 9   and the period 1850 to 2020, respectively, with the baseline set to 1850-1900 (see Section 1.4.1). As also
10   indicated by Figure 3.4, the spread in simulated absolute temperatures is large (Palmer and Stevens, 2019).
11   But the discussion is based on temperature anomaly time series instead of absolute temperatures because our
12   focus is on evaluation of the simulation of climate change in these models, and also because anomalies are
13   more uniformly distributed and are more easily deseasonalised to isolate long-term trends (see Section
14   1.4.1). CMIP6 models broadly reproduce surface temperature variations over the past millennium, including
15   the cooling that follows periods of intense volcanism (medium confidence) (Figure 3.2c). Simulated GMST
16   anomalies are well within the uncertainty range of temperature reconstructions (medium confidence) since
17   about the year 1300, except for some short periods immediately following large volcanic eruptions, for
18   which simulations driven by different forcing datasets disagree (Figure 3.2c). Before the year 1300, larger
19   disagreements between models and temperature reconstructions can be expected because forcing and
20   temperature reconstructions are increasingly uncertain further back in time, but specific causes have not been
21   identified conclusively (Ljungqvist et al., 2019; PAGES 2k Consortium, 2019) (medium confidence). For the
22   historical period, results for CMIP6 shown in Figure 3.4 suggest that the qualitative history of surface
23   temperature increase is well reproduced, including the increase in warming rates beginning in the 1960s and
24   the temporary cooling that follows large volcanic eruptions.
26   Although virtually all CMIP6 modelling groups report improvements in their model’s ability to simulate
27   current climate compared to the CMIP5 version (Gettelman et al., 2019; Golaz et al., 2019; Mauritsen et al.,
28   2019; Swart et al., 2019; Voldoire et al., 2019b; Wu et al., 2019c; Bock et al., 2020; Boucher et al., 2020a;
29   Dunne et al., 2020), it does not necessarily follow that the simulation of temperature trends is also improved
30   (Bock et al., 2020; Fasullo et al., 2020). The CMIP6 multi-model ensemble encompasses observed warming
31   and the multi-model mean tracks those observations within 0.2°C over most of the historical period.
32   Figure 3.4 confirms the findings of Papalexiou et al. (2020), who highlighted based on 29 CMIP6 models
33   that most models replicate the period of slow warming between 1942 and 1975 and the late twentieth century
34   warming (1975–2014). The CMIP6 multi-model mean is cooler over the period 1980-2000 than both
35   observations and CMIP5 (Bock et al., 2020; Flynn and Mauritsen, 2020; Gillett et al., 2021; Figure 3.4).
36   Biases of several tenths of a degree in some CMIP6 models over that period may be due to an overestimate
37   in aerosol radiative forcing (Andrews et al., 2020; Dittus et al., 2020; Flynn and Mauritsen, 2020) (see also
38   Section 6.3.5, Figure 6.8, and Section 7.3.3). Papalexiou et al. (2020), Stolpe et al. (2020) and Tokarska et al.
39   (2020) all report that CMIP6 models on average overestimate warming from the 1970s or 1980s to the
40   2010s, although quantitative conclusions depend on which observational dataset is compared against (see
41   also Table 2.4). However, Figure 3.4, which includes a larger number of models than available to those
42   studies, indicates that the average CMIP6 model tracks observed warming better than CMIP5 models after
43   the year 2000. The CMIP6 multi-model mean GSAT warming between 1850-1900 and 2010-2019 and
44   associated 5-95% range is 1.09°C (0.66 to 1.64°C). Cross-Chapter Box 2.3 assessed GSAT warming over the
45   same period at 1.06°C (0.88 to 1.21°C). So some CMIP6 models simulate a warming that is smaller than the
46   assessed observed range, and other CMIP6 models simulate a warming that is larger. That overestimated
47   warming may be an early symptom of overestimated equilibrium climate sensitivities (ECS) in some CMIP6
48   models (Meehl et al., 2020; Schlund et al., 2020) (see also Section 7.5.6), and has implications for
49   projections of GSAT changes (see Chapter 4) (Liang et al., 2020; Nijsse et al., 2020; Tokarska et al., 2020;
50   Ribes et al., 2021). In some models, a large ECS and a strong aerosol forcing lead to too large a mid-20th
51   century cooling followed by overestimated warming rates in the late 20th century when aerosol emissions
52   decrease (Golaz et al., 2019; Flynn and Mauritsen, 2020). Temperature biases are driven by both model
53   physics and prescribed forcing, which is a challenge for model development.
55   Chylek et al. (2020) argue that CMIP5 models overestimate the temperature response to volcanic eruptions.
     Do Not Cite, Quote or Distribute                      3-15                                       Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1   Lehner et al. (2016), Rypdal (2018) and Stolpe et al. (2020) point instead to missed compensating effects on
 2   surface temperature change associated with the El-Nino Southern Oscillation (ENSO) or the Atlantic
 3   Multidecadal Oscillation (AMO). An alternative view sees those ENSO and AMO responses as expressions
 4   of changes in climate feedbacks driven by the geographical pattern of SST changes (Andrews et al., 2018).
 5   At least one model is able to reproduce such pattern effects (Gregory and Andrews, 2016). Errors in the
 6   volcanic forcing prescribed in simulations, including for CMIP6 (Rieger et al., 2020), also introduce
 7   differences with the observed temperature response, independently of the quality of the model physics. In
 8   addition, comparisons of the modelled temperature response to large eruptions over the past millennium to
 9   temperature reconstructions based on tree rings show a much better agreement (Lücke et al., 2019; Zhu et al.,
10   2020a) than comparisons to the annual, multi-temperature proxy reconstructions shown in Figure 3.2c. These
11   considerations, and Figures 3.2c and 3.4, suggest that CMIP6 models do not systematically overestimate the
12   cooling that follows large volcanic eruptions (see also Cross-Chapter Box 4.1).
14   When interpreting model simulations of historical temperature change, it is important to keep in mind that
15   some models are tuned towards representing the observed trend in global mean surface temperature over the
16   historical period (Hourdin et al., 2017). In Figure 3.4 the CMIP6 models that are documented to be tuned to
17   reproduce observed warming, typically by tuning aerosol forcing or factors that influence the model’s ECS,
18   are marked with an asterisk. Such tuning of a model can strongly impact its temperature projections
19   (Mauritsen and Roeckner, 2020). However, Bock et al. (2020) reported that there is no statistically
20   significant difference in multi-model mean GSAT between the models that had been tuned based on
21   observed warming compared to those which had not. Moreover, only two of thirteen models used for the
22   DAMIP simulations on which CMIP6 attribution studies are based were tuned towards historical warming
23   (Bock et al., 2020; Gillett et al., 2021). Further, tuning is done on globally averaged quantities, so does not
24   substantially change the spatio-temporal pattern of response on which many regression-based attribution
25   studies are based (Bock et al., 2020). Therefore, we assess with high confidence that the tuning of a small
26   number of CMIP6 models to observed warming has not substantially influenced attribution results assessed
27   in this chapter.
32   Figure 3.4: Observed and simulated time series of the anomalies in annual and global mean near-surface air
33               temperature (GSAT). All anomalies are differences from the 1850–1900 time-mean of each individual
34               time series. The reference period 1850–1900 is indicated by grey shading. (a) Single simulations from
35               CMIP6 models (thin lines) and the multi-model mean (thick red line). Observational data (thick black
36               lines) are HadCRUT5, and are blended surface temperature (2 m air temperature over land and sea
37               surface temperature over the ocean). All models have been subsampled using the HadCRUT5
38               observational data mask. Vertical lines indicate large historical volcanic eruptions. CMIP6 models which
39               are marked with an asterisk are either tuned to reproduce observed warming directly, or indirectly by
40               tuning equilibrium climate sensitivity. Inset: GSAT for each model over the reference period, not masked
41               to any observations. (b). Multi-model means of CMIP5 (blue line) and CMIP6 (red line) ensembles and
42               associated 5 to 95 percentile ranges (shaded regions). Observational data are HadCRUT5, Berkeley Earth,
43               NOAAGlobalTemp-Interim and Kadow et al. (2020). Masking was done as in (a). CMIP6 historical
44               simulations are extended with SSP2-4.5 simulations for the period 2015-2020 and CMIP5 simulations are
45               extended with RCP4.5 simulations for the period 2006-2020. All available ensemble members were used
46               (see Section 3.2). The multi-model means and percentiles were calculated solely from simulations
47               available for the whole time span (1850-2020). Figure is updated from Bock et al. (2020), their Figures 1
48               and 2. / CC BY4.0 Further details on data sources and
49               processing are available in the chapter data table (Table 3.SM.1).
51   [END FIGURE 3.4 HERE]
54   The reliance of detection and attribution studies on climate models (see Section 3.2) requires that those
55   models simulate realistic statistics of internal variability on multi-decadal timescales. An incorrect estimate
56   of variability in models would affect confidence in the conclusions from detection and attribution. AR5
57   found that CMIP5 models simulate realistic variability in global-mean surface temperature on decadal time
     Do Not Cite, Quote or Distribute                        3-16                                        Total pages: 202
     Final Government Distribution                       Chapter 3                                      IPCC AR6 WGI

 1   scales, with variability on multi-decadal time scales being more difficult to evaluate because of the short
 2   observational record (Flato et al., 2013). Since AR5, new work has characterized the contributions of
 3   variability in different ocean areas to SST variability, with tropical modes of variability like ENSO dominant
 4   on time scales of 5 to 10 years, while longer time scales see the variance maxima move poleward to the
 5   North Atlantic, North Pacific, and Southern oceans (Monselesan et al., 2015). There may, however, be
 6   sizeable, two-way interdependencies between ENSO and sea surface temperature variability in different
 7   basins (Kumar et al., 2014; Cai et al., 2019), and ENSO’s influence on global surface temperature variability
 8   may not be confined only to decadal timescales (Triacca et al., 2014). Studies based on large ensembles of
 9   20th and 21st century climate change simulations confirm that internal variability has a substantial influence
10   on global warming trends over periods shorter than 30-40 years (Kay et al., 2015; Dai and Bloecker, 2019).
11   Although the equatorial Pacific seems to be the main source of internal variability on decadal timescales,
12   Brown et al. (2016) linked diversity in modelled oceanic convection, sea ice, and energy budget in high-
13   latitude regions to overall diversity in modelled internal variability.
15   Interest in internal variability since publication of the AR5 stems in part from its importance in
16   understanding the slower global surface temperature warming over the early 21st century (see Cross-Chapter
17   Box 3.1). Evidence coming mostly from paleo studies is mixed on whether CMIP5 models underestimate
18   decadal and multi-decadal variability in global mean temperature. Schurer et al. (2013) found good
19   agreement between internal variability derived from paleo reconstructions, estimated as the fraction of
20   variance that is not explained by forced responses, and modelled variability, although the subset of CMIP5
21   models they used may have been associated with larger variability than the full CMIP5 ensemble. PAGES 2k
22   Consortium (2019) found that the largest 51-year trends in both reconstructions of global mean temperature
23   and fully forced climate simulations over the period 850 to 1850 were almost identical. Zhu et al. (2019a)
24   showed agreement in the modelled and reconstructed temporal spectrum of global surface temperatures on
25   annual to multi-millennial timescales. However, they suggest that decadal-to-centennial variability is partly
26   forced by slow orbital changes that predate the last millennium. This is consistent with Gebbie and Huybers
27   (2019), who showed that the deep ocean has been out of equilibrium over that period. Laepple and Huybers
28   (2014) found good agreement between modelled and proxy-derived decadal ocean temperature variability,
29   but underestimates of variance by models by at least a factor of ten at centennial timescales because models
30   underestimate the difference between the warm and cold periods of the last millennium. Parsons et al. (2020)
31   found that some CMIP6 models exhibit much higher multidecadal variability in GSAT than CMIP5 models,
32   with indications that variability in these models is also higher than that from proxy reconstructions. CMIP6
33   models may not share the underestimation by CMIP5 models of variability in decadal to multidecadal modes
34   of variability, such as Pacific decadal variability (Section 3.7.6) (England et al., 2014; Thompson et al.,
35   2014; Schurer et al., 2015) and Atlantic Multidecadal Variability (AMV, which may be partly forced, see
36   Section 3.7.7) but this assessment is limited by the small number of available studies. For the SH, Hegerl et
37   al. (2018) found an instance of internal variability in the early 20th century larger than that modelled, but
38   indicated that could be an observational issue. Friedman et al. (2020) found biases in interhemispheric SST
39   contrast in some models that may be consistent with underestimated cooling after early-20th century
40   eruptions or underestimated Pacific decadal variability, but could also be due to an imperfect separation
41   between internal variability and forced signal in the observations. Figure 3.2c, updated from PAGES 2k
42   Consortium (2019), compares modelled temperatures to reconstructions over the last millennium. It indicates
43   that models reproduce the observed variability well, at least for the timescales between 20 and 50 years that
44   paleo reconstructions typically resolve and that the figure represents. In summary, decadal GMST variability
45   simulated in CMIP6 models spans the range of residual decadal variability in large-scale reconstructions
46   (medium evidence, low agreement).
48   In addition, new literature suggests that anthropogenic forcing itself may locally increase or decrease
49   variability in surface temperatures (Screen et al., 2014; Qian and Zhang, 2015; Brown et al., 2017; Park et
50   al., 2018; Santer et al., 2018; Weller et al., 2020). Those studies imply limitations in the use of pre-industrial
51   control simulations to quantify the role of unforced variability over the historical period. Some recent
52   attribution studies (Gillett et al., 2021; Ribes et al., 2021) have estimated variability from ensembles of
53   forced simulations instead, which would be expected to resolve any such changes in variability.
55   Figure 3.5 shows the standard deviation of zonal-mean surface temperature in CMIP6 pre-industrial control
     Do Not Cite, Quote or Distribute                       3-17                                       Total pages: 202
     Final Government Distribution                         Chapter 3                                      IPCC AR6 WGI

 1   simulations and observed temperature datasets. Results are consistent with those based on CMIP5 models,
 2   which showed the largest model spread where variability is also large, in the tropics and mid- to high-
 3   latitudes (Flato et al., 2013). Modelled variability is within a factor two of observed variability over most of
 4   the globe. The apparent overestimation of high latitude variability in models compared to observations may
 5   be due to interpolation and infilling over data sparse high latitude regions in the observational products
 6   shown here (Jones, 2016).
11   Figure 3.5: The standard deviation of annually averaged zonal-mean near-surface air temperature. This is
12               shown for four detrended observed temperature datasets (HadCRUT5, Berkeley Earth,
13               NOAAGlobalTemp-Interim and Kadow et al. (2020), for the years 1995-2014) and 59 CMIP6 pre-
14               industrial control simulations (one ensemble member per model, 65 years) (after Jones et al., 2013). For
15               line colours see the legend of Figure 3.4. Additionally, the multi-model mean (red) and standard deviation
16               (grey shading) are shown. Observational and model datasets were detrended by removing the least-
17               squares quadratic trend. Further details on data sources and processing are available in the chapter data
18               table (Table 3.SM.1).
20   [END FIGURE 3.5 HERE]
23   The previous paragraph took an ensemble-mean view of model performance, but individual models disagree
24   on unforced variability. Figure 3.6 illustrates the large differences in GSAT variability in unforced CMIP6
25   pre-industrial control simulations, following the method of Parsons et al. (2020). Surface temperatures in
26   pre-industrial conditions are especially variable in the ten models highlighted in Figure 3.6a, and some
27   models substantially exceed the variability seen in CMIP5 models (Parsons et al., 2020). Figure 3.6b shows
28   that the distribution of warming trends simulated by CMIP6 models in historical simulations is clearly
29   distinct from that simulated in unforced piControl simulations. Still, the unforced variability of the five most
30   variable models approaches half that observed over the historical period under anthropogenically forced
31   conditions (Figure 3.6c) (Parsons et al., 2020; Ribes et al., 2021). For the Centre National de la Recherche
32   Météorologique (CNRM) models, which are among the most variable, the large, low-frequency variability is
33   attributed to strong simulated Atlantic Multidecadal Variability (Séférian et al., 2019; Voldoire et al.,
34   2019b), which is difficult to disprove because of the short observational record (Cassou et al., 2018;
35   Section 3.7.7). But, importantly, patterns of temperature variability simulated by even the most variable
36   models differ from the pattern of forced temperature change (Parsons et al., 2020). Taken together, this
37   discussion and Figures 3.2, 3.5 and 3.6 indicate that the statistics of internal variability in models compare
38   well in most cases to observational estimates and temperature proxy reconstructions, though some CMIP6
39   models appear to have higher multidecadal variability than CMIP5 models or proxy reconstructions. When
40   used in attribution studies, models with overestimated variability would increase estimated uncertainties and
41   make results statistically conservative.
46   Figure 3.6: Simulated internal variability of global surface air temperature (GSAT) versus observed changes.
47               (a) Time series of 5-year running mean GSAT anomalies in 45 CMIP6 pre-industrial control (unforced)
48               simulations. The 10 most variable models in terms of 5-year running mean GSAT are coloured according
49               to the legend on Figure 3.4. (b) Histograms of GSAT changes in CMIP6 historical simulations (extended
50               by SSP2-4.5 simulations) from 1850-1900 to 2010-2019 are shown by pink shading in (c), and GSAT
51               changes from the first 51 years average to the last 20 years average of 170-year overlapping segments of
52               the pre-industrial control simulations shown in (a) are shown by blue shading. GMST changes in
53               observational datasets for the same period are indicated by black vertical lines. (c) Observed GMST
54               anomaly time series relative to the 1850-1900 average. Black lines represent the 5-year running means
55               while grey lines show unfiltered annual time series. Further details on data sources and processing are
56               available in the chapter data table (Table 3.SM.1).
     Do Not Cite, Quote or Distribute                        3-18                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                     IPCC AR6 WGI

 1   [END FIGURE 3.6 HERE]
 4   In summary, there is high confidence that CMIP6 models reproduce observed large-scale mean surface
 5   temperature patterns and internal variability as well as their CMIP5 predecessors, but with little evidence for
 6   reduced biases. CMIP6 models also reproduce historical GSAT changes similarly to their CMIP5
 7   counterparts (medium confidence). However, in spite of model imperfections, there is very high confidence
 8   that biases in surface temperature trends and variability simulated by the CMIP5 and CMIP6 ensembles are
 9   small enough to support detection and attribution of human-induced warming.
11   Detection and attribution
12   Looking at periods preceding the instrumental record, AR5 assessed with high confidence that the 20th
13   century annual mean surface temperature warming reversed a 5000-year cooling trend in NH mid-to-high
14   latitudes caused by orbital forcing, and attributed the reversal to anthropogenic forcing with high confidence
15   (see also Section Since AR5, the combined response to solar, volcanic and greenhouse gas forcing
16   was detected in all NH continents (PAGES 2k-PMIP3, 2015) over the period 864 to 1840. In contrast, the
17   effect of those forcings was not detectable in the SH (Neukom et al., 2018). Global and NH temperature
18   changes from reconstructions over this period have been attributed mostly to volcanic forcing (Schurer et al.,
19   2014a; McGregor et al., 2015; Otto-Bliesner et al., 2016; PAGES 2k Consortium, 2019; Büntgen et al.,
20   2020), with a smaller role for changes in greenhouse forcing, and solar forcing playing a minor role (Schurer
21   et al., 2014b; PAGES 2k Consortium, 2019).
23   Focusing now on warming over the historical period, AR5 assessed that it was extremely likely that human
24   influence was the dominant cause of the observed warming since the mid-20th century, and that it was
25   virtually certain that warming over the same period cannot be explained by internal variability alone. Since
26   AR5 many new attribution studies of changes in global surface temperature have focused on methodological
27   advances (see also Section 3.2). Those advances include better accounting for observational and model
28   uncertainties, and internal variability (Ribes and Terray, 2013; Hannart, 2016; Ribes et al., 2017; Schurer et
29   al., 2018); formulating the attribution problem in a counterfactual framework (Hannart and Naveau, 2018);
30   and reducing the dependence of the attribution on uncertainties in climate sensitivity and forcing (Otto et al.,
31   2015; Haustein et al., 2017, 2019). Studies now account for the uncertainties in the statistics of internal
32   variability, either explicitly (Hannart, 2016; Hannart and Naveau, 2018; Ribes et al., 2021) or implicitly
33   (Ribes and Terray, 2013; Schurer et al., 2018; Gillett et al., 2021), thus addressing concerns about over-
34   confident attribution conclusions. Accounting for observational uncertainty increases the range of warming
35   attributable to greenhouse gases by only 10 to 30% (Jones and Kennedy, 2017; Schurer et al., 2018). While
36   some attribution studies estimate attributable changes in globally-complete GSAT (Schurer et al., 2018;
37   Gillett et al., 2021; Ribes et al., 2021), others attribute changes in observational GMST, but this makes little
38   difference to attribution conclusions (Schurer et al., 2018). Moreover, based on a synthesis of observational
39   and modelling evidence, Cross-Chapter Box 2.3 assesses that the current best estimate of the scaling factor
40   between GMST and GSAT is one, and therefore attribution studies of GMST and GSAT are here treated
41   together in deriving assessed warming ranges. Studies also increasingly validate their multi-model
42   approaches using imperfect model tests (Schurer et al., 2018; Gillett et al., 2021; Ribes et al., 2021).
43   Alternative techniques, based purely on statistical or econometric approaches, without the need for climate
44   modelling, have also been applied (Estrada et al., 2013; Stern and Kaufmann, 2014; Dergiades et al., 2016)
45   and match the results of physically-based methods. The larger range of attribution techniques and
46   improvements to those techniques increase confidence in the results compared to AR5.
48   In contrast, studies published since the AR5 indicate that closely constraining the separate contributions of
49   greenhouse gas changes and aerosol changes to observed temperature changes remains challenging.
50   Attribution of warming to greenhouse gas forcing is made as early as the end of the 19th century (Schurer et
51   al., 2014b; Owens et al., 2017; PAGES 2k Consortium, 2019). Hegerl et al. (2019) found that volcanism
52   cooled global temperatures by about 0.1°C between 1870 and 1910, then a lack of volcanic activity warmed
53   temperatures by about 0.1°C between 1910 and 1950, with aerosols cooling temperatures throughout the 20th
54   century, especially between 1950 and 1980 when the estimated range of aerosol cooling was about 0.1 to
55   0.5°C. Jones et al. (2016a) attributed a warming of 0.87 to 1.22°C per century over the period 1906 to 2005
     Do Not Cite, Quote or Distribute                      3-19                                       Total pages: 202
     Final Government Distribution                          Chapter 3                                       IPCC AR6 WGI

 1   to greenhouse gases, partially offset by a cooling of −0.54 to −0.22°C per century attributed to aerosols. But
 2   they also found that detection of the greenhouse gas or the aerosol signal often fails, because of uncertainties
 3   in modelled patterns of change and internal variability. That point is illustrated by Figure 3.7, which shows
 4   two- and three-way fingerprinting regression coefficients for 13 CMIP6 models and the corresponding
 5   attributable warming ranges, derived using HadCRUT4 (Gillett et al., 2021). Regression coefficients with an
 6   uncertainty range that includes zero mean that detection has failed. Models with regression coefficients
 7   significantly less than 1 significantly overpredict the temperature response to the corresponding forcing.
 8   Conversely, models with regression coefficients significantly greater than 1 underpredict the response to
 9   these forcings. While estimates of warming attributable to anthropogenic influence derived using individual
10   models are generally consistent, estimates of warming attributable to greenhouse gases and aerosols
11   separately based on individual models are not all consistent, and detection of the aerosol influence fails more
12   often than that of greenhouse gases. Hence, results of recent studies emphasize the need to use multi-model
13   means to better constrain estimates of GSAT changes attributable to greenhouse gas and aerosol forcing
14   (Schurer et al., 2018; Gillett et al., 2021; Ribes et al., 2021).
19   Figure 3.7: Regression coefficients and corresponding attributable warming estimates for individual CMIP6
20               models. Upper panels show regression coefficients based on a two-way regression (left) and three-way
21               regression (right), of observed 5-yr mean global mean masked and blended surface temperature
22               (HadCRUT4) onto individual model response patterns, and a multi-model mean, labelled ‘Multi’.
23               Anthropogenic, natural, greenhouse gas, and other anthropogenic (aerosols, ozone, land-use change)
24               regression coefficients are shown. Regression coefficients are the scaling factors by which the model
25               responses must be multiplied to best match observations. Regression coefficients consistent with one
26               indicate a consistent magnitude response in observations and models, and regression coefficients
27               inconsistent with zero indicate a detectable response to the forcing concerned. Lower panels show
28               corresponding observationally-constrained estimates of attributable warming in globally-complete GSAT
29               for the period 2010-2019, relative to 1850-1900, and the horizontal black line shows an estimate of
30               observed warming in GSAT for this period. Figure is adapted from Gillett et al. (2021), their Extended
31               Data Figure 3. Further details on data sources and processing are available in the chapter data table (Table
32               3.SM.1).
34   [END FIGURE 3.7 HERE]
37   Figure 3.8 compares attributable trends in globally-complete GSAT for the period 2010-2019 compared to
38   1850-1900 from three detection and attribution studies, two of which use CMIP6 multi-model means (Gillett
39   et al., 2021; Ribes et al., 2021), and an estimate based on assessed effective radiative forcing and transient
40   and equilibrium climate sensitivity (see Section The reference period 1850-1900 is used to assess
41   attributable temperature changes because this is when the earliest gridded surface temperature records start,
42   this is when the CMIP6 historical simulations start, this is the earliest base period used in attribution
43   literature, and this is a reference period used in IPCC SR1.5 and earlier reports. It should however be noted
44   that Cross-Chapter Box 1.2 assesses with medium confidence that there was an anthropogenic warming with
45   a likely range of 0.0°C-0.2°C between 1750 and 1850-1900. Figure 3.8 also shows the GSAT changes
46   directly simulated in response to these forcings in thirteen CMIP6 models. In spite of their different
47   methodologies and input datasets, the three attribution approaches yield very similar results, with the
48   anthropogenic attributable warming range encompassing observed warming, and the natural attributable
49   warming being close to zero. The warming driven by greenhouse gas increases is offset in part by cooling
50   due to other anthropogenic forcing agents, mostly aerosols, although uncertainties in these contributions are
51   larger than the uncertainty in the net anthropogenic warming, as discussed above. Estimates based on
52   physical understanding of forcing and ECS made by Chapter 7 are close to estimates from attribution studies,
53   despite being the products of a different approach. This agreement enhances confidence in the magnitude and
54   causes of attributable surface temperature warming.

     Do Not Cite, Quote or Distribute                         3-20                                          Total pages: 202
     Final Government Distribution                          Chapter 3                                       IPCC AR6 WGI

 3   Figure 3.8: Assessed contributions to observed warming, and supporting lines of evidence. Shaded bands show
 4               assessed likely ranges of temperature change in GSAT, 2010-2019 relative to 1850-1900, attributable to
 5               net human influence, well-mixed greenhouse gases, other human forcings (aerosols, ozone, and land-use
 6               change), natural forcings, and internal variability, and the 5-95% range of observed warming. Bars show
 7               5-95% ranges based on (left to right) Haustein et al. (2017), Gillett et al. (2021) and Ribes et al. (2021),
 8               and crosses show the associated best estimates. No 5-95% ranges were provided for the Haustein et al.
 9               (2017) greenhouse gas or other human forcings contributions. The Ribes et al. (2021) results were
10               updated using a revised natural forcing time series, and the Haustein et al. (2017) results were updated
11               using HadCRUT5. The Chapter 7 best estimates and ranges are derived using assessed forcing time series
12               and a two-layer energy balance model as described in Section Coloured symbols show the
13               simulated responses to the forcings concerned in each of the models indicated. Further details on data
14               sources and processing are available in the chapter data table (Table 3.SM.1).
16   [END FIGURE 3.8 HERE]
19   The AR5 found high confidence for a major role for anthropogenic forcing in driving warming over each of
20   the inhabited continents, except for Africa where they found only medium confidence because of limited data
21   availability (Bindoff et al., 2013). At the hemispheric scale, Friedman et al. (2020) and Bonfils et al. (2020)
22   detected an anthropogenically forced response of inter-hemispheric contrast in surface temperature change,
23   which has a complex time evolution but showing the NH cooling more than the SH until around 1975 but
24   then warming after that. Bonfils et al. (2020) attribute the NH reversal to a combination of reduced aerosol
25   forcing and greenhouse gas induced warming of NH land masses. Friedman et al. (2020) found that CMIP5
26   models simulate the correct sign of the inter-hemispheric contrast when forced with all forcings but
27   underestimate its magnitude. Figure 3.9 shows global surface temperature change in CMIP6 all-forcing and
28   natural-only simulations globally, averaged over continents, and separately over land and ocean surfaces.
29   All-forcing simulations encompass observed temperature changes for all regions, while natural-only
30   simulations fail to do so in recent decades except in Antarctica, based on the annual means shown. As stated
31   above, warming results from a partial offset of greenhouse warming by aerosol cooling. That offset is
32   stronger over land than ocean. Regionally, models show a large range of possible temperature responses to
33   greenhouse gas and aerosol forcing, which complicates single-forcing attribution. A more detailed discussion
34   of regional attribution can be found in Section 10.4. Over global land surfaces, Chan and Wu (2015) used
35   CMIP5 simulations to attribute a warming trend of 0.3 (2.5%–97.5% confidence interval: 0.2–0.36) °C per
36   decade to anthropogenic forcing, with natural forcing only contributing 0.05 (0.02–0.06) °C per decade.
37   Accounting for unsampled sources of uncertainty and the availability of only a single study, their result
38   suggests that it is very likely that human influence is the main driver of warming over land.
43   Figure 3.9: Global, land, ocean and continental annual mean near-surface air temperatures anomalies in
44               CMIP6 models and observations. Time series are shown for CMIP6 historical anthropogenic and
45               natural (brown), natural-only (green), greenhouse gas only (grey) and aerosol only (blue) simulations
46               (multi-model means shown as thick lines, and shaded ranges between the 5th and 95th percentiles) and for
47               HadCRUT5 (black). All models have been subsampled using the HadCRUT5 observational data mask.
48               Temperatures anomalies are shown relative to 1950-2010 for Antarctica and relative to 1850–1900 for
49               other continents. CMIP6 historical simulations are expand by the SSP2-4.5 scenario simulations. All
50               available ensemble members were used (see Section 3.2). Regions are defined by Iturbide et al. (2020).
51               Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).
53   [END FIGURE 3.9 HERE]
56   In summary, since the publication of the AR5, new literature has emerged which better accounts for
57   methodological and climate model uncertainties in attribution studies (Ribes et al., 2017; Hannart and
     Do Not Cite, Quote or Distribute                          3-21                                         Total pages: 202
     Final Government Distribution                            Chapter 3                                    IPCC AR6 WGI

 1   Naveau, 2018) and concluded that anthropogenic warming is approximately equal to observed warming over
 2   the 1951-2010 period. The IPCC SR1.5 reached the same conclusion for 2017 relative to 1850-1900 based
 3   on anthropogenic warming and associated uncertainties calculated using the method of Haustein et al.
 4   (2017). Moreover, the improved understanding of the causes of the apparent slowdown in warming over the
 5   beginning of the 21st century and the difference in simulated and observed warming trends over this period
 6   (Cross-Chapter Box 3.1) further improve our confidence in the assessment of the dominant anthropogenic
 7   contribution to observed warming. In deriving our assessments, these considerations are balanced against
 8   new literature that raises questions about the ability of some models to simulate variability in surface
 9   temperatures over a range of time scales (Laepple and Huybers, 2014; Parsons et al., 2017; Friedman et al.,
10   2020), and the finding that some CMIP6 models exhibit substantially higher multidecadal internal variability
11   than that seen in CMIP5, which remains to be fully understood (Parsons et al., 2020; Ribes et al., 2021).
12   Further, uncertainties in simulated aerosol-cloud interactions are still large (Section, resulting in
13   very diverse spatial responses of different climate models to aerosol forcing and differences in the historical
14   global mean temperature evolution and in diagnosed cooling attributable to aerosol (Figure 3.8). Moreover,
15   like previous generations of coupled model simulations, historical and single forcing CMIP6 simulations
16   follow a common experimental design (Eyring et al., 2016a; Gillett et al., 2016b) and are thus all driven by
17   the same common set of forcings, even though these forcings are uncertain. Hence, forcing uncertainty is not
18   directly accounted for in most of the attribution and model evaluation studies assessed here, although this
19   limitation can to some extent be addressed by comparing with previous generation multi-model ensembles or
20   individual model studies using different sets of forcings.
22   The IPCC SR1.5 best estimate and likely range of anthropogenic attributable GMST warming was 1.0±0.2°C
23   in 2017 with respect to the period 1850-1900. Here, the best estimate is expressed in terms of GSAT and is
24   calculated as the average of the three estimates shown in Figure 3.9, yielding a value of 1.07°C. Ranges for
25   attributable GSAT warming are derived by finding the smallest ranges with a precision of 0.1°C which span
26   all of the 5-95% ranges from the attribution studies shown in Figure 3.9. These ranges are then assessed as
27   likely rather than very likely because the studies may underestimate the importance of the structural
28   limitations of climate models, which probably do not represent all possible sources of internal variability; use
29   too simple climate models, which may underestimate the role of internal variability; or underestimate model
30   uncertainty, especially when using model ensembles of limited size and inter-dependent models, for example
31   through common errors in forcings across models, as discussed above. This leads to a likely range for
32   anthropogenic attributable warming in 2010-2019 relative to 1850-1900 of 0.8 to 1.3°C in terms of GSAT.
33   Consequently, that range encompasses the best estimate and very likely range of observed GSAT warming of
34   1.06 °C [0.88 to 1.21 °C] over the same period (Cross-Chapter Box 2.3). There is medium confidence that the
35   best estimate and likely ranges of attributable warming expressed in terms of GMST are equal to those for
36   GSAT (Cross-Chapter Box 2.3). Repeating the process for other time periods leads to the best estimates and
37   likely ranges listed in Table 3.1. GSAT change attributable to natural forcings is −0.1 to 0.1°C. The likely
38   range of GSAT warming attributable to greenhouse gases is assessed in the same way to be 1.0 to 2.0°C
39   while the GSAT change attributable to aerosols, ozone and land-use change is −0.8 to 0.0°C. Progress in
40   attribution techniques allows the important advance of attributing observed surface temperature warming
41   since 1850-1900, instead of since 1951 as was done in the AR5.
46   Table 3.1: Estimates of warming in GSAT attributable to human influence for different periods in °C, all relative to
47              the 1850-1900 base period. Uncertainty ranges are 5-95% ranges for individual studies and likely ranges for
48              the assessment. The results shown in the table use the methods described in the three studies indicated, but
49              applied to additional periods and the warming trend. Ribes et al. (2021) results were updated using a
50              corrected natural forcing time series, and Haustein et al. (2017) results were updated to use HadCRUT5.
                             1986-2005         1995-2014               2006-2015       2010-2019           Warming rate
      Ribes et al. (2021)    0.65 (0.52 –      0.82 (0.69 –            0.94 (0.81 –    1.03 (0.89 –        0.23 (0.18 –
                             0.77)             0.94)                   1.08)           1.17)               0.29)

     Do Not Cite, Quote or Distribute                           3-22                                       Total pages: 202
     Final Government Distribution                      Chapter 3                                       IPCC AR6 WGI

      Gillett et al. (2021)   0.63 (0.32 –   0.84 (0.63 –         0.98 (0.74 –       1.11 (0.92 –       0.36 (0.30 –
                              0.94)          1.06)                1.22)              1.30)              0.41)
      Haustein et al.         0.73 (0.58 –   0.88 (0.75 –         0.98 (0.87 –       1.06 (0.94 –       0.23 (0.19 –
      (2017)                  0.82)          0.98)                1.10)              1.22)              0.35)
      Assessment              0.68 (0.3 –    0.85 (0.6 – 1.1)     0.97 (0.7 – 1.3)   1.07 (0.8 – 1.3)   0.2 (0.1 – 0.3)
 2   [END TABLE 3.1 HERE]
 5   The IPCC AR5 assessed the likely range of the contribution of internal variability to GMST warming to be
 6   −0.1 to 0.1°C over the period 1951-2010. Since then, several studies have downplayed the contribution of
 7   internal modes of variability to global temperature variability, often by arguing for a forced component to
 8   those internal modes (Mann et al., 2014; Folland et al., 2018; Haustein et al., 2019; Liguori et al., 2020).
 9   Haustein et al. (2017) found a 5-95% confidence interval of −0.09 to +0.12 °C for the contribution of internal
10   variability to warming between 1850–1879 and 2017. Ribes et al. (2021) imply a contribution of internal
11   variability of −0.02 ± 0.16°C to warming between 2010-2019 and 1850-1900, assuming independence
12   between errors in the observations and in the estimate of the forced response. Based on these studies, but
13   allowing for unsampled sources of error, we assess the likely range of the contribution of internal variability
14   to GSAT warming between 2010-2019 and 1850-1900 to be −0.2 to 0.2°C.
16   The IPCC SR1.5 gave a likely range for the human-induced warming rate of 0.1°C to 0.3°C per decade in
17   2017, with a best estimate of 0.2°C per decade (Allen et al., 2018). Table 3.1 lists the estimates of
18   attributable anthropogenic warming rate over the period 2010-2019 by the three studies that underpin the
19   assessment of GSAT warming (Haustein et al., 2017; Gillett et al., 2021; Ribes et al., 2021). Estimates from
20   Haustein et al. (2017), based on observed warming, and Ribes et al. (2021), based on CMIP6 simulations
21   constrained by observed warming, are in good agreement. Gillett et al. (2021), also based on CMIP6 models,
22   corresponds to a larger anthropogenic attributable warming rate, because of a smaller warming rate attributed
23   to natural forcing than in Ribes et al. (2021). This disagreement does not support a decrease in uncertainty
24   compared to the SR1.5 assessment. So the range for anthropogenic attributable surface temperature warming
25   rate of 0.1°C to 0.3°C per decade is again assessed to be likely, with a best estimate of 0.2°C per decade.
28 Upper-Air Temperature
30   Chapter 2 assessed that the troposphere has warmed since at least the 1950s, that it is virtually certain that
31   the stratosphere has cooled, and that there is medium confidence that the upper troposphere in the tropics has
32   warmed faster than the near-surface since at least 2001 (Section The AR5 assessed that
33   anthropogenic forcings, dominated by greenhouse gases, likely contributed to the warming of the troposphere
34   since 1961 and that anthropogenic forcings, dominated by the depletion of the ozone layer due to ozone-
35   depleting substances, very likely contributed to the cooling of the lower stratosphere since 1979. Since the
36   AR5, understanding of observational uncertainties in the radiosonde and satellite data has improved with
37   more available data and longer coverage, and differences between models and observations in the tropical
38   atmosphere have been investigated further.
40   Tropospheric temperature
41   The AR5 assessed with low confidence that most, though not all, CMIP3 (Meehl et al., 2007) and CMIP5
42   (Taylor et al., 2012) models overestimated the observed warming trend in the tropical troposphere during the
43   satellite period 1979-2012, and that a third to a half of this difference was due to an overestimate of the SST
44   trend during this period (Flato et al., 2013). Since the AR5, additional studies based on CMIP5 and CMIP6
45   models show that this warming bias in tropospheric temperatures remains. Recent studies have investigated
46   the role of observational uncertainty, the model response to external forcings, the influence of the time
47   period considered, and the role of biases in SST trends in contributing to this bias.
49   Several studies since AR5 have continued to demonstrate an inconsistency between simulated and observed

     Do Not Cite, Quote or Distribute                      3-23                                         Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 1   temperature trends in the tropical troposphere, with models simulating more warming than observations
 2   (Mitchell et al., 2013, 2020, Santer et al., 2017a, 2017b; McKitrick and Christy, 2018; Po-Chedley et al.,
 3   2021). Santer et al. (2017b) used updated and improved satellite retrievals to investigate model performance
 4   in simulating the tropical mid- to upper-troposphere trends, and removed the influence of stratospheric
 5   cooling by regression. These factors were found to reduce the size of the discrepancy in mid- to upper-
 6   tropospheric temperature trends between models and observations over the satellite era, but a discrepancy
 7   remained. Santer et al. (2017a) found that during the late 20th century, the discrepancies between simulated
 8   and satellite-derived mid- to upper-tropospheric temperature trends were consistent with internal variability,
 9   while during most of the early 21st century, simulated tropospheric warming is significantly larger than
10   observed, which they relate to systematic deficiencies in some of the external forcings used after year 2000
11   in the CMIP5 models. However, in CMIP6, differences between simulated and observed upper tropospheric
12   temperature trends persist despite updated forcing estimates (Mitchell et al., 2020). Figure 3.10 shows that
13   CMIP6 models forced by combined anthropogenic and natural forcings overestimate temperature trends
14   compared to radiosonde data (Haimberger et al., 2012) throughout the tropical troposphere (Mitchell et al.,
15   2020). Over the 1979-2014 period, models are more consistent with observations in the lower troposphere,
16   and least consistent in the upper troposphere around 200 hPa, where biases exceed 0.1°C per decade. Several
17   studies using CMIP6 models suggest that differences in climate sensitivity may be an important factor
18   contributing to the discrepancy between the simulated and observed tropospheric temperature trends
19   (McKitrick and Christy, 2020; Po-Chedley et al., 2021), though it is difficult to deconvolve the influence of
20   climate sensitivity, changes in aerosol forcing and internal variability in contributing to tropospheric
21   warming biases (Po-Chedley et al., 2021). Another study found that the absence of a hypothesized negative
22   tropical cloud feedback could explain half of the upper troposphere warming bias in one model (Mauritsen
23   and Stevens, 2015).
25   Mitchell et al. (2013) and Mitchell et al. (2020) found a smaller discrepancy in tropical tropospheric
26   temperature trends in models forced with observed SSTs (see also Figure 3.10a), and CMIP5 models and
27   observations were found to be consistent below 150 hPa when viewed in terms of the ratio of temperature
28   trends aloft to those at the surface (Mitchell et al., 2013). Flannaghan et al. (2014) and Tuel (2019) showed
29   that most of the tropospheric temperature difference between CMIP5 models and the satellite-based trend
30   over the 1970-2018 period is due to respective differences in SST warming trends in regions of deep
31   convection, and Po-Chedley et al. (2021) show that CMIP6 models with a more realistic SST simulation in
32   the central and eastern Pacific show a better performance than other models. Though systematic biases still
33   remain, this indicates that the bias in tropospheric temperature warming in models is in part linked to surface
34   temperature warming biases, especially in the lower troposphere.
36   In summary, studies continue to find that CMIP5 and CMIP6 model simulations warm more than
37   observations in the tropical mid- and upper-troposphere over the 1979-2014 period (Mitchell et al., 2013,
38   2020, Santer et al., 2017a, 2017b; Suárez-Gutiérrez et al., 2017; McKitrick and Christy, 2018), and that
39   overestimated surface warming is partially responsible (Mitchell et al., 2013; Po-Chedley et al., 2021). Some
40   studies point to forcing errors in the CMIP5 simulations in the early 21st century as a possible contributor
41   (Mitchell et al., 2013; Sherwood and Nishant, 2015; Santer et al., 2017a), but CMIP6 simulations use
42   updated forcing estimates yet generally still warm more than observations. Although accounting for internal
43   variability and residual observational errors can reconcile models with observations to some extent (Suárez-
44   Gutiérrez et al., 2017; Mitchell et al., 2020), some studies suggest that climate sensitivity also plays a role
45   (Mauritsen and Stevens, 2015; McKitrick and Christy, 2020; Po-Chedley et al., 2021). Hence, we assess with
46   medium confidence that CMIP5 and CMIP6 models continue to overestimate observed warming in the upper
47   tropical troposphere over the 1979-2014 period by at least 0.1°C per decade, in part because of an
48   overestimate of the tropical SST trend pattern over this period.
53   Figure 3.10: Observed and simulated tropical mean temperature trends through the atmosphere. Vertical
54                profiles of temperature trends in the tropics (20°S-20°N) for three periods: (a) 1979-2014 (b) 1979-1997
55                (ozone depletion era) (c) 1998-2014 (ozone stabilisation era). The black lines show trends in the RICH

     Do Not Cite, Quote or Distribute                         3-24                                         Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1                1.7 (long dashed) and RAOBCORE 1.7 (dashed) radiosonde datasets (Haimberger et al., 2012), and in the
 2                ERA5/5.1 reanalysis (solid). Grey envelopes are centred on the RICH 1.7 trends, but show the uncertainty
 3                based on 32 RICH-obs members of version 1.5.1 of the dataset, which used version 1.7.3 of the RICH
 4                software but with the parameters of version 1.5.1. ERA5 was used as reference for calculating the
 5                adjustments between 2010 and 2019, and ERA-Interim was used for the years before that. Red lines show
 6                trends in CMIP6 historical simulations from one realization of 60 models. Blue lines show trends in 46
 7                CMIP6 models that used prescribed, rather than simulated, sea surface temperatures (SSTs). Figure is
 8                adapted from Mitchell et al. (2020), their Figure 1. Further details on data sources and processing are
 9                available in the chapter data table (Table 3.SM.1).
11   [END FIGURE 3.10 HERE]
14   The AR5 assessed as likely that anthropogenic forcings, dominated by greenhouse gases, contributed to the
15   warming of the troposphere since 1961 (Bindoff et al., 2013). Since then, there has been further progress in
16   detecting and attributing tropospheric temperature changes. Mitchell et al. (2020) used CMIP6 models to
17   find that the main driver of tropospheric temperature changes are greenhouse gases. Previous detection of the
18   anthropogenic influence on tropospheric warming may have overestimated uncertainties: Pallotta and Santer
19   (2020) found that CMIP5 climate models overestimate the observed natural variability in global mean
20   tropospheric temperature on timescales of 5-20 years. Nevertheless, Santer et al. (2019) found that stochastic
21   uncertainty is greater for tropospheric warming (8-15 years) than stratospheric cooling (1-3 years) because of
22   larger noise and slower recovery time from the Pinatubo eruption in the troposphere. The detection time of
23   the anthropogenic signal in the tropospheric warming can be affected by both the model climate sensitivity
24   and the model response to aerosol forcing. Volcanic forcing is also important, as models that do not consider
25   the influence of volcanic eruptions in the early 21st century overestimate the observed tropospheric warming
26   since 1998 (Santer et al., 2014). Changes in the amplitude of the seasonal cycle of tropospheric temperatures
27   have also been attributed to human influences. Santer et al. (2018) found that satellite data and climate
28   models driven by anthropogenic forcing show consistent amplitude increases at mid-latitudes in both
29   hemispheres, amplitude decreases at high latitudes in the SH, and small changes in the tropics.
31   In summary, these studies confirm the dominant role of human activities in tropospheric temperature trends.
32   We therefore assess that it is very likely that anthropogenic forcing, dominated by greenhouse gases, was the
33   main driver of the warming of the troposphere since 1979.
35   Stratospheric temperature
36   The AR5 concluded that the CMIP5 models simulated a generally realistic evolution of lower stratospheric
37   temperatures (Bindoff et al., 2013; Flato et al., 2013), which was better than the CMIP3 models, in part
38   because they generally include time-varying ozone concentrations, unlike many of the CMIP3 models.
39   Nonetheless, it was noted that there was a tendency for the simulations to underestimate stratospheric
40   cooling compared to observations. Bindoff et al. (2013) concluded that it was very likely that anthropogenic
41   forcing, dominated by stratospheric ozone depletion by chemical reactions involving trace species known as
42   ozone-depleting substances (ODS), had contributed to the cooling of the lower stratosphere since 1979.
43   Increased greenhouse gases cause near-surface warming but cooling of stratospheric temperatures.
45   For the lower stratosphere, a debate has been ongoing since the AR5 between studies finding that models
46   underestimate the cooling of stratospheric temperature (Santer et al., 2017b), in part because of
47   underestimated stratospheric ozone depletion (Eyring et al., 2013; Young et al., 2013), and studies finding
48   that lower stratospheric temperature trends are within the range of observed trends (Young et al., 2013;
49   Maycock et al., 2018). Different observational data and different time periods explain the different
50   conclusions. Aquila et al. (2016) used forced chemistry-climate models with prescribed SST to investigate
51   the influence of different forcings on global stratospheric temperature changes. They found that in the lower
52   stratosphere, the simulated cooling trend due to increasing greenhouse gases was roughly constant over the
53   satellite era, while changes in ODS concentrations amplified that stratospheric cooling trend during the era of
54   increasing ozone depletion up until the mid-1990s, with a flattening of the temperature trend over the
55   subsequent period over which stratospheric ozone has stabilised (Section Mitchell et al. (2020)
56   showed that while models simulate realistic trends in tropical lower stratospheric temperature over the whole
     Do Not Cite, Quote or Distribute                       3-25                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   1979-2014 period when compared with radiosonde data, they tend to overestimate the cooling trend over the
 2   ozone depletion era (1979-1997) and underestimate it over the ozone recovery era (1998-2014) (Figure 3.10b
 3   and 3.10c). They speculate that those disagreements are due to poor representations of stratospheric ozone
 4   forcing.
 6   Upper stratospheric temperature changes were not assessed in the context of attribution or model evaluation
 7   in AR5, but this is an area where there has been considerable progress over recent years (see Section
 8 Simulated temperature changes in chemistry-climate models show good consistency with the
 9   reprocessed dataset from NOAA STAR but are less consistent with the revised UK Met Office record
10   (Karpechko et al., 2018b). The latter still shows stronger cooling than simulated in chemistry-climate models
11   (Maycock et al., 2018). Reanalyses, which assimilate AMSU and SSU datasets, indicate an upper-
12   stratospheric cooling from 1979 to 2009 of about 3°C at 5 hPa and 4°C at 1 hPa that agrees well with the
13   cooling in simulations with prescribed SST and using CMIP5 forcings (Simmons et al., 2014). Mitchell
14   (2016) used regularized optimal fingerprinting techniques to carry out an attribution analysis of annual mid
15   to upper stratospheric temperature in response to external forcings. They found that anthropogenic forcing
16   has caused a cooling of approximately 2-3°C in the upper stratosphere over the period of 1979-2015, with
17   greenhouse gases contributing two thirds of this change and ozone depletion contributing one third. They
18   find a large upper stratospheric temperature change in response to volcanic forcing (0.4-0.6 °C for Mount
19   Pinatubo) but that change is still smaller than the lower-stratospheric signal. Aquila et al. (2016) found that
20   the cooling of the middle and upper stratosphere after 1979 is mainly due to changes in greenhouse gas
21   concentrations. Volcanic eruptions and the solar cycle were not found to affect long-term stratospheric
22   temperature trends but to have short-term influences.
24   In summary, based on the latest updates to satellite observations of stratospheric temperature, we assess that
25   simulated and observed trends in global mean temperature through the depth of the stratosphere are more
26   consistent than based on previous datasets, but some differences remain (medium confidence). Studies
27   published since the AR5 increase our confidence in the simulated stratospheric temperature response to
28   greenhouse gas and ozone changes, and support an assessment that it is extremely likely that stratospheric
29   ozone depletion due to ozone-depleting substances was the main driver of the cooling of the lower
30   stratosphere between 1979 and the mid-1990s, as expected from physical understanding. Similarly, revised
31   observations and new studies support an assessment that it is extremely likely that anthropogenic forcing,
32   both from increases in greenhouse gas concentration and depletion of stratospheric ozone due to ozone-
33   depleting substances, was the main driver of upper stratospheric cooling since 1979.
38   Cross-Chapter Box 3.1:          Global Surface Warming over the Early 21st Century
40   Contributors: Christophe Cassou (France), Yu Kosaka (Japan), John Fyfe (Canada), Nathan Gillett (Canada),
41   Edward Hawkins (UK), Blair Trewin (Australia)
43   AR5 found that the rate of global mean surface temperature (GMST) increase inferred from observations
44   over the 1998-2012 period was lower than the rate of increase over the 1951-2012 period, and lower than the
45   ensemble mean increase in historical simulations from CMIP5 climate models extended by RCP scenario
46   simulations beyond 2005 (Flato et al., 2013). This apparent slowdown of surface global warming compared
47   to the 62-year rate was assessed with medium confidence to have been caused in roughly equal measure by a
48   cooling contribution from internal variability and a reduced trend in external forcing (particularly associated
49   with solar and volcanic forcing) in the AR5 based on expert judgement (Flato et al., 2013). In AR5 it was
50   assessed that almost all CMIP5 simulations did not reproduce the observed slower warming, and that there
51   was medium confidence that the trend difference from the CMIP5 ensemble mean was to a substantial degree
52   caused by internal variability with possible contributions from forcing error and model response uncertainty.
53   This Cross-Chapter Box assesses new findings from observational products and statistical and physical
54   models on trends over the 1998-2012 period considered in AR5.
     Do Not Cite, Quote or Distribute                     3-26                                      Total pages: 202
     Final Government Distribution                     Chapter 3                                    IPCC AR6 WGI

 1   Updated observational and reanalyses data sets and comparison with model simulations
 2   Since the AR5, there have been version updates and new releases of most observational GMST data sets
 3   (Cross-Chapter Box 2.3). All the updated products now available consistently find stronger positive trends
 4   for 1998-2012 than those assessed in AR5 (Cowtan and Way, 2014; Karl et al., 2015; Hausfather et al.,
 5   2017; Medhaug et al., 2017; Simmons et al., 2017; Risbey et al., 2018). Simmons et al. (2017) reported that
 6   the 1998-2012 GMST trends in the updated observational and reanalysis data sets available at that time range
 7   from 0.06 °C to 0.14°C per decade, compared with the 0.05 °C per decade on average as reported in AR5,
 8   whilst the latest data products reported in Chapter 2 Table 2.4 show GMST or global mean near-surface air
 9   temperature (GSAT) trends over that period ranging from 0.12°C to 0.14°C per decade. The lowest trend in
10   Simmons et al. (2017) is from HadCRUT4, now superseded by HadCRUT5, which shows a trend of 0.12 ºC
11   per decade. The upward revision is mainly due to improved sea surface temperature (SST) data sets and
12   infilling of surface temperature in locations with missing records in observational products, mainly in the
13   Arctic (see Cross-Chapter Box 2.3 for details).
15   With these updates, all the observed trends assessed here lie within the 10th-90th percentile range of the
16   simulated trends in the CMIP5 and CMIP6 simulations (Cross-Chapter Box 3.1, Figure 1a). This result is
17   insensitive to whether model GSAT (based on surface air temperature) or GMST (based on a blend of
18   surface air temperature over land and sea ice and SST over open ocean) is used, and to whether or not
19   masking with the observational data coverage is applied. Therefore, the observed 1998-2012 trend is
20   consistent with both the CMIP5 or CMIP6 multi-model ensemble of trends over the same period (high
21   confidence).
23   Internal variability
24   All the observation-based GMST and GSAT trends are lower than the multi-model mean GMST and GSAT
25   trends of both CMIP5 and CMIP6 for 1998-2012 (Cross-Chapter Box 3.1, Figure 1a). This suggests a
26   possible cooling contribution from internal variability during this period. This is supported by initialized
27   decadal hindcasts, which account for the phase of the multidecadal modes of variability (Sections 3.7.6 and
28   3.7.7), and which better reproduce observed global mean SST and GSAT trends than uninitialized historical
29   simulations (Guemas et al., 2013; Meehl et al., 2014).
31   Studies since AR5 identify Pacific Decadal Variability (PDV) as the leading mode of variability associated
32   with unforced decadal GSAT fluctuations, with additional influence from Atlantic Multidecadal Variability
33   (Brown et al., 2015; Dai et al., 2015; Steinman et al., 2015; Pasini et al., 2017; Annex IV.2.6, IV.2.7). PDV
34   transitioned from positive (El Niño-like) to negative (La Niña-like) phases during the slow warming period
35   (Figure 3.39f, Cross-Chapter Box 3.1, Figure 1c). Model ensemble members that capture the observed
36   slower decadal warming under transient forcing, and time segments of model simulations that show decadal
37   GSAT decreases under fixed radiative forcing, also feature negative PDV trends (Maher et al., 2014; Meehl
38   et al., 2011, 2013, 2014; Middlemas and Clement, 2016; Cross-Chapter Box 3.1, Figure 1d), suggesting the
39   influence of PDV. This is confirmed by statistical models with the PDV-GSAT relationship estimated from
40   observations and model simulations (Schmidt et al., 2014; Meehl et al., 2016b; Hu and Fedorov, 2017),
41   selected ensemble members and time segments from model simulations where PDV by chance evolves in
42   phase with observations over the slow warming period (Huber and Knutti, 2014; Risbey et al., 2014), and
43   coupled model experiments in which PDV evolution is constrained to follow the observations (Kosaka and
44   Xie, 2013, 2016; England et al., 2014; Watanabe et al., 2014; Delworth et al., 2015). Part of the PDV trend
45   may be driven by anthropogenic aerosols (Smith et al., 2016); however, this result is model-dependent, and
46   internally-driven PDV dominates the forced PDV signal in the CMIP6 multi-model ensemble (Section
47   3.7.6). It is also notable that there is large uncertainty in the magnitude of the PDV influence on GSAT
48   across models (Deser et al., 2017a; Wang et al., 2017a) and among the studies cited above. In addition to
49   PDV, contributions to the reduced warming trend from wintertime Northern Hemisphere atmospheric
50   internal variability, particularly associated with a trend towards the negative phase of the Northern Annular
51   Mode/North Atlantic Oscillation (Annex IV.2.1; Guan et al., 2015; Saffioti et al., 2015; Iles and Hegerl,
52   2017) or the Cold Ocean-Warm Land pattern (Molteni et al., 2017; Yang et al., 2020) have been suggested,
53   leading to regional continental cooling over a large part of Eurasia and North America (Li et al., 2015; Deser
54   et al., 2017; Gan et al., 2019; Cross-Chapter Box 3.1, Figure 1c).
     Do Not Cite, Quote or Distribute                     3-27                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                     IPCC AR6 WGI

 1   Such internally-driven variation of decadal GSAT trends is not unique to the 1998-2012 period (Section
 2; Lovejoy, 2014; Roberts et al., 2015; Dai and Bloecker, 2019). Due to the nature of internal
 3   variability, surface temperature changes over the 1998-2012 period are regionally- and seasonally-varying
 4   (Trenberth et al., 2014; Zang et al., 2019; Cross-Chapter Box 3.1, Figure 1c). Further, there was no
 5   slowdown in the increasing occurrence of hot extremes over land (Seneviratne et al., 2014). Thus, the
 6   internally-driven slowdown of GSAT increase does not correspond to slowdown of warming everywhere on
 7   the Earth’s surface.
 9   Updated forcing
10   CMIP5 historical simulations driven by observed forcing variations ended in 2005 and were extended with
11   RCP scenario simulations for model-observation comparisons beyond that date. Post AR5 studies based on
12   updated external forcing show that while no net effect of updated anthropogenic aerosols is found on GSAT
13   trends (Murphy, 2013; Gettelman et al., 2015; Oudar et al., 2018), natural forcing by moderate volcanic
14   eruptions in the 21st century (Haywood et al., 2014; Ridley et al., 2014; Santer et al., 2014) and a prolonged
15   solar irradiance minimum around 2009 compared to the normal 11-year cycle (Lean, 2018) yield a negative
16   contribution to radiative forcing, which was missing in CMIP5 (Figure 2.2). This explains part of the
17   difference between observed and CMIP5 trends, as shown based on EMIC simulations (Huber and Knutti,
18   2014; Ridley et al., 2014), statistical and mathematical models (Schmidt et al., 2014; Lean, 2018), and
19   process-based climate models (Santer et al., 2014). However, in a single climate model study by Thorne et
20   al. (2015), updating most forcings (greenhouse gas concentrations, solar irradiance, and volcanic and
21   anthropogenic aerosols) available when the study was done makes no significant difference to the 1998-2012
22   GMST trend from that obtained with original CMIP5 forcing. Potential underestimation of volcanic
23   (negative) forcing may have played a role (Outten et al., 2015). In the multi-model ensemble mean, the
24   1998-2012 GMST trends are almost equal in CMIP5 and CMIP6 (Cross-Chapter Box 3.1, Figure 1a),
25   suggesting compensation by a higher transient climate response and equilibrium climate sensitivity in
26   CMIP6 than CMIP5 (Section 7.5.6). To summarize, while there is medium confidence that natural forcing
27   that was missing in CMIP5 contributed to the difference of observed and simulated GMST trends,
28   confidence remains low in the quantitative contribution of net forcing updates.
30   Energy budget and heat redistribution
31   The early 21st century slower warming was observed in atmospheric temperatures, but the heat capacity of
32   the atmosphere is very small compared to that of the ocean. Although there is noticeable uncertainty among
33   observational products (Su et al., 2017a) and observation quality changes through time, global ocean heat
34   content continued to increase during the slower surface warming period (very high confidence), at a rate
35   consistent with CMIP5 and CMIP6 historical simulations (Sections, and There is
36   high confidence that the Earth’s energy imbalance was larger in the 2000s than in the 1985-1999 period
37   (Section, consistent with accelerating ocean heat uptake in the past two decades (Section
38   Internal decadal variability is mainly associated with redistribution of heat within the climate system (Yan et
39   al., 2016c; Drijfhout, 2018) while associated top of the atmosphere radiation anomalies are weak (Palmer
40   and McNeall, 2014). Heat redistribution in the top 350 m of the Indian and Pacific Oceans has been found to
41   be the main contributor to reduced surface warming during the slower surface warming period (Lee et al.,
42   2015; Nieves et al., 2015; Liu et al., 2016a), consistent with the simulated signature of PDV (England et al.,
43   2014; Maher et al., 2018a; Gastineau et al., 2019). Below 700 m, enhanced heat uptake over the slower
44   surface warming period is observed mainly in the North Atlantic and Southern Ocean (Chen and Tung,
45   2014), though whether this is a response to forcing or a unique signature of the slow GMST warming has
46   been questioned (Liu et al., 2016b).
48   Summary and implications
49   With updated observation-based GMST data sets and forcing, improved analysis methods, new modelling
50   evidence and deeper understanding of mechanisms, there is very high confidence that the slower GMST and
51   GSAT increase inferred from observations in the 1998-2012 period was a temporary event induced by
52   internal and naturally-forced variability that partly offset the anthropogenic warming trend over this period.
53   Global ocean heat content continued to increase throughout this period, and the slowdown was only evident
54   in the atmosphere and at the surface (very high confidence). Considering all the sources of uncertainties, it is
55   impossible to robustly identify a single cause of the early 2000s slowdown (Hedemann et al., 2017; Power et
     Do Not Cite, Quote or Distribute                      3-28                                      Total pages: 202
     Final Government Distribution                            Chapter 3                                         IPCC AR6 WGI

 1   al., 2017); rather, it should be interpreted as due to a combination of several factors (Huber and Knutti, 2014;
 2   Schmidt et al., 2014; Medhaug et al., 2017).
 4   A major El Niño event in 2014-2016 led to three consecutive years of record annual GMST with unusually
 5   strong heat release from the Northwestern Pacific Ocean (Yin et al., 2018), which marked the end of the
 6   slower warming period (Hu and Fedorov, 2017; Su et al., 2017b; Cha et al., 2018). The past 5-year period
 7   (2016-2020) is the hottest 5-year period in the instrumental record up to 2020 (high confidence). This rapid
 8   warming was accompanied by a PDV shift toward its positive phase (Su et al., 2017b; Cha et al., 2018). A
 9   higher rate of warming following the 1998-2012 period is consistent with the predictions in AR5 Box 9.2
10   (Flato et al., 2013) and with a statistical prediction system (Sévellec and Drijfhout, 2018). Initialized decadal
11   predictions show higher GMST trends in the early 2020s compared to uninitialized simulations (Thoma et
12   al., 2015; Meehl et al., 2016c).
14   While some recent studies find that internal decadal GSAT variability may become weaker under GSAT
15   warming, associated in part with reduced amplitude PDV (Section; Brown et al., 2017), the
16   weakening is small under a realistic range of warming. A large volcanic eruption would temporarily cool
17   GSAT (Cross-Chapter Box 4.1). Thus, there is very high confidence that reduced and increased GMST and
18   GSAT trends at decadal timescales will continue to occur in the 21st century (Meehl et al., 2013; Roberts et
19   al., 2015; Medhaug and Drange, 2016). However, such internal or volcanically forced decadal variations in
20   GSAT trend have little affect on the centennial warming (England et al., 2015; Cross-Chapter Box 4.1).
25   Cross-Chapter Box 3.1, Figure 1: 15-year trends of surface global warming for 1998-2012 and 2012-2026. (a, b)
26   GSAT and GMST trends for 1998-2012 (a) and 2012-2026 (b). Histograms are based on GSAT in historical
27   simulations of CMIP6 (red shading, extended by SSP2-4.5) and CMIP5 (grey shading; extended by RCP4.5). Filled and
28   open diamonds at the top represent multi-model ensemble means of GSAT and GMST trends, respectively. Diagonal
29   lines show histograms of HadCRUT5.0.1.0. Triangles at the top of (a) represent GMST trends of Berkeley Earth,
30   GISTEMP, Kadow et al. (2020) and NOAAGlobalTemp-Interim, and the GSAT trend of ERA5. Selected CMIP6
31   members whose 1998-2012 trends are lower than the HadCRUT5.0.1.0 mean trend are indicated by purple shading (a)
32   and (b). In (a), model GMST and GSAT, and ERA5 GSAT are masked to match HadCRUT data coverage. (c-d) Trend
33   maps of annual near-surface temperature for 1998-2012 based on HadCRUT5.0.1.0 mean (c) and composited surface
34   air temperature trends of subsampled CMIP6 simulations (d) that are included in purple shading area in (a). In (c), cross
35   marks indicate trends that are not significant at the 10% level based on t-tests with serial correlation taken into account.
36   Ensemble size used for each of the histograms and the trend composite is indicated at the top right of each of panels
37   (a,b,d). Model ensemble members are weighted with the inverse of the ensemble size of the same model, so that
38   individual models are equally weighted. Further details on data sources and processing are available in the chapter data
39   table (Table 3.SM.1).
46   3.3.2    Precipitation, Humidity and Streamflow
48   Paleoclimate context
49   A fact hindering detection and attribution studies in precipitation and other hydrological variables is the large
50   internal variability of these fields relative to the anthropogenic signal. This low signal-to-noise ratio hinders
51   the emergence of the anthropogenic signal from natural variability. Moreover, the sign of the change
52   depends on location and time of the year. Paleoclimate records provide valuable context for observed trends
53   in the 20th and 21st century and assist with the attribution of these trends to human influence (see also Section
54 By nature, hydrological proxy data represent regional conditions, but taken together can represent
55   large-scale patterns. As an example of how paleorecords have helped assessing the origin of changes, we
56   consider some, mainly subtropical, regions which have experienced systematic drying in recent decades (see

     Do Not Cite, Quote or Distribute                           3-29                                           Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   also Section Paleoclimate simulations of monsoons are assessed in Section
 3   Records of tree ring width have provided evidence that recent prolonged dry spells in the Levant and Chile
 4   are unprecedented in the last millennium (high confidence) (Cook et al., 2016a; Garreaud et al., 2017). East
 5   Africa, has also been drying in recent decades (Rowell et al., 2015; Hoell et al., 2017a), a trend that is
 6   unusual in the context of the sedimentary paleorecord spanning the last millennium (Tierney et al., 2015).
 7   This may be a signature of anthropogenic forcing but cannot as of yet be distinguished from natural
 8   variability (Hoell et al., 2017b; Philip et al., 2018). Likewise, tree rings indicate that the 2012-2014 drought
 9   in the southwestern United States was exceptionally severe in the context of natural variability in the last
10   millennium, and may have been exacerbated by the contribution of anthropogenic temperature rise (medium
11   confidence) (Griffin and Anchukaitis, 2014; Williams et al., 2015). Furthermore, Williams et al. (2020) used
12   a combination of hydrological modelling and tree-ring reconstructions to show that the period from 2000 to
13   2018 was the driest 19-year span in southwestern North America since the late 1500s. Nonetheless, tree rings
14   also indicate the presence of prolonged megadroughts in western North America throughout the last
15   millennium that were more severe than 20th and 21st century events (high confidence) (Cook et al., 2004,
16   2010, 2015). These were associated with internal variability (Coats et al., 2016; Cook et al., 2016b) and
17   indicate that large-magnitude changes in the water cycle may occur irrespective of anthropogenic influence
18   (see also McKitrick and Christy, 2019).
20   Paleoclimate records also allow for model evaluation under conditions different from present-day. AR5
21   concluded that models can successfully reproduce to first-order patterns of past precipitation changes during
22   the Last Glacial Maximum (LGM) and mid-Holocene, though simulated precipitation changes during the
23   mid-Holocene tended to be underestimated (Flato et al., 2013). Further analysis of CMIP5 models confirmed
24   these results but has also revealed systematic offsets from the paleoclimate record (DiNezio and Tierney,
25   2013a; Hargreaves and Annan, 2014; Harrison et al., 2014, 2015; Bartlein et al., 2017; Scheff et al., 2017;
26   Tierney et al., 2017a). Harrison et al. (2014) concluded that CMIP5 models do not perform better in
27   simulating rainfall during the LGM and mid-Holocene than earlier model versions despite higher resolution
28   and complexity. However, prescribing changes in vegetation and dust was found to improve the match to the
29   paleoclimate record (Pausata et al., 2016; Tierney et al., 2017b) suggesting that vegetation feedbacks in the
30   CMIP5 models may be too weak (low confidence) (Hopcroft et al., 2017). Brierley et al. (2020) compared
31   the latitudinal gradient of annual precipitation changes in the European-African sector simulated by CMIP6
32   models for the mid-Holocene with pollen-based reconstructions and showed that models generally reproduce
33   the direction of changes seen in the reconstructions (Figure 3.11). They do not show a robust signal in area
34   averaged rainfall over most European regions where quantitative reconstructions exist, which is not
35   incompatible with reconstructions. Over the Sahara/Sahel and West Africa regions, where reconstructions
36   suggest positive anomalies during the mid-Holocene, both CMIP5 and CMIP6 models also simulate a
37   rainfall increase, but it is much weaker (see also Section Overall, however, large discrepancies
38   remain between simulations and reconstructions.
40   Liu et al. (2018) evaluated the terrestrial moisture changes that occurred during the LGM and concluded that
41   the multi-model median from CMIP5 is consistent with available paleo-records in some regions, but not in
42   others. CMIP5 models accurately reproduce an increase in moisture in the western United States, related to
43   an intensified winter storm track and decreased evaporative demand (Oster et al., 2015; Ibarra et al., 2018;
44   Lora, 2018). On the other hand, CMIP5 models show a wide variety of responses in the tropical Indo-Pacific
45   region, with only a few matching the pattern of change inferred from the paleoclimate record (DiNezio and
46   Tierney, 2013b; DiNezio et al., 2018). The variable response across models is related to the effect of the
47   exposure of the tropical shelves during glacial times, which variously intensifies or weakens convection in
48   the rising branch of the Walker cell, depending on model parameterization (DiNezio et al., 2011). For the
49   Last Interglacial, CMIP6 models reproduce the proxy-based increased precipitation relative to pre-industrial
50   in the North African, South Asian and North American regions, but not in Australia (Scussolini et al., 2019).
52   In summary, there is medium confidence that CMIP5 and CMIP6 models can reproduce broad aspects of
53   precipitation changes during paleo reference periods, but large discrepancies remain. Further assessment of
54   model performance and comparison between CMIP5 and CMIP6 during past climates can be found in Section
     Do Not Cite, Quote or Distribute                      3-30                                      Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 5   Figure 3.11: Comparison between simulated annual precipitation changes and pollen-based reconstructions at
 6                the Mid-Holocene (6,000 years ago). The area-averaged changes over five regions (Iturbide et al., 2020)
 7                as simulated by CMIP6 models (individually identifiable, one ensemble member per model) and CMIP5
 8                models (blue) are shown, stretching from the tropics to high-latitudes. All regions contain multiple
 9                quantitative reconstructions: their interquartile range are shown by boxes and with whiskers for their full
10                range excluding outliers. Figure is adapted from Brierley et al. (2020). Further details on data sources and
11                processing are available in the chapter data table (Table 3.SM.1).
13   [END FIGURE 3.11 HERE]
16 Atmospheric Water Vapour
18   The AR5 concluded that an anthropogenic contribution to increases in specific humidity is found with
19   medium confidence at and near the surface. A levelling off of atmospheric water vapour over land in the last
20   two decades that needed better understanding, and remaining observational uncertainties, precluded a more
21   confident assessment (Bindoff et al., 2013). Sections and show that there have been
22   significant advances in the understanding of the processes controlling land surface humidity. In particular,
23   there has been a focus on the role of oceanic moisture transport and land-atmosphere feedbacks in explaining
24   the observed trends in relative humidity.
26   Water vapour is the most important natural greenhouse gas and its amount is expected to increase in a global
27   warming context leading to further warming. Particularly important are changes in the upper troposphere
28   because there water vapour regulates the strength of the water-vapour feedback (Section CMIP5
29   models have been shown to have a wet bias in the tropical upper troposphere and a dry bias in the lower
30   troposphere, with the former bias and model spread being larger than the latter (Jiang et al., 2012; Tian et al.,
31   2013). Tian et al. (2013) also showed that in comparison to the AIRS specific humidity, CMIP5 models have
32   the well-known double intertropical convergence zone (ITCZ) bias in the troposphere from 1000 hPa to 300
33   hPa, especially in the tropical Pacific. Water vapour biases in models are dominated by errors in relative
34   humidity throughout the troposphere, which are in turn closely related to errors in large scale circulation;
35   temperature errors dominate near the tropopause (Takahashi et al., 2016). Section 7.4.2 discusses this topic
36   in more detail for CMIP6 models. However, Schroeder et al. (2019) show that the majority of well-
37   established water vapour records are affected by inhomogeneity issues and thus should be used with caution
38   (see also Section A comparison of trends in column water vapour path for 1998-2019 in satellite
39   data, a reanalysis, CMIP5 and CMIP6 simulations averaged over the near-global ocean reveals that while on
40   average model trends are higher than those in observations and reanalysis, the latter lie within the multi-
41   model range (Figure 3.12).
46   Figure 3.12: Column water vapour path trends (%/decade) for the period 1998-2019 averaged over the near-
47                global ocean (50°S-50°N). The figure shows satellite data (RSS) and ERA5.1 reanalysis, as well as
48                CMIP5 (sky blue) and CMIP6 (brown) historical simulations. All available ensemble members were used
49                (see Section 3.2). Fits to the model trend probability distributions were performed with kernel density
50                estimation. Figure is updated from Santer et al. (2007). Further details on data sources and processing are
51                available in the chapter data table (Table 3.SM.1).
53   [END FIGURE 3.12 HERE]
56   The detection and attribution of tropospheric water vapour changes can be traced back to Santer et al. (2007),
     Do Not Cite, Quote or Distribute                          3-31                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   who used estimates of the atmospheric water vapour from satellite-based Special Sensor Microwave Imager
 2   (SSM/I) and from CMIP3 historical climate simulations. They provide evidence of human-induced
 3   moistening of the troposphere, and the simulated human fingerprint pattern was detectable at the 5% level by
 4   2002 in water vapour satellite data (from 1988 to 2006). The observed changes matched the historical
 5   simulations forced by greenhouse gas changes and other anthropogenic forcings, and not those due to natural
 6   variability alone. Then, Santer et al. (2009) repeated this study with CMIP5 models, and found that the
 7   detection and attribution conclusions were not sensitive to model quality. These results demonstrate that the
 8   human fingerprint is governed by robust and basic physical processes, such as the water vapour feedback.
 9   Finally, Chung et al. (2014) extended this line of research by focusing on the global-mean water vapour
10   content in the upper troposphere. Using satellite-based observations and sets of CMIP5 climate simulations
11   run under various climate-forcing options, they showed that the observed moistening trend of the upper
12   troposphere over the 1979-2005 period could not be explained by internal variability alone, but is attributable
13   to a combination of anthropogenic and natural forcings. This increase in water vapour is accompanied by a
14   reduction in mid-tropospheric relative humidity and clouds in the subtropics and mid-latitude in both models
15   and observations related to changes in the Hadley cell (Lau and Kim, 2015; also Section
17   Dunn et al. (2017) confirmed earlier findings that global mean surface relative humidity increased during
18   1973-2000, followed by a steep decline (also reported in Willet et al., 2014) until 2013, and specific
19   humidity correspondingly increased and then remained approximately constant (see also Section,
20   with none of the CMIP5 models capturing this behaviour. They noted biases in the mean state of the CMIP5
21   models’ surface relative humidity (and ascribe the failure to the representation of land surface processes and
22   their response to CO2 forcing), concluding that these biases preclude any detection and attribution
23   assessment. On the other hand, Byrne and O’Gorman (2018) show that the positive trend in specific
24   humidity continued in recent years and can be detected over land and ocean from 1979 to 2016. Moreover,
25   they provide a theory suggesting that the increase in annual surface temperature and specific humidity as
26   well as the decrease in relative humidity observed over land are linked to warming over the neighbouring
27   ocean. They also point out that the negative trend in relative humidity over land regions is quite uncertain
28   and requires further investigation. A recent study has also identified an anthropogenically-driven decrease in
29   relative humidity over the NH midlatitude continents in summer during 1979-2014, which was
30   underestimated by CMIP5 models (Douville and Plazzotta, 2017). Furthermore, in a modelling study
31   Douville et al. (2020) showed that this decrease in boreal summer relative humidity over midlatitudes is
32   related not only to global ocean warming, but also to the physiological effect of CO2 on plants in the land
33   surface model.
35   In summary, we assess that it is likely that human influence has contributed to moistening in the upper
36   troposphere since 1979. Also, there is medium confidence that human influence contributed to a global
37   increase in annual surface specific humidity, and medium confidence that it contributed to a decrease in
38   surface relative humidity over midlatitude NH continents during summertime.
41 Precipitation
43   AR5 concluded that there was medium confidence that human influence had contributed to large-scale
44   precipitation changes over land since 1950, including an increase in the NH mid to high latitudes. Moreover,
45   AR5 concluded that observational uncertainties and challenges in precipitation modelling precluded a more
46   confident assessment (Bindoff et al., 2013). Overall, they found that large-scale features of mean
47   precipitation in CMIP5 models are in modest agreement with observations, but there are systematic errors in
48   the tropics (Flato et al., 2013).
50   Since AR5, Li et al. (2016b) found that CMIP5 models simulate the large scale patterns of annual mean land
51   precipitation and seasonality well, as well as reproducing qualitatively the observed zonal mean land
52   precipitation trends for the period 1948-2005: models capture the drying trends in the tropics and at 45°S and
53   the wetting trend in the NH mid-to-high latitudes, but the amplitudes of the changes are much smaller than
54   observed. Land precipitation was found to show enhanced seasonality in observations (Chou et al., 2013),
55   qualitatively consistent with the simulated response to anthropogenic forcing (Dwyer et al., 2014). However,
     Do Not Cite, Quote or Distribute                     3-32                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   models do not appear to reproduce the zonal mean trends of seasonality over the period 1948-2005, nor the
 2   two-dimensional distributions of trends of annual precipitation and seasonality over land, but differences
 3   may be explainable by internal variability (Li et al., 2016b). However, observed trends in seasonality depend
 4   on data set used (Li et al., 2016b; Marvel et al., 2017), and Marvel et al. (2017) found that observed changes
 5   in the annual cycle phase are consistent with model estimates of forced changes. These phase changes are
 6   mainly characterized by earlier onset of the wet season on the equatorward flanks of the extratropical storm
 7   tracks, particularly in the SH. Box 8.2 assesses regional changes in water cycle seasonality.
 9   The CMIP5 models have also been shown to adequately simulate the mean and interannual variability of the
10   global monsoon (Section, but maintain the double ITCZ bias in the equatorial Pacific (Lee and
11   Wang, 2014a; Tian, 2015; Ni and Hsu, 2018). Despite the ITCZ bias, CMIP5 models have been used to
12   detect in reanalysis a southward shift in the ITCZ prior to 1975, followed by a northward shift in the ITCZ
13   after 1975, in response to forced changes in inter-hemispheric temperature contrast (Bonfils et al., 2020;
14   Friedman et al., 2020) (Sections and, Figure 8.11). CMIP5 models perform better than CMIP3
15   models, in particular regarding the global monsoon domain and intensity (Lee and Wang, 2014b).
17   In observations at time scales less than a day intermittent rainfall fluctuations dominate variability, but
18   CMIP5 models systematically underestimate them (Covey et al., 2018). Moreover, as noted in previous
19   generation models, CMIP5 models produce rainfall too early in the day (Covey et al., 2016). Also, models
20   overpredict precipitation frequency but have weaker intensity, although comparison with observed data sets
21   is complex as the latter present large differences in intensity among them (Herold et al., 2016; Pendergrass
22   and Deser, 2017; Trenberth et al., 2017). Regarding trends in precipitation intensity, models have also been
23   shown to reproduce the compensation between increasing heavy precipitation and decreasing light to
24   moderate rainfall (Thackeray et al., 2018), a characteristic found in the observational record (Gu and Adler,
25   2018). Regional performance is further assessed in Chapters 8 and the Atlas, while precipitation extremes are
26   considered in Chapter 11.
28   The simulation of annual mean rainfall patterns in the CMIP6 models reveals minor improvements compared
29   to those of CMIP5 models (Figure 3.13). The persistent biases include the double ITCZ in the tropical
30   Pacific (seen as bands of excessive rainfall at both sides of the equatorial Pacific in Figure 3.13b,d) and the
31   southward-shifted ITCZ in the equatorial Atlantic, which have been linked to the meridional pattern of SST
32   bias (Zhou et al., 2020a) and the reduced sensitivity of precipitation to local SST (Good et al., 2021). Tian
33   and Dong (2020) also found that all three generations of CMIP models share similar systematic annual mean
34   precipitation errors in the tropics, but that the double ITCZ bias is slightly reduced in CMIP6 models in
35   comparison to CMIP3 and CMIP5 models. They also found some improvement in the overly intense Indian
36   ocean ITCZ and the too dry South American continent except over the Andes. Fiedler et al. (2020) identified
37   improvements in the tropical mean spatial correlations and root mean square error of the climatology as well
38   as in the day-to-day variability, but found little change across CMIP phases in the double ITCZ bias and
39   diurnal cycle. The CMIP6 models reproduce better the domain and intensity of the global monsoon (see
40   Section Moreover, CMIP6 models better represent the storm tracks (Priestley et al., 2020; also
41   Section, thereby reducing the precipitation biases in the North Atlantic and midlatitudes of the SH
42   (Figure 3.13b,d). As a result, pattern correlations between simulated and observed annual mean precipitation
43   range between 0.80 and 0.92 for CMIP6, compared to a range of 0.79 to 0.88 for CMIP5 (Bock et al., 2020).
44   This relative improvement may be related to increased model resolution, as found when comparing biases in
45   the mean of the HighResMIP models with the mean of the corresponding lower-resolution versions of the
46   same models (see Figure 3.13e,f), particularly in the tropics and extratropical storm tracks. In agreement, a
47   recent study using several coupled models showed that increasing the atmospheric resolution leads to a
48   strong decrease in the precipitation bias in the tropical Atlantic ITCZ (Vannière et al., 2019) (see further
49   discussion in Section Based on these results we assess that despite some improvements, CMIP6
50   models still have deficiencies in simulating precipitation patterns, particularly over the tropical ocean (high
51   confidence).
     Do Not Cite, Quote or Distribute                     3-33                                      Total pages: 202
     Final Government Distribution                         Chapter 3                                      IPCC AR6 WGI

 1   Figure 3.13: Annual-mean precipitation rate (mm day–1) for the period 1995–2014. (a) Multi-model (ensemble)
 2                mean constructed with one realization of the CMIP6 historical experiments from each model. (b) Multi-
 3                model mean bias, defined as the difference between the CMIP6 multi-model mean and precipitation
 4                analyses from the Global Precipitation Climatology Project (GPCP) version 2.3 (Adler et al., 2003). (c)
 5                Multi-model mean of the root mean square error calculated over all months separately and averaged with
 6                respect to the precipitation analyses from GPCP v2.3. (d) Multi-model-mean bias, calculated as the
 7                difference between the CMIP6 multi-model mean and the precipitation analyses from GPCP v2.3. Also
 8                shown is the multi-model mean bias as the difference between the multi-model mean of (e) high
 9                resolution and (f) low resolution simulations of four HighResMIP models and the precipitation analyses
10                from GPCP v2.3. Uncertainty is represented using the advanced approach: No overlay indicates regions
11                with robust signal, where ≥66% of models show change greater than variability threshold and ≥80% of all
12                models agree on sign of change; diagonal lines indicate regions with no change or no robust signal, where
13                <66% of models show a change greater than the variability threshold; crossed lines indicate regions with
14                conflicting signal, where ≥66% of models show change greater than variability threshold and <80% of all
15                models agree on the sign of change. For more information on the advanced approach, please refer to the
16                Cross-Chapter Box Atlas.1. Dots in panel e) marks areas where the bias in high resolution versions of the
17                HighResMIP models is lower in at least 3 out of 4 models than in the corresponding low resolution
18                versions. Further details on data sources and processing are available in the chapter data table (Table
19                3.SM.1).
21   [END FIGURE 3.13 HERE]
24   Recent studies comparing observations and CMIP5 simulations have shown that tropical volcanic eruptions
25   induce a significant reduction in global precipitation, particularly over the wet tropics, including the global
26   monsoon regions (Iles and Hegerl, 2014; Paik and Min, 2017; Paik et al., 2020a). Reconstructions and
27   modelling studies also suggest a distinct remote influence of volcanic forcing such that large volcanoes
28   erupting in one hemisphere can enhance global monsoon precipitation in the other hemisphere (Liu et al.,
29   2016a; Zuo et al., 2019). The climatic effect of volcanic eruptions is further assessed in Cross-Chapter Box
30   4.1.
32   An intensification of the wet-dry zonal mean patterns, consisting of the wet tropical and mid-latitude bands
33   becoming wetter, and the dry subtopics becoming drier is expected in response to greenhouse gas and ozone
34   changes (Section However, detecting these changes is complicated by model errors in locating the
35   main features of rainfall patterns. To deal with this issue, Marvel and Bonfils (2013) identified in each
36   CMIP5 historical simulation the latitudinal peaks and troughs of the rainfall latitudinal patterns, measured
37   the amplification and shift of these patterns in a pattern-based fingerprinting study, and found that the
38   simultaneous amplification and shift in zonal precipitation patterns are detectable in Global Precipitation
39   Climatology Project (GPCP) observations over the 1979-2012 period. Similarly, Bonfils et al. (2020) found
40   that the intensification of wet-dry zonal patterns identified in CMIP5 historical simulations is detectable in
41   reanalyses over the 1950-2014 period (see also Figure 8.11).
43   Based on long-term island precipitation records, Polson et al. (2016) identified significant increases in
44   precipitation in the tropics and decreases in the subtropics, which are consistent with those simulated by the
45   CMIP5 models. Moreover, results from Polson and Hegerl (2017) give support to an intensification of the
46   water cycle according to the wet-gets-wetter, dry-gets-drier paradigm over tropical land areas as well. Other
47   studies suggest that this paradigm does not necessarily hold over dry regions where moisture is limited
48   (Greve et al., 2014; Kumar et al., 2015, see also Section Polson and Hegerl (2017) explained this
49   discrepancy by taking into account the seasonal and interannual movement of the regions (Allan, 2014). A
50   follow-up study using CMIP6 models also found that the observed strengthening contrast of precipitation
51   over wet and dry regions was detectable, although the increase was significantly larger in observations than
52   in the multi-model mean. The change was attributed to a combination of anthropogenic and natural forcings,
53   with anthropogenic forcings detectable in multi-signal analyses (Figure 3.14; Schurer et al. (2020)).
     Do Not Cite, Quote or Distribute                        3-34                                         Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 1   Figure 3.14: Wet (a) and dry (b) region tropical mean (30°S-30°N) annual precipitation anomalies. Observed data
 2                are shown with black lines (GPCP), ERA5 reanalysis in grey, single model simulations results are shown
 3                with light blue/red lines (CMIP6), and multi-model-mean results are shown with dark blue/red lines
 4                (CMIP6). Wet and dry region annual anomalies are calculated as the running mean over 12 months
 5                relative to a 1988-2020 base period. The regions are defined as the wettest third and driest third of the
 6                surface area, calculated for the observations and for each model separately for each season (following
 7                Polson and Hegerl, 2017). Scaling factors (panels c,d) are calculated for the combination of the wet and
 8                dry region mean, where the observations, reanalysis and all the model simulations are first standardised
 9                using the mean standard deviation of the pre-industrial control simulations. Two total least squares
10                regression methods are used: noise in variables (following Polson and Hegerl, 2017) which estimates a
11                best estimate and a 5-95% confidence interval using the pre-industrial controls (circle and thick green
12                line) and the pre-industrial controls with double the variance (thin green line); and a bootstrap method
13                (DelSole et al., 2019) (5-95% confidence interval shown with a purple line and best estimate with a
14                purple circle). Panel (c) shows results for GPCP and panel (d) for ERA5. Figure is adapted from Schurer
15                et al. (2020). Further details on data sources and processing are available in the chapter data table (Table
16                3.SM.1).
18   [END FIGURE 3.14 HERE]
21   Global land precipitation has likely increased since the middle of the 20th century (medium confidence), while
22   there is low confidence in trends in land data prior to 1950 and over the ocean during the satellite era due to
23   disagreement between datasets (Section Figure 3.15a shows the time evolution of the global mean
24   land precipitation since 1950, as well as the trend during the period. Adler et al. (2017) found no significant
25   trend in the global mean precipitation during the satellite era, consistent with model simulations (Wu et al.,
26   2013) and physical understanding of the energy budget (Section 8.2.1). This has been suggested to be due to
27   the negative effect of anthropogenic sulphates that opposed the positive influence of rising global mean
28   temperatures due to greenhouse gases (Salzmann, 2016; Richardson et al., 2018). The precipitation change
29   expected from ocean warming is also partly offset by the fast atmospheric adjustment to increasing
30   greenhouse gases (Section 8.2.1). Over the ocean, the negligible trend may be due to the cancelling effects of
31   CO2 and aerosols (Richardson et al., 2018).
33   A gridpoint based analysis of annual precipitation trends over land regions since 1901 (Knutson and Zeng,
34   2018) comparing observed and simulated trends found that detectable anthropogenic increasing trends have
35   occurred prominently over many middle to high latitude regions of the NH and subtropics of the SH. The
36   observed trends in many cases are significantly stronger than modelled in the CMIP5 historical runs for the
37   1901-2010 period (though not for 1951-2010), which may be due to disagreement between observed data
38   sets (Section, and/or suggest possible deficiencies in models.
40   The observed precipitation increase in the NH high latitudes over the period 1966-2005 was attributed to
41   anthropogenic forcing by a study using CMIP5 models (Wan et al., 2015) supporting the AR5 assessment.
42   Initial results from CMIP6 also support the role of anthropogenic forcing in the precipitation increase
43   observed in NH high latitudes (see Figure 3.15c): the observed positive trend detected for the band 60°N-
44   90°N can only be reproduced when anthropogenic forcing is included, although models tend to simulate
45   overall a larger positive trend. A similar positive trend, but less significant, is also detected between 30°N-
46   60°N, while in the southern mid-latitudes no trend is simulated (see Figure 3.15d, f).
51   Figure 3.15: Observed and simulated time series of anomalies in zonal average annual mean precipitation. a), c)-
52                f) Evolution of global and zonal average annual mean precipitation (mm day-1) over areas of land where
53                there are observations, expressed relative to the base-line period of 1961–1990, simulated by CMIP6
54                models (one ensemble member per model) forced with both anthropogenic and natural forcings (brown)
55                and natural forcings only (green). Multi-model means are shown in thick solid lines and shading shows
56                the 5-95% confidence interval of the individual model simulations. The data is smoothed using a low pass
57                filter. Observations from three different data sets are included: gridded values derived from Global
     Do Not Cite, Quote or Distribute                          3-35                                          Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1                Historical Climatology Network (GHCN V2) station data, updated from Zhang et al. (2007), data from
 2                the Global Precipitation Climatology Product (GPCP L3 v2.3, Huffman and Bolvin (2013)) and from the
 3                Climate Research Unit (CRU TS4.02, Harris et al. (2014)). Also plotted are boxplots showing
 4                interquartile and 5-95% ranges of simulated trends over the period for simulations forced with both
 5                anthropogenic and natural forcings (brown) and natural forcings only (blue). Observed trends for each
 6                observational product are shown as horizontal lines. Panel b) shows annual mean precipitation rate (mm
 7                day-1) of GHCN V2 for the years 1950-2014 over land areas used to compute the plots. Further details on
 8                data sources and processing are available in the chapter data table (Table 3.SM.1).
10   [END FIGURE 3.15 HERE]
13   For the SH extratropics, Solman and Orlanski (2016) found that the observed summertime rainfall increase
14   over high latitudes and decrease over mid-latitudes over the period 1979-2010 are quasi-zonally symmetric
15   and related to changes in eddy activity. The latter were in turn found to be associated with the poleward shift
16   of the westerlies due mostly to ozone depletion. Positive rainfall trends in the subtropics, particularly over
17   southeastern South America (see also Section and northern and central Australia, have been also
18   attributed to stratospheric ozone depletion (Kang et al., 2011; Gonzalez et al., 2014) and greenhouse gases
19   (Vera and Diaz, 2014; Saurral et al., 2019). During austral winter, wetting at high latitudes and drying at
20   mid-latitudes are not zonally homogenous, due to both changes in eddy activity and increased lower
21   troposphere humidity. Solman and Orlanski (2016) associated these climate changes with increases in
22   greenhouse gas concentration levels. Recently, Blazquez and Solman (2017) have shown that CMIP5 models
23   represent very well the dynamical forcing and the frequency of frontal precipitation in the SH winter
24   extratropics, but the amount of precipitation due to fronts is overestimated. Chapters 10 and 11 validate in
25   more detail the simulation of fronts in climate models (Sections and
27   Over the ocean, observations show coherent large-scale patterns of fresh ocean regions becoming fresher and
28   salty ocean regions saltier across the globe, which has been related through modelling studies to changes in
29   precipitation minus evaporation and is consistent with the wet-gets-wetter, dry-gets-drier paradigm (see
30   Sections and, Durack et al., 2012, 2013; Skliris et al., 2014; Durack, 2015; Hegerl et al., 2015;
31   Levang and Schmitt, 2015; Zika et al., 2015; Grist et al., 2016; Cheng et al., 2020).
33   Overall, studies published since AR5 provide further evidence of an anthropogenic influence on
34   precipitation, and therefore we now assess that it is likely that human influence has contributed to large-scale
35   precipitation changes observed since the mid-20th century. New attribution studies strengthen previous
36   findings of a detectable increase in mid to high latitude land precipitation over the NH (high confidence).
37   There is medium confidence that human influence has contributed to a strengthening of the zonal mean wet
38   tropics-dry subtropics contrast, and that tropical rainfall changes follow the wet-gets-wetter, dry-gets-drier
39   paradigm. There is also medium confidence that ozone depletion has increased precipitation over the
40   southern high latitudes and decreased it over southern midlatitudes during austral summer. Owing to
41   observational uncertainties and inconsistent results between studies, we conclude that there is low confidence
42   in the attribution of changes in the seasonality of precipitation.
45 Streamflow
47   Streamflow is to-date the only variable of the terrestrial water cycle with enough in-situ observations to
48   allow for detection and attribution analysis at continental to global scales. Based on evidence from a few
49   formal detection and attribution studies, particularly on the timing of peak streamflow, and the qualitative
50   evaluation of studies reporting on observed and simulated trends, AR5 concluded that there is medium
51   confidence that anthropogenic influence on climate has affected streamflow in some middle and high latitude
52   regions. AR5 also noted that observational uncertainties are large and that often only a limited number of
53   models were considered.
55   Section assesses that there have not been significant trends in global average streamflow over the
56   last century, though regional trends have been observed, driven in part by internal variability. Only a limited
     Do Not Cite, Quote or Distribute                       3-36                                         Total pages: 202
     Final Government Distribution                       Chapter 3                                      IPCC AR6 WGI

 1   number of studies have systematically compared observed streamflow trends at continental to global scales
 2   with changes simulated by global circulation models in a detection and attribution setting. Yang et al. (2017)
 3   did not find a significant correlation between observed runoff changes and changes simulated in CMIP5
 4   models in most grid cells, consistent with the assessment that observed changes are dominated by internal
 5   variability. In a pan-European assessment, Gudmundsson et al. (2017) attribute the spatio-temporal pattern of
 6   decreasing streamflow in southern Europe and increasing streamflow in northern Europe to anthropogenic
 7   climate change, but also concluded that additional effects of human water withdrawals could not be
 8   excluded. Focussing on continental runoff during 1958-2004, Alkama et al. (2013) found a significant
 9   change only when using reconstructed data over all rivers, and a large uncertainty in the estimate of the
10   global streamflow trend due to opposing changes over different continents. Gedney et al. (2014) detected the
11   influence of aerosols on streamflow in North America and Europe, with aerosols having driven an increase
12   in streamflow due to reduced evaporation (see Section for details on processes). There is also
13   evidence for a detectable anthropogenic contribution toward earlier winter-spring streamflows in the north-
14   central US (Kam et al., 2018) and in western Canada (Najafi et al., 2017). From a model evaluation
15   perspective, Sheffield et al. (2013) reported that CMIP5 models reproduce spatial variations in runoff in
16   North America well, though they tend to underestimate it.
18   Recently, Gudmundsson et al. (2021) performed a global detection and attribution study on streamflow and
19   found that some regions are drying and others are wetting. Moreover, the simulated streamflow trends are
20   consistent with observations only if externally forced climate change is considered, and the simulated effects
21   of water and land management cannot reproduce the observed trends. The effects of volcanic eruptions in
22   driving reduced streamflow have also been detected in the wet tropics (Iles and Hegerl, 2015; Zuo et al.,
23   2019).
25   In summary, there is medium confidence that anthropogenic climate change has altered local and regional
26   streamflow in various parts of the world and that the associated global-scale trend pattern is inconsistent with
27   internal variability. Moreover, human interventions and water withdrawals, while affecting streamflow,
28   cannot explain the observed spatio-temporal trends (medium confidence).
33   Cross-Chapter Box 3.2:          Human Influence on Large-scale Changes in Temperature and
34                                   Precipitation Extremes
36   Contributors: Nathan Gillett (Canada), Seung-Ki Min (Republic of Korea), Krishnan Raghavan (India), Ying
37   Sun (China), Xuebin Zhang (Canada)
39   Understanding how temperature and precipitation extremes have changed at large scales and the causes of
40   these changes is an important part of our overall assessment of human influence on the climate system.
41   Chapter 11 assesses changes in extremes and their causes, while this Cross-Chapter Box summarizes
42   relevant assessments and supporting evidence in Chapters 8 and 11 and relates changes in extremes to mean
43   changes on global and continental scales.
45   Attribution of temperature extremes
46   One important aspect of various indicators of temperature extremes is their connection to mean temperature
47   at local, regional and global scales. For example, the highest daily temperature in a summer is often highly
48   correlated with the summer mean temperature. Model projections show that changes in temperature extremes
49   are often closely related to shifts in mean temperature (Seneviratne et al., 2016; Kharin et al., 2018). It is thus
50   no surprise that changes in temperature extremes are consistent with warming mean temperature, with
51   warming leading to more hot extremes and fewer cold extremes. Given the attribution of mean warming to
52   human influence (Section 3.3.1), and the connection between changes in mean and extreme temperatures, it
53   is to be expected that anthropogenic forcing has also influenced temperature extremes.
55   Chapter 11 assesses that there is high confidence that climate models can reproduce the mean state and
     Do Not Cite, Quote or Distribute                       3-37                                       Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   overall warming of temperature extremes observed globally and in most regions, although the magnitude of
 2   the trends may differ, and the ability of models to capture observed trends in temperature-related extremes
 3   depends on the metric evaluated, the way indices are calculated, and the time periods and spatial scales
 4   considered (Section 11.3.3). There has been widespread evidence of human influence on various aspects of
 5   temperature extremes, at global, continental, and regional scales. This includes attribution to human
 6   influence of observed changes in intensity, frequency, and duration and other relevant characteristics at
 7   global and continental scales (Section 11.3.4). The left panel of Cross-Chapter Box 3.2, Figure 1 clearly
 8   shows that long-term changes in the global mean annual maximum daily maximum temperature can be
 9   reproduced by both CMIP5 and CMIP6 models forced with the combined effect of natural and
10   anthropogenic forcings, but cannot be reproduced by simulations under natural forcing alone. Consistent
11   with the assessment for global mean temperature (Section 3.3.1), aerosol changes are found to have offset
12   part of the greenhouse gas induced increase in hot extremes globally and over most continents over the 1951-
13   2015 period (Hu et al., 2020; Seong et al., 2021), though greenhouse gas and aerosol influences are less
14   clearly separable in observed changes in cold extremes.
16   Chapter 11 assesses that it is virtually certain that human-induced greenhouse gas forcing is the main
17   contributor to the observed increase in the likelihood and severity of hot extremes and the observed decrease
18   in the likelihood and severity of cold extremes on global scales, and very likely that this applies on most
19   continents.
21   Attribution of precipitation extremes
22   An important piece of evidence supporting the SREX and AR5 assessment that there is medium confidence
23   that anthropogenic forcing has contributed to a global scale intensification of heavy precipitation during the
24   second half of the 20th century is the evidence for anthropogenic influence on other aspects of the global
25   hydrological cycle. The most significant aspect of that is the increase in atmospheric moisture content
26   associated with warming which should, in general, lead to enhanced extreme precipitation, particularly
27   associated with enhanced convergence in tropical and extratropical cyclones (Sections, 11.4.1). Such
28   a connection is supported by the fact that annual maximum one-day precipitation increases with global mean
29   temperature at a rate similar to the increase in the moisture holding capacity in response to warming, both in
30   observations and in model simulations. Additionally, models project an increase in extreme precipitation
31   across global land regions even in areas in which total annual or seasonal precipitation is projected to
32   decrease.
34   The overall performance of CMIP6 models in simulating extreme precipitation intensity and frequency is
35   similar to that of CMIP5 models (high confidence), and there is high confidence in the ability of models to
36   capture the large-scale spatial distribution of precipitation extremes over land (Section 11.4.3). Evidence of
37   human influence on extreme precipitation has become stronger since the AR5. Considering changes in
38   precipitation intensity averaged over all wet days, there is high confidence that daily mean precipitation
39   intensities have increased since the mid-20th century in a majority of land regions, including Europe, North
40   America and Asia, and it is likely that such an increase is mainly due to anthropogenic emissions of
41   greenhouse gases (Section and 11.4.4). Section 11.4.4 also finds a larger fraction of land showing
42   enhanced extreme precipitation and a larger probability of record-breaking one-day precipitation than
43   expected by chance, which can only be explained when anthropogenic greenhouse gas forcing is considered.
44   The right panel of Cross-Chapter Box 3.2, Figure 1 demonstrates the consistency in global average annual
45   maximum daily precipitation in the observations and model simulations under combined anthropogenic and
46   natural forcing, and inconsistency with natural forcing alone. While there is more evidence in the literature to
47   quantify the net anthropogenic influence on extreme precipitation than the influence of individual forcing
48   components, a dominant contribution of greenhouse gas forcing to the long-term intensification of extreme
49   precipitation on global and continental scales has recently been quantified separately from the influence of
50   anthropogenic aerosol and natural forcings (Dong et al., 2020; Paik et al., 2020b).
52   Chapter 11 assesses that it is likely that human influence, in particular due to greenhouse gas forcing, is the
53   main driver of the observed intensification of heavy precipitation in global land regions during recent
54   decades (Section 11.4.4).
     Do Not Cite, Quote or Distribute                      3-38                                       Total pages: 202
     Final Government Distribution                      Chapter 3                                       IPCC AR6 WGI

 4   Cross-Chapter Box 3.2, Figure 1:     Comparison of observed and simulated changes in global mean
 5                                        temperature and precipitation extremes. Time series of globally averaged 5-
 6                                        year mean anomalies of the annual maximum daily maximum temperature
 7                                        (TXx in °C) and annual maximum 1-day precipitation (Rx1day as standardized
 8                                        probability index in %) during 1953-2017 from the HadEX3 observations and
 9                                        the CMIP5 and CMIP6 multi-model ensembles with natural and human forcing
10                                        (upper) and natural forcing only (lower). For CMIP5, historical simulations for
11                                        1953-2005 are combined with corresponding RCP4.5 scenario runs for 2006-
12                                        2017. For CMIP6, historical simulations for 1953-2014 are combined with
13                                        SSP2-4.5 scenario simulations for 2015-2017. Numbers in brackets represents
14                                        the number of models used. The time-fixed observational mask has been
15                                        applied to model data throughout the whole period. Grid cells with more than
16                                        70% data availability during 1953-2017 plus data for at least 3 years during
17                                        2013-2017 are used. Coloured lines indicate multi-model means, while shading
18                                        represents 5th-95th percentile ranges, based on all available ensemble members
19                                        with equal weight given to each model (Section 3.2). Anomalies are relative to
20                                        1961-1990 means. Figure is updated from Seong et al. (2021), their Figure 3
21                                        and Paik et al. (2020), their Figure 3. Further details on data sources and
22                                        processing are available in the chapter data table (Table 3.SM.1).
29   3.3.3   Atmospheric Circulation
31 The Hadley and Walker Circulations
33   The tropical tropospheric circulation features meridional and zonal overturning circulations, called Hadley
34   and Walker circulations. In the zonal mean, the downwelling branch of the Hadley circulation cell is located
35   in the subtropics and is often used as an indicator of the meridional extent of the tropics. In the equatorial
36   zonal-vertical section, the major rising branch of the Walker circulation is located over the Maritime
37   continent with secondary ascending regions over northern South America and Africa. The zonal component
38   of the surface trade winds over most of the equatorial Pacific and Atlantic is associated with the Walker
39   circulation. The present section assesses the zonal-mean Hadley cell extent and the Pacific Walker
40   circulation strength. Regional and water cycle aspects of these circulations are assessed in more detail in
41   Section 8.3.2.
43   AR5 found medium confidence that the depletion of stratospheric ozone had contributed to Hadley cell
44   widening in the SH in austral summer (Bindoff et al., 2013). It also noted that in contrast to a simulated
45   weakening in response to greenhouse gas forcing, the Walker circulation had actually strengthened since the
46   early 1990s, precluding any detection of human influence.
48   Hadley cell extent
49   Grise et al. (2019) found that a metric based on surface zonal winds, which are well constrained by surface
50   observations, best compares reanalyses with CMIP5 models. With this method and new reanalysis products,
51   the CMIP5 historical simulation ensembles simulate comparable mean states and variability of the
52   subtropical edge latitude of the Hadley cells to those observed (Grise et al., 2019).
54   Chapter 2 assesses that there has been a very likely widening of the Hadley circulation since the 1980s
55   (Section The CMIP5 (Davis and Birner, 2017; Grise et al., 2018) and CMIP6 (Grise and Davis,
56   2020) historical simulation ensembles span the observed trend of the zonal-mean Hadley cell edges since the
57   1980s (Figure 3.16a-c). Studies based on CMIP5 models find a contribution from human influence to the
     Do Not Cite, Quote or Distribute                      3-39                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   observed widening trend, especially in the SH (Gerber and Son, 2014a; Staten et al., 2018, 2020; Grise et al.,
 2   2019; Jebri et al., 2020), which is confirmed in CMIP6 (Figure 3.16b-c; Grise and Davis, 2020).
 4   In the annual mean, internal variability, including Pacific Decadal Variability (PDV; Annex IV.2.6),
 5   contributed to the observed zonal-mean Hadley cell expansion since 1980 comparably with human influence
 6   (Allen et al., 2014; Allen and Kovilakam, 2017; Mantsis et al., 2017; Amaya et al., 2018; Grise et al., 2018).
 7   Indeed, the ensemble-mean expansion in historical simulations is significantly weaker than in most of the
 8   reanalyses shown in Figure 3.16a-c, while the Atmospheric Model Intercomparison Project (AMIP)
 9   simulations forced by observed SSTs (Figure 3.16a-c) show stronger trends than historical coupled
10   simulations on average (Davis and Birner, 2017; Grise et al., 2018; Nguyen et al., 2015). The human-induced
11   change has not yet clearly emerged out of the internal variability range in the NH expansion (Quan et al.,
12   2018; Grise et al., 2019), whereas the trend in the annual-mean SH edge is outside the 5-95th percentile range
13   of internal variability in CMIP6 in 3 out of the 4 reanalyses (Figure 3.16b). For the SH summer when the
14   simulated human influence is strongest, the 1981-2000 trend in 3 out of the 4 reanalyses falls outside the 5-
15   95th percentile range of internal variability (Grise et al., 2018, 2019; Tao et al., 2016).
17   In CMIP5 simulations, greenhouse gas increases and, in austral summer, stratospheric ozone depletion,
18   contribute to the SH expansion (Gerber and Son, 2014b; Nguyen et al., 2015; Tao et al., 2016a; Kim et al.,
19   2017b), but the ozone influence is not significant in available CMIP6 simulations (Figure 3.16b-c). Since the
20   2000s, the stabilization or slight recovery of stratospheric ozone (Section is consistent with the
21   smaller observed trends (Banerjee et al., 2020). While many CMIP5 models underrepresent the magnitude of
22   the PDV, implying potential overconfidence on the detection of human influence on the Hadley cell
23   expansion, this is less the case for the CMIP6 models (Section 3.7.6). However, the mechanism underlying
24   the Hadley Cell expansion remains unclear (Staten et al., 2018, 2020), precluding a process-based validation
25   of the simulated human influence.
27   Walker circulation strength
28   CMIP5 models reproduce the mean state of the Walker circulation with reasonable fidelity, evidenced by the
29   spatial pattern correlations of equatorial zonal mass stream function between models and observations being
30   larger than 0.88 (Ma and Zhou, 2016a). CMIP5 historical simulations on average simulate a significant
31   weakening of the Pacific Walker circulation over the 20th century (DiNezio et al., 2013; Sandeep et al., 2014;
32   Kociuba and Power, 2015), which is also seen in CMIP6 (Figure 3.16d). This weakening is accompanied by
33   reduction of convective activity over the Maritime Continent and enhancement over the central equatorial
34   Pacific (DiNezio et al., 2013; Sandeep et al., 2014; Kociuba and Power, 2015). In the CMIP6 simulations,
35   greenhouse gas forcing induces this weakening (Figure 3.16d), which is consistent with theories based on
36   radiative-convective equilibrium (Vecchi et al., 2006; Vecchi and Soden, 2007) and thermodynamic air-sea
37   coupling (Xie et al., 2010), but inconsistent with a theory highlighting the ocean dynamical effect which
38   suggests a strengthening in response to greenhouse gas increases (Clement et al., 1996; Seager et al., 2019;
39   see also Section Seager et al. (2019) attributed this inconsistency to equatorial Pacific SST biases
40   in the models (Section However, observational and reanalysis data sets disagree on the sign of
41   trends in the Walker Circulation strength over the 1901-2010 period (Figure 3.16d,), and Section
42   assesses low confidence in observed long-term Walker Circulation trends. The observational uncertainty
43   remains high in the trends since the 1950s (Tokinaga et al., 2012; L’Heureux et al., 2013), though both
44   CMIP5 and CMIP6 historical simulations span trends of all but one observational data set (Figure 3.16e). For
45   this period, external influence simulated in CMIP6 is insignificant due to a partial compensation of forced
46   responses to greenhouse gases and aerosols and large internal decadal variability (Figure 3.16e). It is notable
47   that while AMIP simulations on average show strengthening over both the periods, those simulations are
48   forced by one reconstruction of SST, which itself is subject to uncertainty before the 1970s (Deser et al.,
49   2010; Tokinaga et al., 2012).
51   Observational SST products indicate that the equatorial zonal SST gradient from the western to the eastern
52   equatorial Pacific has strengthened since 1870 (Section While CMIP5 historical simulations on
53   average simulate a weakening, large ensemble simulations span the observed strengthening since the 1950s
54   (Watanabe et al., 2021) suggesting an important contribution from internal variability. Coats and Karnauskas
55   (2017) also find that the anthropogenic influence on the SST gradient is yet to emerge out of internal
     Do Not Cite, Quote or Distribute                      3-40                                      Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1   variability even on centennial timescales.
 3   Trends since the 1980s in in-situ and satellite observations and reanalyses exhibit strengthening of the
 4   Pacific Walker circulation and SST gradient (L’Heureux et al., 2013; Boisséson et al., 2014; England et al.,
 5   2014; Kociuba and Power, 2015; Ma and Zhou, 2016; Section; Figure 3.16f). AMIP simulations
 6   reproduce this strengthening (Boisséson et al., 2014; Ma and Zhou, 2016; Figure 3.16d), indicating a
 7   dominant role of SST changes. However, all reanalysis trends lie outside the 5-95% range of simulated
 8   CMIP6 historical Walker circulation trends over this period (Figure 3.16f), consistent with CMIP5 results
 9   (England et al., 2014; Kociuba and Power, 2015). This may be in part caused by the underestimation of the
10   PDV magnitude especially in CMIP5 models (Kociuba and Power, 2015; Chung et al., 2019; Section 3.7.6),
11   but also suggests a potential error in simulating the forced changes of the Walker circulation. Specifically,
12   anthropogenic and volcanic aerosol changes over this period may have driven a strengthening (DiNezio et
13   al., 2013; Takahashi and Watanabe, 2016a; Hua et al., 2018). This aerosol influence may be indirect via
14   Atlantic Multidecadal Variability (AMV; Annex IV.2.7) through inter-basin teleconnections (McGregor et
15   al., 2014; Li et al., 2015d; Chikamoto et al., 2016; Kucharski et al., 2016; Ruprich-Robert et al., 2017),
16   which may be underestimated in models due to SST biases in the equatorial Atlantic (Section;
17   McGregor et al., 2018). Note also the large uncertainty in aerosol influence on the Walker circulation (Kuntz
18   and Schrag, 2016; Hua et al., 2018; Oudar et al., 2018), which is also seen in CMIP6 (Figure 3.16f).
20   Paleoclimate data from the Pliocene epoch suggest that there was a reduction in the zonal SST gradient in the
21   tropical Pacific under a similar CO2 concentration as today (Cross-Chapter Box 2.4; Section
22   Tierney et al. (2019) found that this weaker gradient compared to pre-industrial, which suggests a weaker
23   Walker circulation, is captured by climate models under Pliocene CO2 levels, in agreement with the CMIP6
24   response to greenhouse gas forcing (Figure 3.16d), though the magnitude of this effect varies strongly
25   between models (Corvec and Fletcher, 2017).
27   Summary
28   It is likely that human influence has contributed to the poleward expansion of the zonal mean Hadley cell in
29   the SH since the 1980s. This assessment is supported by studies since AR5, which consistently find human
30   influence from greenhouse gas increases on the expansion, with additional influence from ozone depletion in
31   austral summer. For the strong ozone depletion period of 1981-2000, human influence is detectable in the
32   summertime poleward expansion in the SH (medium confidence). By contrast, there is medium confidence
33   that the expansion of the zonal mean Hadley cell in the NH is within the range of internal variability, with
34   contributions from PDV and other internal variability. The causes of the observed strengthening of the
35   Pacific Walker circulation over the 1980-2014 period are not well understood, since the observed
36   strengthening trend is outside the range of variability simulated in the coupled models (medium confidence).
37   Large observational uncertainty, lack of understanding of the mechanism underlying the poleward Hadley
38   cell expansion, and contradicting theories on the greenhouse gas influence and uncertainty in the aerosol
39   influence on the Walker circulation strength, limit confidence in these assessments.
44   Figure 3.16: Model evaluation and attribution of changes in Hadley cell extent and Walker circulation strength.
45                (a-c) Trends in subtropical edge latitude of the Hadley cells in (a) the Northern Hemisphere for 1980-
46                2014 annual mean and (b-c) Southern Hemisphere for (b) 1980-2014 annual mean and (c) 1980/81-
47                1999/2000 December-January-February mean. Positive values indicate northward shifts. (d-f) Trends in
48                the Pacific Walker circulation strength for (d) 1901-2010, (e) 1951-2010 and (f) 1980-2014. Positive
49                values indicate strengthening. Based on CMIP5 historical (extended with RCP4.5), CMIP6 historical,
50                AMIP, pre-industrial control, and single forcing simulations along with HadSLP2 and reanalyses. Pre-
51                industrial control simulations are divided into non-overlapping segments of the same length as the other
52                simulations. White boxes and whiskers represent mean, interquartile ranges and 5th and 95th percentiles,
53                calculated after weighting individual members with the inverse of the ensemble of the same model, so
54                that individual models are equally weighted (Section 3.2). The filled boxes represent the 5-95%
55                confidence interval on the multi-model mean trends of the models with at least 3 ensemble members, with
56                dots indicating the ensemble means of individual models. The edge latitude of the Hadley cell is defined

     Do Not Cite, Quote or Distribute                        3-41                                        Total pages: 202
     Final Government Distribution                        Chapter 3                                       IPCC AR6 WGI

 1                where the surface zonal wind velocity changes sign from negative to positive, as described in the
 2                Appendix of Grise et al. (2018). The Pacific Walker circulation strength is evaluated as the annual mean
 3                difference of sea level pressure between 5ºS-5ºN, 160ºW-80ºW and 5ºS-5ºN, 80ºE-160ºE. Further details
 4                on data sources and processing are available in the chapter data table (Table 3.SM.1).
 6   [END FIGURE 3.16 HERE]
 9 Global Monsoon
11   Monsoons are seasonal transitions of regimes in atmospheric circulation and precipitation with the annual
12   cycle of solar insolation, in association with redistribution of moist static energy (Wang and Ding, 2008;
13   Wang et al., 2014b; Biasutti et al., 2018). The global monsoon can be defined to encompasses all monsoon
14   systems based on precipitation contrast in the solstice seasons (Wang and Ding, 2008; Figure 3.17). All
15   regional monsoons are intimately connected to the global tropical atmospheric overturning by mass
16   (Trenberth et al., 2000), momentum and energy budgets (Biasutti et al., 2018; Geen et al., 2020).
17   Assessments of regional monsoon changes are made in Section and Sections and 10.6.3.
19   AR5 assessed that CMIP5 models simulated monsoons better than CMIP3 models but that biases remained
20   in domains and intensity (high confidence) (Flato et al., 2013). There were no detection and attribution
21   assessment statements on the decreasing trend of global monsoon precipitation over land from the 1950s to
22   the 1980s or the increasing trend of global monsoon precipitation afterwards. In the paleoclimate context, it
23   was determined with high confidence that orbital forcing produces strong interhemispheric rainfall variability
24   evident in multiple types of proxies (Masson-Delmotte et al., 2013a).
26   Paleoclimate proxy evidence shows that the global monsoon has varied with orbital forcing and greenhouse
27   gases (Section; Mohtadi et al., 2016; Seth et al., 2019). These large-magnitude intensifications and
28   weakenings in the global monsoon involved in some cases orders-of-magnitude changes in precipitation
29   locally (Harrison et al., 2014; Tierney et al., 2017c). Paleoclimate modelling and limited data from past
30   climate states with high CO2 suggest that precipitation intensifies in the monsoon domain under elevated
31   greenhouse gases, providing context for present and future trends (Passey et al., 2009; Haywood et al., 2013;
32   Zhang et al., 2013b). In model simulations of the mid-Pliocene, when globally averaged temperature was
33   higher than present day, precipitation was larger in West African and South and East Asian monsoons than
34   under pre-industrial conditions, consistent with proxy evidence (Zhang et al., 2015; Sun et al., 2016, 2018;
35   Corvec and Fletcher, 2017; Li et al., 2018a). Prescott et al. (2019) and Zhang et al. (2019) find an important
36   role for orbital forcing and CO2 in the mid-Pliocene monsoon expansion and intensification. Models are also
37   able to capture interhemispherically contrasting monsoon changes in the Last Interglacial in response to
38   orbital forcing and greenhouse gases, with wetter West African and Asian monsoons and a drier South
39   American monsoon as seen in proxies (Govin et al., 2014; Gierz et al., 2017; Pedersen et al., 2017). In
40   overall agreement with proxy evidence, a model with transient forcing simulates wetting and drying
41   respectively of the Southern and NH monsoons during the last deglaciation, with an important contribution
42   from Atlantic Meridional Overturning Circulation (AMOC) slowdown (Otto-Bliesner et al., 2014; Mohtadi
43   et al., 2016).
45   During the mid-Holocene, global monsoons were stronger especially in the NH with an expansion of the
46   West African monsoon domain in response to orbital forcing (Biasutti et al., 2018; Section
47   Simulations of the mid-Holocene with CMIP5 and CMIP6 models qualitatively capture the stronger NH
48   monsoon (Jiang et al., 2015; Brierley et al., 2020), mainly driven by atmospheric circulation changes
49   (D’Agostino et al., 2019). However, the models underestimate the monsoon expansion found in proxy
50   reconstructions (Perez-Sanz et al., 2014; Harrison et al., 2015; Tierney et al., 2017d), which may be linked to
51   mean biases in the monsoon domain (Brierley et al., 2020) and may be improved by imposing vegetation and
52   dust changes (Pausata et al., 2016). The models simulate the weaker SH monsoon during the mid-Holocene
53   (D’Agostino et al., 2020), consistent with proxy evidence (Section These studies indicate that
54   models can qualitatively reproduce past global monsoon changes seen in proxies, though issues remain in
55   quantitatively reproducing proxy observations. Studies of last millennium simulations show that simulated

     Do Not Cite, Quote or Distribute                        3-42                                         Total pages: 202
     Final Government Distribution                     Chapter 3                                   IPCC AR6 WGI

 1   global monsoon precipitation increases with global mean temperature, while changes in monsoon circulation
 2   and hemispheric monsoon precipitation depend on forcing sources (Liu et al., 2012; Chai et al., 2018).
 3   Compared to greenhouse gas and solar variations, volcanic forcing is more effective in changing the global
 4   monsoon precipitation over the last millennium (Chai et al., 2018).
 6   Reproducing monsoons in terms of domain, precipitation amount, and timings of onset and retreat over the
 7   historical period also remains difficult. While CMIP5 historical simulations correctly capture global
 8   monsoon domains and intensity based on summer and winter precipitation differences, they underestimate
 9   the extent and intensity of East Asian and North American monsoons while overestimating them over the
10   tropical western North Pacific (Lee and Wang, 2014b; Yan et al., 2016a). Wang et al. (2020) reported that
11   CMIP6 models simulate the global monsoon domain and precipitation better (Figure 3.17a,b), albeit with
12   biases in annual mean precipitation and the timings of onset and withdrawal of the SH monsoon. Notable
13   inter-model differences were identified in CMIP5, with the multi-model ensemble mean outperforming
14   individual models (Lee and Wang, 2014b). Common biases were identified across CMIP5 models in moist
15   static energy and upper-tropospheric temperature associated with the South Asian summer monsoon, which
16   may arise from overly smoothed model topography (Boos and Hurley, 2012). However, in atmospheric
17   models with increasing resolution approaching 20 km, improvements in monsoon precipitation are not
18   universal across regions and models, and overall improvements are unclear (Johnson et al., 2016; Ogata et
19   al., 2017; Zhang et al., 2018b).
21   In instrumental records, global summer monsoon precipitation intensity (measured by summer precipitation
22   averaged over the monsoon domain) decreased from the 1950s to 1980s, followed by an increase (Section
23; Figure 3.17c), arising mainly from variations in Northern Hemispheric land monsoons. A CMIP5
24   multi-model study by Zhang et al. (2018b) found that observed 1951-2004 trends of the global and NH
25   summer land monsoon precipitation intensity are well captured by historical simulations, and CMIP6 models
26   show similar results for global land summer monsoon precipitation (Figure 3.17c). However, the 1960s peak
27   in the NH summer monsoon circulation is outside the 5th-95th percentile range of CMIP5 and CMIP6
28   historical simulations for two out of three reanalyses (Figure 3.17d). Modelling studies show that greenhouse
29   gas increases act to enhance NH summer monsoon precipitation intensity (Liu et al., 2012; Polson et al.,
30   2014; Chai et al., 2018; Zhang et al., 2018b). Since the mid-20th century, however, modelling studies show
31   that this effect was overwhelmed by the influence of anthropogenic aerosols in CMIP5 (Polson et al., 2014;
32   Guo et al., 2015; Zhang et al., 2018c; Giannini and Kaplan, 2019) and in CMIP6 (Zhou et al., 2020b).
33   Weakening of the monsoon circulation and reduction of moisture availability are important in this aerosol
34   influence (Zhou et al., 2020b). Besides these human influences, the global monsoon is sensitive to internal
35   variability and natural forcing including ENSO and volcanic aerosols on interannual time scales and PDV
36   and AMV on decadal to multidecadal time scales (Wang et al., 2013a, 2018; Liu et al., 2016a; Jiang and
37   Zhou, 2019; Zuo et al., 2019); though AMV in the 20th century may have been partly driven by aerosols, see
38   Section 3.7.7. Indeed, AMIP simulations better reproduce the observed multidecadal variations of the global
39   monsoon precipitation and circulation (Figure 3.17c,d). Zhang et al. (2018b) find that the multi-model
40   ensemble mean trend of global land monsoon precipitation in historical simulations, dominated by
41   anthropogenic aerosol forcing contributions, emerges out of the 90% range of internally-driven trends in pre-
42   industrial control simulations. However, it should be noted that CMIP5 models tend to underrepresent the
43   PDV magnitude (Section 3.7.6), suggesting potential overconfidence on the detection of the forced signal.
44   An observed enhancement in global summer monsoon precipitation since the 1980s is accompanied by an
45   intensification of the NH summer monsoon circulation (Figure 3.17c,d). These trends appear to be at the
46   extreme of the range of the CMIP6 historical simulation ensemble but well captured by AMIP simulations
47   (Figure 3.17c,d). While the precipitation increase is consistent with greenhouse gas forcing, the circulation
48   intensification is opposite to the simulated response to greenhouse gas forcing, and these enhancements have
49   been attributed to PDV and AMV (Wang et al., 2013b; Kamae et al., 2017).
51   In summary, while greenhouse gas increases acted to enhance the global land monsoon precipitation over the
52   20th century (medium confidence), consistent with projected future enhancement (Section 4.5.1), this
53   tendency was overwhelmed by anthropogenic aerosols from the 1950s to the 1980s, which contributed to
54   weakening of global land summer monsoon precipitation intensity for this period (medium confidence).
55   There is medium confidence that the intensification of global monsoon precipitation and NH summer
     Do Not Cite, Quote or Distribute                    3-43                                      Total pages: 202
     Final Government Distribution                         Chapter 3                                      IPCC AR6 WGI

 1   monsoon circulation since the 1980s is dominated by internal variability. These assessments are supported
 2   respectively by multi-model detection and attribution studies which find an important role for anthropogenic
 3   aerosols in the weakening trend, and studies that identify a role for AMV and PDV in inducing the NH
 4   summer monsoon circulation enhancement since the 1980s. Supported by multi-model simulations that are
 5   qualitatively consistent with proxy evidence, there is high confidence that orbital forcing contributed to
 6   stronger NH monsoon precipitation in the mid-Pliocene and mid-Holocene than pre-industrial. While CMIP5
 7   models can capture the domain and precipitation intensity of the global monsoon, biases remain in their
 8   regional representations, and they are unsuccessful in quantitatively reproducing changes in paleo
 9   reconstructions (high confidence). CMIP6 models reproduce the domain and precipitation intensity of the
10   global monsoon observed over the instrumental period better than CMIP5 models (medium confidence).
11   However, CMIP5 and CMIP6 models fail to fully capture the variations of the NH summer monsoon
12   circulation (Figure 3.17d), but there is low confidence to this assessment due to a lack of evidence in the
13   literature.
18   Figure 3.17: Model evaluation of global monsoon domain, intensity, and circulation. (a-b) Climatological
19                summer-winter range of precipitation rate, scaled by annual mean precipitation rate (shading) and 850
20                hPa wind velocity (arrows) based on (a) GPCP and ERA5 and (b) a multi-model ensemble mean of
21                CMIP6 historical simulations for 1979-2014. Enclosed by red lines is the monsoon domain based on the
22                definition by Wang and Ding (2008). (c-d) 5-year running mean anomalies of (c) global land monsoon
23                precipitation index defined as the percentage anomaly of the summertime precipitation rate averaged over
24                the monsoon regions over land, relative to its average for 1979-2014 (the period indicated by light grey
25                shading) and (d) the tropical monsoon circulation index defined as the vertical shear of zonal winds
26                between 850 and 200 hPa levels averaged over 0º-20ºN, from 120ºW eastward to 120ºE in NH summer
27                (Wang et al., 2013; m s–1) in CMIP5 historical and RCP4.5 simulations, CMIP6 historical and AMIP
28                simulations. Summer and winter are defined for individual hemispheres: May through September for NH
29                summer and SH winter, and November through March for NH winter and SH summer. The number of
30                models and ensembles are given in the legend. The multi-model ensemble mean and percentiles are
31                calculated after weighting individual members with the inverse of the ensemble size of the same model,
32                so that individual models are equally weighted irrespective of ensemble size. Further details on data
33                sources and processing are available in the chapter data table (Table 3.SM.1).
35   [END FIGURE 3.17 HERE]
38 Extratropical Jets, Storm Tracks and Blocking
40   Extratropical jets are wind maxima in the upper troposphere which are often associated with storms,
41   blocking, and weather extremes. Blocking refers to long-lived, stationary high-pressure systems that are
42   often associated with a poleward displacement of the jet, causing cold spells in winter and heat waves in
43   summer (e.g., Sousa et al., 2018). Sections,, and 11.7.2 discuss these features in more detail.
45   AR5 concluded that models were able to capture the general characteristics of extratropical cyclones and
46   storm tracks, although it also noted that most models underestimated cyclone intensity, that biases in cyclone
47   frequency were linked to biases in sea surface temperatures, and that resolution can play a significant role in
48   the quality of the simulation of storms (Flato et al., 2013). Similarly, AR5 found with high confidence that
49   simulation of blocking was improved due to increases in resolution. AR5 did not specifically assess changes
50   in Southern-Hemisphere storm track characteristics or blocking.
52   Since AR5, new research using CMIP5 and CMIP6 models has confirmed that increasing the model
53   resolution improves the simulation of cyclones and blocking in all seasons albeit with some exceptions and
54   caveats (Zappa et al., 2013; Davini et al., 2017; Schiemann et al., 2017, 2020; Davini and D’Andrea, 2020;
55   Priestley et al., 2020). New research also finds that model performance with respect to the simulation of
56   cyclones and that of blocking events are correlated (Zappa et al., 2014), suggesting biases in both are aspects

     Do Not Cite, Quote or Distribute                        3-44                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   of the same underlying problem in models (Figure 3.18). In the North Pacific basin the annual-mean
 2   blocking frequency is now well simulated compared to earlier evaluations, but substantial errors in the
 3   blocking frequency remain in the Euro-Atlantic sector (Dunn-Sigouin and Son, 2013; Davini and D’Andrea,
 4   2016, 2020; Mitchell et al., 2017; Woollings et al., 2018; Figure 3.18). While there is a resolution
 5   dependence in the size of this bias, even at very high resolution blocking in the Euro-Atlantic sector remains
 6   underestimated (Schiemann et al., 2017), and there is evidence of a compensation of errors as the resolution
 7   is increased (Davini et al., 2017). Davini and D’Andrea (2020) show that while the simulation of blocking
 8   improves with increasing resolution in CMIP3, CMIP5, and CMIP6 models, other factors contribute to
 9   biases, particularly to the underestimation of Euro-Atlantic blocking (Schiemann et al., 2020). The
10   persistence of blocking events, typically underestimated, has not improved from CMIP5 to CMIP6
11   (Schiemann et al., 2020). Chapter discusses the implications of the biases discussed here for
12   regional climate.
14   For the North Pacific storm track CMIP6 simulations exhibit large remaining underestimations of cyclone
15   frequencies during summer (June to August), which for the low-resolution models have essentially remained
16   unchanged versus CMIP5, and there is only a small resolution dependence of this bias (Priestley et al.,
17   2020). During winter (December to February), both CMIP5 and CMIP6 models tend to place the North
18   Pacific storm track too far equatorward (Yang et al., 2018b; Priestley et al., 2020), leading to an
19   overestimation of cyclones between 30 and 40°N in the Pacific and an underestimation to the north of this.
20   Both low- and high-resolution models show this pattern, but low-resolution models generally simulate fewer
21   cyclones throughout the North Pacific (Priestley et al., 2020).
23   In winter, the North Atlantic storm track remains displaced to the south and east in many models (Harvey et
24   al., 2020), leading to underestimations of cyclone frequencies near the North American coast and
25   overestimations in the eastern North Atlantic. Higher-resolution CMIP6 models perform slightly better in
26   this regard than low-resolution models. In summer (June to August), cyclone frequencies throughout the
27   extratropical North Atlantic, which were substantially underestimated in CMIP5, have improved in CMIP6
28   high-resolution models. In low-resolution CMIP6 models, the problem is essentially unchanged (Priestley et
29   al., 2020); this is associated with generally underestimated variability of sea level pressure in CMIP models
30   (Harvey et al., 2020).
32   For the SH (not considered in AR5), (Priestley et al., 2020) find considerable improvement in the placement
33   of the Southern Ocean storm track during summer (December to February) in CMIP6 models versus CMIP5,
34   consistent with a more realistic annual mean surface wind maximum latitude in CMIP6 than in CMIP5
35   (Goyal et al., 2021). Relative to CMIP5, both low- and high-resolution CMIP6 models have increased track
36   densities south of about 55°S and decreases between about 40 and 55°S, in better agreement with
37   observations than CMIP5 models (Parsons et al., 2016; Patterson et al., 2019). CMIP5 models and high-
38   resolution CMIP6 models simulate a storm track that is positioned too far equatorward, although the bias is
39   smaller in the high-resolution models. By contrast, the low-resolution CMIP6 models simulate a storm track
40   that is slightly too far poleward on average (Priestley et al., 2020). In winter (June to August), the biases
41   found in CMIP5 are only slightly improved on in CMIP6, with models continuing to underestimate the broad
42   maximum cyclone track density in the south-eastern Indian Ocean and overestimate the minimum in the
43   south-western South Pacific (Priestley et al., 2020).
45   There is only one contiguous blocking region in the SH, with the blocking frequency maximizing in the
46   South Pacific and minimizing in the southern Indian Ocean regions (Parsons et al., 2016; Patterson et al.,
47   2019). CMIP5 simulations agree relatively well with ERA-Interim in this region regarding the distribution of
48   blocking events (Parsons et al., 2016). Individual models exhibit considerable biases in the blocking
49   frequency; however only in austral summer do Patterson et al. (2019) find a systematic, multi-model
50   underestimation of the blocking frequency in and around the Tasman Sea. The blocking frequency is
51   anticorrelated with the amplitude of the SAM. Ozone depletion, through stratosphere-troposphere coupling,
52   may have caused an increase in the blocking frequency in the South Atlantic sector (Dennison et al., 2016);
53   this finding requires confirmation using a multi-model approach.
55   In addition to inadequate resolution, blocking and storm track biases in both hemispheres also result from
     Do Not Cite, Quote or Distribute                     3-45                                      Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1   mean state biases, in particular, biases related to the parameterization of orographic effects and to the
 2   misrepresentation of the Gulf Stream SST front (Anstey et al., 2013; Berckmans et al., 2013; Davini and
 3   D’Andrea, 2016b; O’Reilly et al., 2016b; Pithan et al., 2016; Schiemann et al., 2017). Overall SST biases
 4   have been suggested to have only a weak relevance to blocking (Davini and D’Andrea, 2016b).
 6   Section assesses that the total number of extratropical cyclones has likely increased since the 1980s
 7   in the NH (low confidence), but with fewer deep cyclones particularly in summer. This observed reduction in
 8   cyclone activity by about 4% per decade in the NH in summer (Chang et al., 2016; Section may be
 9   associated with human-induced warming. CMIP5 historical simulations generally reproduce a reduction but
10   underestimate its magnitude (Chang et al., 2016). Furthermore, feedback mechanisms associated with clouds
11   may be responsible for substantial inter-model spread (Chang et al., 2016; Voigt and Shaw, 2016). In boreal
12   winter, recent studies have suggested a potential influence of the rapid Arctic warming on observed
13   intensification of NH storm track activity in the past few decades, while other studies question this
14   possibility (Cross-chapter Box 10.1).
16   Section assesses that the extratropical jets and cyclone tracks have likely been shifting poleward in
17   both hemispheres since the 1980s with marked seasonality in trends (medium confidence). For the SH,
18   studies using CMIP5 and other models imply that both ozone depletion and increasing greenhouse gases
19   have caused substantial atmospheric circulation change since the 1960s when concentrations of ozone-
20   depleting substances started to increase (Eyring et al., 2013; Iglesias-Suarez et al., 2016; Karpechko et al.,
21   2018b; Son et al., 2018a). In particular, ozone depletion, during austral summer, has been linked to a
22   poleward shift of the westerly jet and Southern-Hemisphere circulation zones and a southward expansion of
23   the tropics (Kang et al., 2011), which is associated with a strengthening trend of the Southern Annular Mode
24   (SAM; Section 3.7.2). This has been well reproduced by climate models with prescribed historical ozone
25   concentration or interactive ozone chemistry (Gerber and Son, 2014; Son et al., 2018; Figure 3.19).
27   In summary, there is low confidence that an observed decrease in the frequency of NH summertime
28   extratropical cyclones is linked to anthropogenic influence. In the SH, there is high confidence that human
29   influence, in the form of ozone depletion, has contributed to the observed poleward shift of the jet in austral
30   summer, while confidence is low for human influence on historical blocking activity. The low confidence
31   statements are due to the limited number of studies available. The shift of the SH jet is correlated with
32   modulations of the SAM, and justification for the associated high confidence statement on attribution of
33   changes in the SAM is provided in Section 3.7.2. There is medium confidence in model performance
34   regarding the simulation of the extratropical jets, storm track and blocking activity, with increased resolution
35   sometimes corresponding to better performance, but important shortcomings remain, particularly for the
36   Euro-Atlantic sector of the NH. Nonetheless, synthesizing across Sections, there is high
37   confidence that CMIP6 models capture the general characteristics of the tropospheric large-scale circulation.
42   Figure 3.18: Instantaneous Northern-Hemisphere blocking frequency (% of days) in the extended northern
43                winter season (DJFM) for the years 1979-2000. Results are shown for ERA5 reanalysis (black), CMIP5
44                (blue) and CMIP6 (red) models. Coloured lines show multi-model means and shaded ranges show
45                corresponding 5-95% ranges constructed with one realization from each model. Figure is adapted from
46                Davini and D’Andrea (2020), their Figure 12 and following the D’Andrea et al. (1998) definition of
47                blocking. Further details on data sources and processing are available in the chapter data table (Table
48                3.SM.1).
50   [END FIGURE 3.18 HERE]
55   Figure 3.19: Long-term mean (thin black contour) and linear trend (colour) of zonal mean DJF zonal winds
56                over 1985-2014 in the SH. Displayed are (a) ERA5 and (b) CMIP6 multi-model mean (58 CMIP6
     Do Not Cite, Quote or Distribute                       3-46                                        Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 1                models). The solid contours show positive (westerly) and zero long-term mean zonal wind, and the
 2                dashed contours show negative (easterly) long-term mean zonal wind. Only one ensemble member per
 3                model is included. Figure is modified from Eyring et al. (2013), their Figure 12. Further details on data
 4                sources and processing are available in the chapter data table (Table 3.SM.1).
 6   [END FIGURE 3.19 HERE]
 9 Sudden Stratospheric Warming Activity
11   Sudden stratospheric warmings (SSWs) are stratospheric weather events associated with anomalously high
12   temperatures at high latitudes persisting from days to weeks. Section discusses the definition and
13   observational aspects of SSWs. SSWs are often associated with anomalous weather conditions, e.g. winter
14   cold spells, in the lower atmosphere (e.g., Butler et al., 2015a; Baldwin et al., 2021).
16   Seviour et al. (2016) found that stratosphere-resolving CMIP5 models, on average, reproduce the observed
17   frequency of vortex splits (one form of SSWs) but with a wide range of model-specific biases. Models that
18   produce a better mean state of the polar vortex also tend to produce a more realistic SSW frequency (Seviour
19   et al., 2016). The mean sea level pressure anomalies occurring in CMIP5 model simulations when an SSW is
20   underway however differs substantially from those in reanalyses (Seviour et al., 2016). Unlike stratosphere-
21   resolving models, models with limited stratospheric resolution, which make up more than half of the CMIP5
22   ensemble, underestimate the frequency of SSWs (Osprey et al., 2013; Kim et al., 2017a). Taguchi (2017)
23   found a general underestimation in CMIP5 models of the frequency of “major” SSWs (which are associated
24   with a break-up of the polar vortex), an aspect of an underrepresentation in those models of dynamical
25   variability in the stratosphere. Wu and Reichler (2020) found that finer vertical resolution in the stratosphere
26   and a model top above the stratopause tend to be associated with a more realistic SSW frequency in CMIP5
27   and CMIP6 models.
29   Some studies find an increase in the frequency of SSWs under increasing greenhouse gases (e.g. Kim et al.,
30   2017a; Schimanke et al., 2013; Young et al., 2013). However, this behaviour is not robust across ensembles
31   of chemistry-climate models (Mitchell et al., 2012; Ayarzagüena et al., 2018; Rao and Garfinkel, 2021).
32   There is an absence of studies specifically focusing on simulated trends in SSWs during recent decades, and
33   the short record and substantial decadal variability yields low confidence in any observed trends in the
34   occurrence of SSW events in the NH winter (Section Such an absence of a trend and large
35   variability would also be consistent with a recent reconstruction of SSWs extending back to 1850, based on
36   sea level pressure observations (Domeisen, 2019), although this time series has limitations as it is not based
37   on direct observations of SSWs.
39   In summary, an anthropogenic influence on the frequency or other aspects of SSWs has not yet been robustly
40   detected. There is low confidence in the ability of models to simulate any such trends over the historical
41   period because of large natural interannual variability and also due to common substantial biases in the
42   simulated mean state affecting the simulated frequency of SSWs.
45   3.4     Human Influence on the Cryosphere
47   3.4.1    Sea Ice
49 Arctic Sea Ice
51   The AR5 concluded that “anthropogenic forcings are very likely to have contributed to Arctic sea ice loss
52   since 1979” (Bindoff et al., 2013), based on studies showing that models can reproduce the observed decline
53   only when including anthropogenic forcings, and formal attribution studies. Since the beginning of the
54   modern satellite era in 1979, NH sea ice extent has exhibited significant declines in all months with the
55   largest reduction in September (see Figures 3.20 and 3.21 and Section for more details on observed

     Do Not Cite, Quote or Distribute                         3-47                                           Total pages: 202
     Final Government Distribution                        Chapter 3                                     IPCC AR6 WGI

 1   changes). The recent Arctic sea ice loss during summer is unprecedented since 1850 (high confidence), but
 2   as in AR5 and SROCC there remains only medium confidence that the recent reduction is unique during at
 3   least past 1000 years due to sparse observations (Sections and CMIP5 models also
 4   simulate NH sea ice loss over the satellite era but with large differences among models (e.g., Massonnet et
 5   al., 2012; Stroeve et al., 2012). The envelope of simulated ice loss across model simulations encompasses the
 6   observed change, although observations fall near the low end of the CMIP5 and CMIP6 distributions of
 7   trends (Figure 3.20). CMIP6 models on average better capture the observed Arctic sea ice decline, albeit
 8   with large inter-model spread. Notz et al. (2020) found that CMIP6 models better reproduce the sensitivity of
 9   Arctic sea ice area to CO2 emissions and global warming than earlier CMIP models although the models’
10   underestimation of this sensitivity remains. Davy and Outten (2020) also found that CMIP6 models can
11   simulate the seasonal cycle of Arctic sea ice extent and volume better than CMIP5 models. For the
12   assessment of physical processes associated with changes in Arctic sea ice, see Section
14   Since AR5, there have been several new detection and attribution studies on Arctic sea ice. While the
15   attribution literature has mostly used sea ice extent (SIE), it is closely proportional to sea ice area (SIA; Notz,
16   2014), which is assessed in Chapters 2 and 9 and shown in Figures 3.20 and 3.21. Kirchmeier-Young et al.
17   (2017) compared the observed time series of the September SIE over the period 1979-2012 with those from
18   different large ensemble simulations which provide a robust sampling of internal climate variability
19   (CanESM2, CESM1, and CMIP5) using an optimal fingerprinting technique. They detected anthropogenic
20   signals which were separable from the response to natural forcing due to solar irradiance variations and
21   volcanic aerosol, supporting previous findings (Figure 3.21; Kay et al., 2011; Min et al., 2008; Notz and
22   Marotzke, 2012; Notz and Stroeve, 2016). Using selected CMIP5 models and three independently derived
23   sets of observations, Mueller et al. (2018) detected fingerprints from greenhouse gases, natural, and other
24   anthropogenic forcings simultaneously in the September Arctic SIE over the period 1953-2012. They further
25   showed that about a quarter of the greenhouse gas induced decrease in SIE has been offset by an increase
26   due to other anthropogenic forcing (mainly aerosols). Similarly, Gagné et al. (2017a) suggested that the
27   observed increase in Arctic sea ice concentration over the 1950-1975 period was primarily due to the cooling
28   contribution of anthropogenic aerosols forcing based on single model simulations. Gagné et al. (2017b)
29   identified a detectable increase in Arctic SIE in response to volcanic eruptions using CMIP5 models and four
30   observational datasets. Polvani et al. (2020) suggested that ozone depleting substances played a substantial
31   role in the Arctic sea ice loss over the 1955-2005 period.
33   Differences in sea ice loss among the models (Figure 3.20) have been attributed to a number of factors (see
34   also Section These factors include the late 20th century simulated sea ice state (Massonnet et al.,
35   2012), the magnitude of changing ocean heat transport (Mahlstein and Knutti, 2011), and the rate of global
36   warming (e.g., Gregory et al., 2002; Mahlstein and Knutti, 2012; Rosenblum and Eisenman, 2017). Sea ice
37   thermodynamic considerations indicate that the magnitude of sea ice variability and loss depends on ice
38   thickness (Bitz, 2008; Massonnet et al., 2018) and hence the climatology simulated by different models may
39   influence their projections of change (medium confidence), as indicated by the regression lines in Figure
40   3.20.
42   An important consideration in comparing Arctic sea ice loss in models and observations is the role of
43   internal variability (medium confidence). Using ensemble simulations from a single model, Kay et al. (2012)
44   suggested that internal variability could account for about half of the observed September ice loss. More
45   recently, large ensemble simulations have been performed with many more ensemble members (Kay et al.,
46   2015). These enable a more robust characterization of internal variability in the presence of forced
47   anthropogenic change. Using such large ensembles, some studies discussed the influence of internal
48   variability on Arctic sea ice trends (Swart et al., 2015). Song et al. (2016) also compared the trends in the
49   forced and unforced simulations using multiple climate models and found that internal variability explains
50   about 40% of the observed September sea ice melting trend, supporting previous studies (Stroeve et al.,
51   2012). Based on the large ensembles of CESM1 and CanESM2, September Arctic sea ice extent variance
52   first increases and then decreases as SIE declines from its pre-industrial value (Kirchmeier-Young et al.,
53   2017; Mueller et al., 2018) consistent with previous work (Goosse et al., 2009), but neither study found a
54   strong sensitivity of detection and attribution results to the change in variability. Further work has indicated
55   that internally-driven summer atmospheric circulation trends with enhanced atmospheric ridges over
     Do Not Cite, Quote or Distribute                       3-48                                       Total pages: 202
     Final Government Distribution                           Chapter 3                                        IPCC AR6 WGI

 1   Greenland and the Arctic Ocean, which project on the negative phase of the North Atlantic Oscillation
 2   (Section 3.7.1), play an important role in the observed Arctic sea ice loss (Hanna et al., 2015; Ding et al.,
 3   2017). A fingerprint analysis using the CESM large ensemble suggests that this internal variability accounts
 4   for 40-50% of the observed September Arctic sea ice decline (Ding et al., 2019; England et al., 2019).
 5   Internally-generated decadal tropical variability and associated atmospheric teleconnections were suggested
 6   to have contributed to the changing atmospheric circulation in the Arctic and the associated rapid sea ice
 7   decline from 2000 to 2014 (Meehl et al., 2018).
 9   Some recent studies evaluated the human contribution to recent record minimum SIE events in the Arctic.
10   Analysing CMIP5 simulations, Zhang and Knutson (2013) found that the observed 2012 record low in
11   September Arctic SIE is inconsistent with internal climate variability alone. Based on several large
12   ensembles, Kirchmeier-Young et al. (2017) concluded that the observed 2012 SIE minimum cannot be
13   reproduced in a simulation excluding human influence. Fučkar et al. (2016) showed that climate change
14   contributed to the record low March Arctic SIE in 2015, which was accompanied by the record minimum
15   SIE in the Sea of Okhotsk (Paik et al., 2017).
17   Based on the new attribution studies since AR5, we conclude that it is very likely that anthropogenic forcing
18   mainly due to greenhouse gas increases was the main driver of Arctic sea ice loss since 1979. Increases in
19   anthropogenic aerosols have offset part of the greenhouse gas induced Arctic sea ice loss since the 1950s
20   (medium confidence). Despite large differences in the mean sea ice state in the Arctic, Arctic sea ice loss is
21   captured by all CMIP5 and CMIP6 models. Nonetheless, large inter-model differences in the Arctic sea ice
22   decline remain, limiting our ability to quantify forced changes and internal variability contributions.
27   Figure 3.20: Mean (x-axis) and trend (y-axis) in Arctic sea ice area (SIA) in September (left) and Antarctic SIA
28                in February (right) for 1979-2017 from CMIP5 (upper) and CMIP6 (lower) models. All individual
29                models (ensemble means) and the multi-model mean values are compared with the observations
30                (OSISAF, NASA Team, and Bootstrap). Solid line indicates a linear regression slope with corresponding
31                correlation coefficient (r) and p-value provided. Note the different scales used on the y-axis for Arctic and
32                Antarctic SIA. Results remain essentially the same when using sea ice extent (SIE) (not shown). Further
33                details on data sources and processing are available in the chapter data table (Table 3.SM.1).
35   [END FIGURE 3.20 HERE]
38 Antarctic Sea Ice
40   AR5 concluded that “there is low confidence in the attribution of the observed increase in Antarctic SIE
41   since 1979” (Bindoff et al., 2013) due to the limited understanding of the external forcing contribution as
42   well as the role of internal variability. Based on a difference between the first and last decades, Antarctic sea
43   ice cover exhibited a small increase in summer and winter over the 1979-2017 period (Section;
44   Figures 3.20 and 3.21). However, these changes are not statistically significant and starting in late 2016,
45   anomalously low sea ice area has been observed (Section The mean hemispheric sea ice changes
46   result from much larger, but partially compensating, regional changes with increases in the western Ross Sea
47   and Weddell Sea and declines in the Bellingshausen and Amundsen Seas (Hobbs et al., 2016). Observed
48   regional trends have been particularly large in austral autumn (see Section, and also Section
49   for more details of regional changes and related physical processes). Starting in austral spring of 2016, the
50   ice extent decreased strongly (Turner et al., 2017) and has since remained anomalously low (Figure 3.21).
51   This decrease has been associated with anomalous atmospheric conditions associated with teleconnections
52   from warming in the eastern Indian Ocean and a negative Southern Annular Mode (Chenoli et al., 2017;
53   Stuecker et al., 2017; Schlosser et al., 2018; Meehl et al., 2019; Purich and England, 2019; Wang et al.,
54   2019a). A decadal-scale warming of the near-surface ocean that resulted from strengthened westerlies may
55   also have contributed to and helped to sustain the sea ice loss (Meehl et al., 2019). Before satellites and on
56   even longer time scales, very limited observational data and proxy coverage leads to low confidence in all
     Do Not Cite, Quote or Distribute                          3-49                                          Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 1   aspects of Antarctic sea-ice (Section;
 3   CMIP5 climate models generally simulate Antarctic sea ice loss over the satellite era since 1979 (Mahlstein
 4   et al., 2013; Turner et al., 2013) in contrast to the observed change, and CMIP6 models also simulate
 5   Antarctic ice loss (Roach et al., 2020; Figure 3.20 and 3.21). A number of studies have suggested that this
 6   discrepancy may be in part due to the role of internal variability in the observed change (Mahlstein et al.,
 7   2013; Polvani and Smith, 2013; Zunz et al., 2013; Meehl et al., 2016a; Turner et al., 2016), including
 8   teleconnections associated with tropical Pacific variability (Meehl et al., 2016a) and changing surface
 9   conditions resulting from multi-decadal ocean circulation variations (Singh et al., 2019). However, when the
10   spatial pattern is considered, trends in the summer and autumn (from 1979-2005) appear outside the range of
11   internal variability (Hobbs et al., 2015). This suggests that the models may exhibit an unrealistic simulation
12   of the Antarctic sea ice forced response or the internal variability of the system. Discrepancies among the
13   models in simulated sea ice variability (Zunz et al., 2013), the sea ice climatological state (Roach et al.,
14   2018), upper ocean temperature trends (Schneider and Deser, 2018), SH westerly wind trends (Purich et al.,
15   2016), or the sea ice response to Southern Annular Mode variations (Ferreira et al., 2014; Holland et al.,
16   2017; Kostov et al., 2017; Landrum et al., 2017) may all play some role in explaining these differences with
17   the observed trends. Increased fresh water fluxes caused by mass loss of the Antarctic Ice Sheet (either by
18   melting at the front of ice shelves or via iceberg calving) have been suggested as a possible mechanism
19   driving the multidecadal Antarctic sea ice expansion (Bintanja et al., 2015; Pauling et al., 2016) but there is a
20   lack of consensus on this mechanism’s impacts (Pauling et al., 2017). A recent study based on decadal
21   predictability suggests that initializing the state of the Antarctic bottom water cell can reproduce the observed
22   Antarctic sea ice increase (Zhang et al., 2017), consistent with the suggestion that multidecadal variability
23   associated with variations in deep convection has contributed to the observed increase in Antarctic sea ice
24   since 1979 (Latif et al., 2013; Zhang et al., 2017, 2019a) (see also Section
26   There have been several studies which aimed to identify causes of the observed Antarctic SIE changes.
27   Gagné et al. (2015) assessed the consistency of observed and simulated changes in Antarctic SIE for an
28   extended period using recovered satellite-based estimates, and found that the observed trends since the mid-
29   1960s are not inconsistent with model simulated trends. Studies based on the satellite period also indicate
30   that the observed trends are largely within the range of simulated internal variability (Hobbs et al., 2016). A
31   few distinct factors that led to the weak signal-to-noise ratio in Antarctic SIE trends have been further
32   identified, which include large multi-decadal variability (Monselesan et al., 2015), the short observational
33   record (e.g., Abram et al., 2013), and the limited model performance at representing the complex Antarctic
34   climate system as discussed above (Bintanja et al., 2013; Uotila et al., 2014). The short period of
35   comprehensive satellite observations, beginning in 1979, makes it challenging to set the observed increase
36   between 1979 and 2015, or the subsequent decrease, in a long-term context, and to assess whether the
37   difference in trend between observations and models, which mostly simulate long-term decreases, is
38   systematic or a rare expression of internal variability on decadal-to-multidecadal timescales.
40   In conclusion, the observed small increase in Antarctic sea ice extent during the satellite era is not generally
41   captured by global climate models, and there is low confidence in attributing the causes of the change.
46   Figure 3.21: Seasonal evolution of observed and simulated Arctic (left) and Antarctic (right) sea ice area (SIA)
47                over 1979–2017. SIA anomalies relative to the 1979–2000 means from observations (OBS from
48                OSISAF, NASA Team, and Bootstrap, top) and historical (ALL, middle) and natural only (NAT, bottom)
49                simulations from CMIP5 and CMIP6 multi-models. These anomalies were obtained by computing non-
50                overlapping 3-year mean SIA anomalies for March (February for Antarctic SIA), June, September, and
51                December separately. CMIP5 historical simulations are extended by using RCP4.5 scenario simulations
52                after 2005 while CMIP6 historical simulations are extended by using SSP2-4.5 scenario simulations after
53                2014. CMIP5 NAT simulations end in 2012. Numbers in brackets represents the number of models used.
54                The multi-model mean is obtained by taking the ensemble mean for each model first and then averaging
55                over models. Grey dots indicate multi-model mean anomalies stronger than inter-model spread (beyond ±
56                1 standard deviation). Results remain very similar when based on sea ice extent (SIE) (not shown). Units:
     Do Not Cite, Quote or Distribute                         3-50                                        Total pages: 202
     Final Government Distribution                          Chapter 3                                         IPCC AR6 WGI

 1                 106 km2. Further details on data sources and processing are available in the chapter data table (Table
 2                 3.SM.1).
 4   [END FIGURE 3.21 HERE]
 7   3.4.2   Snow Cover
 9   Seasonal snow cover is a defining climate feature of the northern continents. It is therefore of considerable
10   interest that climate models correctly simulate this feature. It is discussed in more detail in Section 9.5.3, and
11   observational aspects of snow cover are assessed in Section
13   AR5 noted the strong linear correlation between NH snow cover extent (SCE) and annual-mean surface air
14   temperature in CMIP5 models. It was assessed as likely that there had been an anthropogenic contribution to
15   observed reductions in NH snow cover since 1970 (Bindoff et al., 2013). AR5 assessed that CMIP5 models
16   reproduced key features of observed snow cover well, including the seasonal cycle of snow cover over
17   northern regions of Eurasia and North America, but had more difficulties in more southern regions with
18   intermittent snow cover. AR5 also found that CMIP5 models underestimate the observed reduction in spring
19   snow cover over this period (see also Brutel-Vuilmet et al., 2013; Thackeray et al., 2016; Santolaria-Otín and
20   Zolina, 2020; Figure 3.22). This behaviour has been linked to how the snow-albedo feedback is represented
21   in models (Thackeray et al., 2018). The CMIP5 multi-model ensemble has been shown to represent the
22   snow-albedo feedback more realistically than CMIP3, although models from some individual modelling
23   centres have not improved or have even got worse (Thackeray et al., 2018). There is still a systematic
24   overestimation of the albedo of boreal forest covered in snow (Li et al., 2016c; Thackeray et al., 2015).
25   Consequently, the snow albedo feedback might have been overestimated by CMIP5 models (Xiao et al.,
26   2017; Section 9.5.3).
31   Figure 3.22: Time series of Northern Hemisphere March-April mean snow cover extent (SCE) from
32                observations, CMIP5 and CMIP6 simulations. The observations (grey lines) are updated Brown-
33                NOAA (Brown and Robinson, 2011), Mudryk et al. (2020), and GLDAS2. CMIP5 (upper) and CMIP6
34                (lower) simulations of the response to natural plus anthropogenic forcing are shown in orange, natural
35                forcing only in green, and the pre-industrial control simulation range is presented in blue. 5-year mean
36                anomalies are shown for the 1923-2017 period with the x-axis representing the centre years of each 5-
37                year mean. CMIP5 all forcing simulations are extended by using RCP4.5 scenario simulations after 2005
38                while CMIP6 all forcing simulations are extended by using SSP2-4.5 scenario simulations after 2014.
39                Shading indicates 5th-95th percentile ranges for CMIP5 and CMIP6 all and natural forcings simulations,
40                and solid lines are ensemble means, based on all available ensemble members with equal weight given to
41                each model (Section 3.2). The blue vertical bar indicates the mean 5th-95th percentile range of pre-
42                industrial control simulation anomalies, based on non-overlapping segments. The numbers in brackets
43                indicate the number of models used. Anomalies are relative to the average over 1971-2000. For models,
44                SCE is restricted to grid cells with land fraction ≥ 50%. Greenland is excluded from the total area
45                summation. Figure is modified from Paik et al. (2020), their Figure 1. Further details on data sources and
46                processing are available in the chapter data table (Table 3.SM.1).
48   [END FIGURE 3.22 HERE]
51   CMIP6 models improve on CMIP5 models in producing slightly increased SCE versus CMIP5, correcting
52   the low bias in CMIP5 mentioned above (Mudryk et al., 2020). The linear relationship noted above between
53   GSAT and SCE also exists in CMIP6 (Mudryk et al., 2020). Like CMIP5, the CMIP6 models capture the
54   negative trend in spring snow cover that has occurred in recent decades (Figure 3.22). However, the median
55   CMIP6 model now produces slightly stronger post-1981 declines in the March-April mean SCE than the
56   CMIP5 median (Mudryk et al., 2020). Until about 1980, the models produce a generally stable March-April
57   SCE, but after that a substantial decline, reaching a loss of about 2 × 106 km2 in 2012-2017 relative to the
     Do Not Cite, Quote or Distribute                          3-51                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   1971-2000 average. Compared to earlier studies which found that models underestimate observed trends for
 2   the 1979-2005 period (Brutel-Vuilmet et al., 2013), both CMIP5 and CMIP6 models show improved
 3   agreement with the observations over the period to 2017 (Figure 3.22). One remaining concern is a failure of
 4   most CMIP6 models to correctly represent the relationship between snow cover extent and snow mass,
 5   reflecting too slow seasonal increases and decreases of SCE in the models (Mudryk et al., 2020).
 7   Several CMIP5 and CMIP6 based studies have consistently attributed the observed NH spring SCE changes
 8   (Section to anthropogenic influences (Rupp et al., 2013; Najafi et al., 2016; Paik and Min, 2020),
 9   with the observed changes being found to be inconsistent with natural variability alone. Similarly, spring
10   snow thickness (Snow Water Equivalent) changes on the scale of the NH have been attributed to greenhouse
11   gas forcing (Jeong et al., 2017). Using individual forcing simulations from multiple CMIP6 models, Paik and
12   Min (2020) detected greenhouse gas influence in the observed decrease of early spring SCE during 1925-
13   2019, which was found to be separable from the responses to other forcings.
15   In summary, it is very likely that anthropogenic influence contributed to the observed reductions in NH
16   springtime snow cover since 1950. CMIP6 models better represent the seasonality and geographical
17   distribution of snow cover than CMIP5 simulations (high confidence). Both CMIP5 and CMIP6 models
18   simulate strong declines in spring-time SCE during recent years, in general agreement with observations,
19   causing the multi-model mean decreasing trend in spring-time SCE to now better agree with observations
20   than in earlier evaluations. Evidence has yet to emerge that interactions between vegetation and snow, found
21   problematic in CMIP5, have indeed improved in CMIP6 models (Section 9.5.3). Such deficiencies in the
22   representation of snow in climate models mean there is medium confidence in the simulation of snow cover
23   over the northern continents in CMIP6 model simulations. The models consistently link snow extent to
24   surface air temperature (Figure 9.24). With warming of near-surface air linked to anthropogenic influence,
25   and particularly to greenhouse gas increases (Section 3.3.1), this provides additional evidence that reductions
26   in snow cover are also caused by human activity.
29   3.4.3   Glaciers and Ice Sheets
31   While Chapter 9 (Sections 9.4 and 9.5) discusses process understanding for glaciers and ice sheets, as well as
32   evaluation of global and regional-scale glacier and ice sheet models, our focus here is on the attribution of
33   large-scale changes in glaciers and ice sheets. Land ice in the form of glaciers has been included in CMIP
34   climate and Earth system models as components of the land surface models for many years. However, their
35   representation is simplified and is omitted altogether in the less complex modelling systems. In CMIP3
36   (Meehl et al., 2007) and CMIP5 (Taylor et al., 2012) land ice area fraction, a component of land surface
37   models, was defined as a time-independent quantity, and in most model configurations was preset at the
38   simulation initialization as a permanent land feature. In CMIP6 considerable progress has been made in
39   improving and evaluating the representation of modelled land ice. For glaciers, an example is the expansion
40   of the Joint UK Land Environment Simulator (JULES) land surface model to enable elevated tiles, and hence
41   more accurately simulate the altitudinal atmospheric effects on glaciers (Shannon et al., 2019). Moreover,
42   standalone glacier models have now been systematically compared in GlacierMIP (Hock et al., 2019a;
43   Marzeion et al., 2020). The Antarctic and Greenland ice sheets were absent in global climate models that
44   pre-date CMIP6 (Eyring et al., 2016a), however some preliminary analyses that used results from CMIP5 to
45   drive standalone ice sheet models were included in AR5 (Church et al., 2013b). For the first time in CMIP,
46   the latest CMIP6 phase includes a coordinated effort to simulate temporally evolving ice sheets within the
47   Ice Sheet Model Intercomparison Project (ISMIP6; Box 9.3; Nowicki et al., 2016). Our understanding of
48   aspects of the global water storage contained in glaciers and ice sheets, and their contribution to sea-level
49   rise, has improved since the AR5 and SROCC (Hock et al., 2019b; Meredith et al., 2019) both in models and
50   observations (see assessment of observations and model evaluation for the Greenland Ice Sheet in Sections
51 and 9.4.1; Antarctica in Sections and 9.4.2; and glaciers in Sections and 9.5.1).
54 Glaciers
     Do Not Cite, Quote or Distribute                     3-52                                      Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   Glaciers are defined as perennial surface land ice masses independent of the Antarctic and Greenland ice
 2   sheets (Sections 9.5, AR5 assessed that anthropogenic influence had likely contributed to the retreat
 3   of glaciers observed since the 1960s (Bindoff et al., 2013), based on a high level of scientific understanding
 4   and robust estimates of observed mass loss, internal variability, and glacier response to climatic drivers.
 5   SROCC (Hock et al., 2019b) concluded that atmospheric warming was very likely the primary driver of
 6   glacier recession.
 8   Simulations of glacier mass changes under climate change rely on glacier models driven by climate model
 9   output, often in collaborative research efforts such as GlacierMIP (Hock et al., 2019a; Marzeion et al., 2020).
10   The GlacierMIP project is a systematic coordinated modelling effort designed to further understanding of
11   glacier loss using global models. While the low resolution and remaining biases of climate model-derived
12   boundary forcing data is a limitation, the release of the Randolph Glacier Inventory (Pfeffer et al., 2014; RGI
13   Consortium, 2017) has supported more sophisticated, systematic and comprehensive modelling of glaciers
14   worldwide (Hock et al., 2019a).
16   A regional study considering 85 NH glacier systems concluded that there is a discernible human influence on
17   glacier mass balance, with glacier model simulations driven by CMIP5 historical and greenhouse gas-only
18   simulations showing a glacier mass loss, whereas those driven by natural-only forced simulations showed a
19   net glacier growth (Hirabayashi et al., 2016). In addition, a study of the role of climate change in glacier
20   retreat using a simple mass-balance model for 37 glaciers worldwide, concluded that observed length
21   changes would not have occurred without anthropogenic climate change, with observed length variations
22   exceeding those associated with internal variability by several standard deviations in many cases (Roe et al.,
23   2017). Roe et al. (2020) used the same model to estimate that at least 85% of cumulative glacier mass loss
24   since 1850 is attributable to anthropogenic influence. While Marzeion et al. (2014) found that anthropogenic
25   influence contributed only 25 ± 35% of glacier mass loss for the period 1851-2010, their naturally-forced
26   simulations exhibited a substantial negative mass balance, which Roe et al. (2020) argued is unrealistic.
27   Moreover, Marzeion et al. (2014) estimated that anthropogenic influence contributed 69 ± 24% of glacier
28   mass loss for the period 1991 to 2010, consistent with a progressively increasing fraction of mass loss
29   attributable to anthropogenic influence found by Roe et al. (2020).
31   In summary, considering together the SROCC assessment that atmospheric warming was very likely the
32   primary driver of glacier recession, the results of Roe et al. (2017, 2020) and our assessment of the dominant
33   role of anthropogenic influence in driving atmospheric warming (Section 3.3.1), we conclude that human
34   influence is very likely the main driver of the near-universal retreat of glaciers globally since the 1990s.
37 Ice Sheets
39   Greenland Ice Sheet
40   AR5 assessed that it is likely that anthropogenic forcing contributed to the surface melting of the Greenland
41   Ice Sheet since 1993 (Bindoff et al., 2013). SROCC did not directly assess the attribution of Greenland Ice
42   Sheet change to anthropogenic forcing, but it did assess with medium confidence that summer melting of the
43   Greenland Ice Sheet has increased to a level unprecedented over at least the last 350 years, which is two-to-
44   fivefold the pre-industrial level (see also Trusel et al.; 2018).
46   Section assesses that Greenland ice sheet mass loss began in the latter half of the 19th century and
47   the rate of loss has increased substantially since the turn of the 21st century (high confidence), and also notes
48   that integration of proxy evidence and modelling indicates that the last time the rate of mass loss was similar
49   to the 20th century rate was during the early Holocene. Models of Greenland Ice Sheet evolution are
50   evaluated in detail in Section, which assesses that there is overall medium confidence in these models.
51   Model evaluation of surface mass balance changes over the Greenland Ice Sheet, including regional aspects,
52   is also assessed in Atlas.11.2.3.
54   Detection and attribution studies of change in the Greenland Ice Sheet remain challenging (Kjeldsen et al.,
55   2015; Bamber et al., 2019). This is in part due to the short observational record (Shepherd et al., 2012, 2018,
     Do Not Cite, Quote or Distribute                      3-53                                       Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   2020; Bamber et al., 2018; Cazenave et al., 2018; Mouginot et al., 2019; Rignot et al., 2019) and the
 2   challenges this poses to the evaluation of modelling efforts (Section 9.4.1). The latter require not only
 3   dynamic ice sheet models, but also appropriate atmospheric and oceanic conditions to use as a boundary
 4   forcing to drive the models (Nowicki and Seroussi, 2018; Barthel et al., 2020). Nonetheless, new literature
 5   since AR5 finds that ice sheet mass balance calculations using reanalysis-driven regional model simulations
 6   of surface mass balance, are found to agree well with the observed decrease in ice sheet mass over the past
 7   twenty years (Fettweis et al., 2020; Sasgen et al., 2020a; Tedesco and Fettweis, 2020a), consistent with
 8   earlier studies (Flato et al., 2013). These studies also show that the exceptional melt events observed in 2012
 9   and 2019 were associated with exceptional atmospheric conditions (Sasgen et al., 2020a; Tedesco and
10   Fettweis, 2020b). These results support the finding that increased surface melting is associated with
11   warming, although atmospheric circulation anomalies, including the summer North Atlantic Oscillation
12   (NAO) and variations in snowfall play an important role in driving interannual variations (Sasgen et al.,
13   2020; Tedesco and Fettweis, 2020; Section Further, a coupled ice-sheet-climate model study found
14   emergence of decreased surface mass balance prior to the present day in coastal locations in Greenland,
15   which dominate the integrated surface mass balance (Fyke et al., 2014), suggesting that observed variations
16   in surface mass balance in these regions might be expected to be distinguishable from internal variability. A
17   CMIP6 simulation of the historical period showed stable Greenland surface mass balance up to the 1990s,
18   after which it declined due to increased melt and runoff, consistent with a downscaled reanalysis
19   (van Kampenhout et al., 2020). Further, all experts surveyed in a structured expert judgement exercise
20   examining the causes of the increase in mass loss from the Greenland Ice Sheet over the last two decades
21   (Bamber et al., 2019) concluded that external forcing was responsible for at least 50% of the mass loss. A
22   comparison of Greenland Ice Sheet mass loss trends from observations and AR5 model projections for the
23   period 2007-2017 found that observed surface mass balance trends were at the top of the AR5 assessed
24   range, while mass loss due to changing ice dynamics was near the centre of the AR5 range (Slater et al.,
25   2020), providing further evidence of consistent anthropogenically-forced mass loss trends in models and
26   observations.
28   Drawing together the evidence from the continued and strengthened observed mass loss, the agreement
29   between anthropogenically forced climate simulations and observations, and historical and paleo evidence
30   for the unusualness of the observed rate of surface melting and mass loss, we assess that it is very likely that
31   human influence has contributed to the observed surface melting of the Greenland Ice Sheet over the past
32   two decades, and there is medium confidence in an anthropogenic contribution to recent overall mass loss
33   from Greenland.
35   Antarctic Ice Sheet
36   AR5 assessed that there was low confidence in attributing the causes of the observed mass loss from the
37   Antarctic ice sheet since 1993 (Bindoff et al., 2013). SROCC assessed that there is medium agreement but
38   limited evidence of anthropogenic forcing of Antarctic mass balance through both surface mass balance and
39   glacier dynamics. It further assessed that Antarctic ice loss is dominated by acceleration, retreat and rapid
40   thinning of the major West Antarctic Ice Sheet outlet glaciers (very high confidence), driven by melting of
41   ice shelves by warm ocean waters (high confidence). Based on updated observations, Chapter 2 assesses that
42   there is very high confidence that the Antarctic Ice Sheet lost mass between 1992 and 2017, and that there is
43   medium confidence that this mass loss has accelerated. Models of Antarctic Ice Sheet evolution are evaluated
44   in detail in Section, which assesses that there is medium confidence in many ice sheet processes
45   inAntarctic ice sheet models, but low confidence in the ocean forcing affecting basal melt rates. CMIP5 and
46   CMIP6 models perform similarly in their simulation of Antarctic surface mass balance (Section,
47   Gorte et al., 2020). Model evaluation of surface mass balance over the Antarctic Ice Sheet, including
48   regional aspects, is also assessed in Atlas.11.1.3.
50   Ice discharge around the West Antarctic Ice Sheet is strongly influenced by variability in basal melt (Jenkins
51   et al., 2018; Hoffman et al., 2019), in particular at decadal and longer time scales (Snow et al., 2017). Basal
52   melt rate variability can be induced by wind-driven ocean current changes, which may partly be of
53   anthropogenic origin via greenhouse gas forcing (Holland et al., 2019). Moreover, ice discharge losses from
54   the Antarctic Ice Sheet over the 2007-2017 period are close to the centre of the model-based range projected
55   in AR5 (Slater et al., 2020). However, expert opinion differs as to whether recent Antarctic ice loss from the
     Do Not Cite, Quote or Distribute                      3-54                                       Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   West Antarctic Ice Sheet has been driven primarily by external forcing or by internal variability, and there is
 2   no consensus (Bamber et al., 2019). Anthropogenic influence on the Antarctic surface mass balance, which
 3   is expected to partially compensate for ice discharge losses through increases in snowfall, is currently
 4   masked by strong natural variability (Previdi and Polvani, 2016; Bodart and Bingham, 2019), and
 5   observations suggest that it has been close to zero over recent years (Slater et al., 2020) (see further
 6   discussion in Section
 8   Overall, there is medium agreement but limited evidence of anthropogenic influence on Antarctic mass
 9   balance through changes in ice discharge.
12   3.5     Human Influence on the Ocean
14   The global ocean plays an important role in the climate system, as it is responsible for transporting and
15   storing large amounts of heat (Sections 3.5.1 and, freshwater (Sections 3.5.2 and and carbon
16   (Sections 3.6.2 and that are exchanged with the atmosphere. Therefore, accurate ocean simulation in
17   climate models is essential for the realistic representation of the climatic response to anthropogenic warming,
18   including rates of warming, sea level rise and carbon uptake, and the representation of coupled modes of
19   climate variability.
21   Ocean model development has advanced considerably since AR5 (Section Ongoing model
22   developments since AR5 have focused on improving the realism of the simulated ocean in coupled models,
23   with horizontal resolutions increasing to 10-100 km (from about 200 km in CMIP5), and increased vertical
24   resolutions in many modelling systems of 0-1 m for near-surface levels (from the highest resolution of 10 m
25   in CMIP5). These developments are aimed at improving the representation of the diurnal cycle and coupling
26   to the atmosphere (e.g. Bernie et al. 2005, 2007, 2008). General improvements to simulated ocean fidelity in
27   response to increasing resolution are expected (Hewitt et al., 2017). The effects of model resolution on the
28   fidelity of ocean models are discussed in more detail in Sections 9.2.2 and 9.2.4.
30   In this section we assess the global and basin-scale properties of the simulated ocean, with a focus on
31   evaluation of the realism of simulated ocean properties, and the detection and attribution of human-induced
32   changes in the ocean over the period of observational coverage. Observed changes to ocean temperature
33   (Section, salinity (Section, sea level (Section and ocean circulation (Section
34   are reported in Chapter 2. A more process-based assessment of ocean changes, alongside the assessment of
35   variability and changes in ocean properties with spatial scales smaller than ocean basin scales is presented in
36   Chapter 9.
39   3.5.1    Ocean Temperature
41   Ocean temperature and ocean heat content are key physical variables considered for climate model
42   evaluation and are primary indicators of a changing ocean climate. This section assesses the performance of
43   climate models in representing the mean state ocean temperature and heat content (Section, with a
44   particular focus on the tropical oceans given the importance of air-sea coupling in these areas (Section
45 This is followed by an assessment of detection and attribution studies of changes in ocean
46   temperature and heat content (Section Changes in global surface temperature are assessed in Section
50 Sea Surface and Zonal Mean Ocean Temperature Evaluation
52   In CMIP3 and CMIP5 models, large SST biases were found in the mid and high latitudes (Flato et al., 2013).
53   In CMIP6, the NH mid-latitude surface temperature biases appear to be marginally improved in the multi-
54   model mean when contrasted to CMIP5 despite large biases remaining in a few models (Figure 3.23a, Figure
55   3.24). There is a decreased spread of the zonal mean SST error between 50°N and 30°S, relative to CMIP5
     Do Not Cite, Quote or Distribute                       3-55                                       Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   (Figure 3.24a). On the other hand, the Southern Ocean’s warm surface temperature bias remains (Figure
 2   3.23a; Beadling et al., 2020), and is on average larger in CMIP6 than in CMIP5 models (Figure 3.23a, Figure
 3   3.24). This warm bias is often associated with persistent overlying atmospheric cloud biases (Hyder et al.,
 4   2018). Several other large biases also appear to remain largely unresolved in CMIP6, particularly warm
 5   biases in excess of 1°C along the equatorial eastern continental boundaries of the tropical Atlantic and
 6   Pacific Oceans (Figure 3.23a).
 8   Overall, the simulated and observed trends in SST patterns are generally consistent for the historical period
 9   (Olonscheck et al., 2020). The CMIP6 models generally represent the observed pattern of trends better than
10   the CMIP5 models, and observed trends fall within the range of simulated trends over a larger area for
11   CMIP6 models than for CMIP5 models (Olonscheck et al., 2020).
13   The CMIP5 multi-model mean zonally averaged subsurface ocean temperature showed warm biases between
14   200 m and 2000 m (mid-depth) over most latitudes, with exceptions in the Southern Ocean (>60°S, 100-
15   2000 m) and upper (0-400 m) Arctic Ocean. Cold biases were simulated near the surface (0-200 m) at most
16   latitudes (Flato et al., 2013). CMIP6 biases are broadly consistent with those reported in CMIP5 for the near-
17   surface (<200 m) and mid-depth (200-2000 m) ocean (Voldoire et al., 2019b; Beadling et al., 2020; Zhu et
18   al., 2020b). The warm bias begins between 100 and 400 m depth in all three basins, however, it is most
19   prominent in the Atlantic Ocean, with a maximum magnitude in the equatorial latitudes, as in CMIP5 (Figure
20   3.25). In the Pacific, the large warm biases are mostly seen in the subtropical regions (30-60°N and 30-
21   60°S). The cool near surface tropical bias is most prominent in the Pacific Ocean and also present in the
22   Atlantic, with a smaller magnitude (Figure 3.25). Relative to CMIP5, the most prominent difference is an
23   increase to the mid-depth (300-2000 m) warm bias in CMIP6 relative to CMIP5, and a change in sign of the
24   bias from cold to warm for the Southern Ocean mid-depth (>60oS) from CMIP5 to CMIP6 (Figure 3.25).
25   Compared to CMIP3 and CMIP5, there is improved agreement between most CMIP6 models and
26   observations in their representation of the zonal mean temperature of the upper 100 m of the Southern Ocean
27   (Beadling et al., 2020).
29   Focusing on the deep ocean (>2000 m), the CMIP6 ensemble mean shows a prominent and consistent warm
30   bias (Figure 3.25), in all basins except the equatorial and northern Pacific, which contrasts to a cold bias seen
31   in CMIP5 (Flato et al., 2013). We note that while an updated observational temperature dataset is used in this
32   assessment (WOA09 was used in AR5, while WOA18, 1981-2010 is used in AR6), the deep ocean warm
33   bias remains and is approaching double the magnitude (~0.5°C) of the equivalent CMIP5 multi-model mean
34   bias, a feature which is particularly prominent in the Atlantic and southern Indian Oceans. Increased
35   horizontal resolution as well as the choice of the vertical coordinate are reported to partly improve these
36   biases in some models (Adcroft et al., 2019; Rackow et al., 2019; Hewitt et al., 2020).
38   Since AR5, there has been growing evidence that the representation of mean surface and deeper ocean
39   temperatures in coupled climate models can be improved by increasing the horizontal resolution both in the
40   ocean and the atmosphere (e.g. Small et al., 2014; Hewitt et al., 2016; Iovino et al., 2016; Roberts et al.,
41   2019). At an ocean resolution of around 1°, which is typical of CMIP6 models, some processes are
42   parameterized rather than explicitly resolved, leading to a compromise in their dynamical representation. An
43   increase in the model resolution allows for processes to be explicitly resolved, and can for example, enhance
44   the simulation of eddies, thus improving simulated vertical eddy transport, and reducing temperature drifts in
45   the deeper ocean (Griffies et al., 2015; von Storch et al., 2016). For some models, the mean absolute error in
46   ocean temperature below 500 m is smaller in the high resolution version compared to the low resolution
47   version, particularly in eddy-active regions such as the North Atlantic (Rackow et al., 2019). Increasing the
48   horizontal resolution of individual climate models often leads to an overall decrease in the surface
49   temperature biases over regions where they persisted through earlier CMIP generations, such as the central
50   and western equatorial Pacific, as well as the North and tropical Atlantic (Figure 3.3e; Roberts et al., 2019;
51   Hewitt et al., 2020). Despite this, as a group the four HighResMIP models included in Figures 3.3e and 3.24
52   do not on average show smaller SST biases than the CMIP6 multi-model mean, demonstrating the
53   importance of factors other than resolution in contributing to SST biases.
55   In summary, there is little improvement in the multi-model mean sea surface and zonal mean ocean
     Do Not Cite, Quote or Distribute                      3-56                                       Total pages: 202
     Final Government Distribution                           Chapter 3                                        IPCC AR6 WGI

 1   temperatures from CMIP5 to CMIP6 (medium confidence). Nevertheless, the CMIP6 models show a
 2   somewhat more realistic pattern of SST trends (low confidence).
 7   Figure 3.23: Multi-model-mean bias of (a) sea surface temperature and (b) near-surface salinity, defined as the
 8                difference between the CMIP6 multi-model mean and the climatology from the World Ocean Atlas
 9                2018. The CMIP6 multi-model mean is constructed with one realization of 46 CMIP6 historical
10                experiments for the period 1995–2014 and the climatology from the World Ocean Atlas 2018 is an
11                average over all available years (1955-2017). Uncertainty is represented using the advanced approach: No
12                overlay indicates regions with robust signal, where ≥66% of models show change greater than variability
13                threshold and ≥80% of all models agree on sign of change; diagonal lines indicate regions with no change
14                or no robust signal, where <66% of models show a change greater than the variability threshold; crossed
15                lines indicate regions with conflicting signal, where ≥66% of models show change greater than the
16                variability threshold and <80% of all models agree on sign of change. For more information on the
17                advanced approach, please refer to the Cross-Chapter Box Atlas.1. Further details on data sources and
18                processing are available in the chapter data table (Table 3.SM.1).
20   [END FIGURE 3.23 HERE]
25   Figure 3.24: Biases in zonal mean and equatorial sea surface temperature (SST) in CMIP5 and CMIP6 models.
26                CMIP6 (red), CMIP5 (blue) and HighResMIP (green) multi-model mean (a) zonally-averaged SST bias;
27                (b) equatorial SST bias; and (c) equatorial SST compared to observed mean SST (black line) for 1979-
28                1999. The inter-model 5th and 95th percentiles are depicted by the respective shaded range. Model
29                climatologies are derived from the 1979-1999 mean of the historical simulations, using one simulation
30                per model. The Hadley Centre Sea Ice and Sea Surface Temperature version 1 (HadISST) (Rayner et al.,
31                2003) observational climatology for 1979-1999 is used as the reference for the error calculation in (a) and
32                (b); and for observations in (c). Further details on data sources and processing are available in the chapter
33                data table (Table 3.SM.1).
35   [END FIGURE 3.24 HERE]
40   Figure 3.25: CMIP6 potential temperature and salinity biases for the global ocean, Atlantic, Pacific and Indian
41                Oceans. Shown in colour are the time-mean differences between the CMIP6 historical multi-model
42                climatological mean and observations, zonally averaged for each basin (excluding marginal and regional
43                seas). The observed climatological values are obtained from the World Ocean Atlas 2018 (WOA18,
44                1981-2010; Prepared by the Ocean Climate Laboratory, National Oceanographic Data Center, Silver
45                Spring, MD, USA), and are shown as labelled black contours for each of the basins. The simulated annual
46                mean climatologies for 1981 to 2010 are calculated from available CMIP6 historical simulations, and the
47                WOA18 climatology utilized synthesized observed data from 1981 to 2010. A total of 30 available
48                CMIP6 models have contributed to the temperature panels (left column) and 28 models to the salinity
49                panels (right column). Potential temperature units are °C and salinity units are the Practical Salinity Scale
50                1978 [PSS-78]. Further details on data sources and processing are available in the chapter data table
51                (Table 3.SM.1).
53   [END FIGURE 3.25 HERE]
56 Tropical Sea Surface Temperature Evaluation
58   Tropical Pacific Ocean
     Do Not Cite, Quote or Distribute                          3-57                                           Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   In CMIP5, mean state biases in the tropical Pacific Ocean including the excessive equatorial cold tongue,
 2   erroneous mean thermocline depth and slope along the equator remained but were improved relative to
 3   CMIP3 (Flato et al., 2013). Misrepresentation of the interaction between the atmosphere and ocean via the
 4   Bjerknes feedback and vertical mixing parameterizations, and a bias in winds were among the suggested
 5   reasons for the persistent biases (Li et al., 2014; Zhu and Zhang, 2018). Moving to CMIP6, a reduction of the
 6   cold bias in the equatorial cold tongue in the central Pacific is found on average in the CMIP6 models
 7   (Figure 3.24b; Grose et al., 2020; Planton et al., 2020), however, this reduced bias is not statistically
 8   significant when considered across the multi-model ensemble (Planton et al., 2020). It is also noteworthy that
 9   the longitude of the 28°C isotherm is closer to observed in CMIP6 than in CMIP5, with a coincident
10   reduction in the CMIP6 inter-model standard deviation (Grose et al., 2020). The latter result implies that
11   there is an improvement in the representation of the tropical Pacific mean state in CMIP6 models.
12   Comparison of biases in individual HighResMIP models with biases in lower resolution versions of the same
13   models indicates that there is no consistent improvement in SST biases in most of the equatorial Pacific with
14   resolution (Figure 3.3e; Bock et al., 2020).
16   Tropical Atlantic Ocean
17   Fundamental features such as the mean zonal SST gradient in the tropical Atlantic were not reproduced in
18   CMIP5 models. Studies have proposed that weaker than observed alongshore winds, underestimation of
19   stratocumulus clouds, coarse model resolution, and insufficient oceanic cooling due to a deeper thermocline
20   depth and weak vertical velocities at the base of the mixed layer in the eastern basin, underpinned these
21   tropical Atlantic SST gradient biases (Hourdin et al., 2015; Richter, 2015). The SST gradient biases still
22   remain in CMIP6. On average the cold bias in the western part of the basin is reduced while the warm bias in
23   the eastern part has slightly increased (Figure 3.24 b,c) (Richter and Tokinaga, 2020a). Several CMIP6
24   models, however, display large reductions in biases of the zonal SST gradient, such that the eastern
25   equatorial Atlantic warm SST bias and associated westerly wind biases are mostly eliminated in these
26   models (Richter and Tokinaga, 2020a). The high resolution (HighResMIP) CMIP6 models show a better
27   representation of the zonal SST gradient (Figure 3.24 b,c), but some lower resolution models also perform
28   well suggesting that resolution is not the only factor responsible for biases in Tropical Atlantic SST (Richter
29   and Tokinaga, 2020a).
31   Tropical Indian Ocean
32   The tropical Indian Ocean mean state is reasonably well simulated both in CMIP5 and CMIP6 (Figure 3.24
33   b,c). However, CMIP5 models show a large spread in the thermocline depth, particularly in the equatorial
34   part of the basin (Saji et al., 2006; Fathrio et al., 2017a), which has been linked to the parameterization of the
35   vertical mixing and the wind structure, leading to a misrepresentation of the ventilation process in some
36   models (Schott et al., 2009; Richter, 2015; Shikha and Valsala, 2018). A common problem with the CMIP5
37   models is therefore a warm bias in the subsurface, mainly at depths around the thermocline, which is also
38   apparent in the CMIP6 models (Figure 3.25g).
40   In the CMIP6 multi-model mean, the western tropical Indian Ocean shows a slightly larger warm bias
41   compared to CMIP5 (Figure 3.24 b,c), which in part could be related to excessive supply of warm water
42   from the Red Sea (Grose et al., 2020; Semmler et al., 2020). The HighResMIP models show decreases in
43   SST bias across the Indian Ocean with increasing resolution (Figure 3.3e; Bock et al., 2020), though as a
44   group the SST biases in the HighResMIP models are no smaller than those of the full CMIP6 ensemble.
46   In summary, the structure and magnitude of multi-model mean ocean temperature biases have not changed
47   substantially between CMIP5 and CMIP6 (medium confidence). Although biases remain in the latest model
48   generation models, the broad consistency between the observed and simulated basin-scale ocean properties
49   suggests that CMIP5 and CMIP6 models are appropriate tools for investigating ocean temperature and ocean
50   heat content responses to forcing. This also provides high confidence in the utility of CMIP-class models for
51   detection and attribution studies, for both ocean heat content (Section and thermosteric sea level
52   applications (Section 3.5.3).
55 Ocean Heat Content Change Attribution
     Do Not Cite, Quote or Distribute                       3-58                                       Total pages: 202
     Final Government Distribution                          Chapter 3                                       IPCC AR6 WGI

 2   The ocean plays an important role as the Earth’s primary energy store. The AR5 and SROCC assessed that
 3   the ocean accounted for more than 90% of the Earth’s energy change since the 1970s (Rhein et al., 2013;
 4   Bindoff et al., 2019). These assessments are consistent with recent studies assessed in Section 7.2 and Cross-
 5   Chapter Box 9.1, which find that 91% of the observed change in Earth’s total energy from 1971 to 2018 was
 6   stored in the ocean (von Schuckmann et al., 2020). The AR5 concluded that anthropogenic forcing has very
 7   likely made a substantial contribution to ocean warming above 700 m, whereas below 700 m, limited
 8   measurements restricted the assessment of ocean heat content changes in AR5 and prevented a robust
 9   comparison between observations and models (Bindoff et al., 2013).
11   With the recent increase in ocean sampling by Argo to 2000 m (Roemmich et al., 2015; Riser et al., 2016;
12   von Schuckmann et al., 2016) and the resulting improvements in estimates of ocean heat content (Abraham
13   et al., 2013; Balmaseda et al., 2013; Durack et al., 2014a; Cheng et al., 2017; von Schuckmann et al., 2020),
14   a more quantitative assessment of the global ocean heat content changes that extends into the intermediate
15   ocean (700-2000 m) over the more recent period (from 2005 to the present) (Durack et al., 2018) can be
16   performed. Observed ocean heat content changes are discussed in Section, where it is reported that it
17   is virtually certain that the global upper ocean (0-700 m) and very likely that the global intermediate ocean
18   (700-2000 m) warmed substantially from 1971 to the present. Further, ocean layer warming contributions are
19   reported as 61% (0-700 m), 31% (700-2000 m) and 8% (>2000 m) for the 1971 to 2018 period (Table 2.7).
20   CMIP5 model simulations replicate this partitioning fairly well for the industrial-era (1865 to 2017)
21   throughout the upper (0-700 m, 65%), intermediate (700-2000 m, 20%) and deep (>2000 m, 15%) layers
22   (Gleckler et al., 2016; Durack et al., 2018). The corresponding warming percentages for the multi-model
23   mean of a subset of CMIP6 simulations over the 1850-2014 period are 58% for the upper, 21% for the
24   intermediate, and 22% for the deep ocean layers (Figure 3.26). These results are consistent with SROCC
25   which assessed that it is virtually certain that both the upper and intermediate ocean warmed from 2004 to
26   2016, with an increased rate of warming since 1993 (Bindoff et al., 2019). The spatial distribution of these
27   changes for different ocean depths is assessed in Section 9.2.2.
32   Figure 3.26: Global ocean heat content in CMIP6 simulations and observations. Time series of observed (black)
33                and simulated (red) global ocean heat content anomalies with respect to 1995-2014 for the full ocean
34                depth (left panel), upper layer - 0 to 700 m (top right panel), intermediate layer -700 to 2000 m (middle
35                right panel), and the abyssal ocean >2000 m (bottom right panel). The best estimate observations (black
36                solid line) for the period of 1971-2018, along with very likely ranges (black shading) are from Section
37       For the models (1860-2014), ensemble members from 15 CMIP6 models are used to calculate the
38                multi-model mean values (red solid line) after averaging across simulations for each independent model.
39                The very likely ranges in the simulations are shown in red shading. Simulation drift has been removed
40                from all CMIP6 historical runs using a contemporaneous portion of the linear fit to each corresponding
41                pre-industrial control run (Gleckler et al., 2012). Units are 1021 Joules (Zettajoules). Further details on
42                data sources and processing are available in the chapter data table (Table 3.SM.1).
44   [END FIGURE 3.26 HERE]
47   The multi-model means of both CMIP5 and CMIP6 historical simulations forced with time varying natural
48   and anthropogenic forcing showsrobust increases in ocean heat content in the upper (0-700 m) and
49   intermediate (700-2000 m) ocean (high confidence) (Figure 3.26) (Cheng et al., 2016, 2019; Gleckler et al.,
50   2016; Bilbao et al., 2019; Tokarska et al., 2019). Temporary (<10 years) surface and subsurface cooling
51   during and after large volcanic eruptions are also captured in the upper-ocean, and global mean ocean heat
52   content (Balmaseda et al., 2013). The ocean heat content increase is also reflected in the corresponding
53   ocean thermal expansion which is a leading contributor to global mean sea level rise (Sections and
54   9.2.4; Box 9.1).
56   For the period 1971-2014, the rate of ocean heat uptake for the global ocean in the CMIP6 models is about
     Do Not Cite, Quote or Distribute                         3-59                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   6.43 [2.08-8.66] ZJ yr-1, with the upper, intermediate and deeper layers respectively accounting for 68%,
 2   16% and 16% of the full depth global heat uptake (Figure 3.26). Overall, the simulated ocean heat content
 3   changes are consistent with the updated and improved observational analyses, within the very likely
 4   uncertainty range defined for each (Domingues et al., 2008; Purkey and Johnson, 2010; Levitus et al., 2012;
 5   Good et al., 2013; Cheng et al., 2017; Ishii et al., 2017; Zanna et al., 2019) (see also Table 2.7, Section
 6 as well as with the ocean components of total Earth heating assessed in Section, Table 7.1.
 7   Nevertheless, large uncertainties remain, particularly in the deeper layers due to the poor temporal and
 8   spatial sampling coverage, particularly in the Atlantic, Southern and Indian Oceans (Garry et al., 2019). The
 9   very likely ranges of the simulated trends for the full ocean depth and below 2000 m fall within the very
10   likely range of observed uptake during the last two decades. In the intermediate layer, the multi-model
11   ensemble mean mostly stays above (below) the observed 5th-95th percentile range before (after) the year
12   2000. For the upper ocean, some individual model realisations show a reduced ocean heat content increase
13   during the 1970s and 1980s, which is then compensated by a greater warming than the observations from the
14   early 1990s. These discrepancies have been linked with a temporary increase in the Southern Ocean deep
15   water formation rate, as well as with the models’ strong aerosol cooling effects and high equilibrium climate
16   sensitivity (Andrews et al., 2019, 2020; Golaz et al., 2019; Dunne et al., 2020; Winton et al., 2020) (see also
17   Section 7.5.6 and Box 7.2).
19   Nevertheless, both the observed estimates and simulations show that the rate of ocean heat uptake has
20   doubled in the past few decades, when contrasted to the rate over the complete 20th century (Figure 3.26),
21   with over a third of the accumulated heat stored below 700 m (Cheng et al., 2016, 2019; Gleckler et al.,
22   2016; Durack et al., 2018). The Southern Ocean shows the strongest ocean heat uptake that penetrates to
23   deeper layers (Section, whereas ocean heat content increases in the Pacific and Indian Oceans
24   largely occur in the upper layers (Bilbao et al., 2019).
26   Since AR5, the attribution of ocean heat content increases to anthropogenic forcing has been further
27   supported by more detection and attribution studies. These studies have shown that contributions from
28   natural forcing alone cannot explain the observed changes in ocean heat content in either the upper or
29   intermediate ocean layers, and a response to anthropogenic forcing is clearly detectable in ocean heat content
30   (Gleckler et al., 2016; Bilbao et al., 2019; Tokarska et al., 2019). Moreover, a response to greenhouse gas
31   forcing is detectable independently of the response to other anthropogenic forcings (Bilbao et al., 2019;
32   Tokarska et al., 2019), which has offset part of the greenhouse gas induced warming. Further evidence is
33   provided by the agreement between observed and simulated changes in global thermal expansion associated
34   with the ocean heat content increase when both natural and anthropogenic forcings are included in the
35   simulations (Section 3.5.3), though internal variability plays a larger role in driving basin-scale thermosteric
36   sea level trends (Bilbao et al., 2015). Over the Southern Ocean, warming is detectable over the late twentieth
37   century and is largely attributable to greenhouse gases (Swart et al., 2018; Hobbs et al., 2021), while other
38   anthropogenic forcings such as ozone depletion have been shown to mitigate the warming in some of the
39   CMIP5 simulations (Swart et al., 2018; Hobbs et al., 2021). The use of the mean temperature above a fixed
40   isotherm rather than fixed depth further strengthens a robust detection of the anthropogenic response in the
41   upper ocean (Weller et al., 2016), and better accounting for internal variability in the upper ocean (Rathore et
42   al., 2020), helps explain reported hemispheric asymmetry in ocean heat content change (Durack et al.,
43   2014a).
45   In summary, there is strong evidence for an improved understanding of the observed global ocean heat
46   content increase. It is extremely likely that human influence was the main driver of the ocean heat content
47   increase observed since the 1970s, which extends into the deeper ocean (very high confidence). Updated
48   observations, like model simulations, show that warming extends throughout the entire water column (high
49   confidence).
52   3.5.2   Ocean Salinity
54   While ocean assessments have primarily focused on temperature changes, improved observational salinity
55   products since the early 2000s have supported more assessment of long-term ocean salinity change and
     Do Not Cite, Quote or Distribute                      3-60                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   variability from AR4 (Bindoff et al., 2007) to AR5 across both models and observations (Flato et al., 2013;
 2   Rhein et al., 2013). AR5 assessed that it was very likely that anthropogenic forcings have made a discernible
 3   contribution to surface and subsurface oceanic salinity changes since the 1960s. The SROCC augmented
 4   these insights, noting that observed high latitude freshening and warming have very likely made the surface
 5   ocean less dense with a stratification increase of between 2.18% and 2.42% from 1970 to 2017 (Bindoff et
 6   al., 2019). A recent observational analysis has expanded on these assessments, suggesting a very marked
 7   summertime density contrast enhancement across the mixed layer base of 6.2% to 11.6% per decade, driven
 8   by changes in temperature and salinity, which is more than six times larger than previous estimates (Sallée et
 9   al., 2021). An idealised ocean modelling study suggests that the enhanced stratification can account for a
10   third of the salinity enhancement signal since 1990 (Zika et al., 2018). Thus, there has been an expansion of
11   observed global- and basin-scale salinity change assessment literature since AR5, with many new studies
12   reproducing the key patterns of long-term salinity change reported in AR5 (Rhein et al., 2013), and linking
13   these through modelling studies to coincident changes in evaporation-precipitation patterns at the ocean
14   surface (Sections 2.3.1, 3.3.2, and 9.2.2).
16   Unlike SSTs, simulated sea surface salinity (SSS) does not provide a direct feedback to the atmosphere.
17   However, some recent work has identified indirect radiative feedbacks through sea-salt aerosol interactions
18   (Ayash et al., 2008; Amiri-Farahani et al., 2019; Wang et al., 2019b) that can act to strengthen tropical
19   cyclones, and increase precipitation (Balaguru et al., 2012, 2016; Grodsky et al., 2012; Reul et al., 2014;
20   Jiang et al., 2019). The absence of a direct feedback is one of the primary reasons why salinity simulation is
21   difficult to constrain in ocean modelling systems, and why deviations from the observed near-surface salinity
22   mean state between models and observations are often apparent (Durack et al., 2012; Shi et al., 2017).
25 Sea Surface and Depth-profile Salinity Evaluation
27   When compared to the routine assessment of simulated SST, simulated SSS has not received the same
28   research attention at global- to basin-scales. For CMIP3, there was reasonable agreement between the basin-
29   scale patterns of salinity, with a comparatively fresher Pacific when contrasted to the salty Atlantic, and
30   basin salinity maxima features aligning well with the corresponding atmospheric evaporation minus
31   precipitation field (Durack et al., 2012). Similar features are also reproduced in CMIP5 along with realistic
32   variability in the upper layers, but less variability than observations at 300 m and deeper, especially in the
33   poorly sampled Antarctic region (Pierce et al., 2012). In a regional study, only considering the Indian Ocean,
34   CMIP5 SSS was assessed and it was shown that model biases were primarily linked to biases in the
35   precipitation field, with ocean circulation biases playing a secondary role (Fathrio et al., 2017b). The sea
36   surface salinity bias in CMIP6 models is shown in Figure 3.23b.
38   For the first time in AR5, alongside global zonal mean temperature, global zonal mean salinity bias with
39   depth was assessed for the CMIP5 models. This showed a strong upper ocean (<300 m) negative salinity
40   (fresh) bias of order 0.3 PSS-78, with a tendency toward a positive salinity (salty) bias (<0.25 PSS-78) in the
41   NH intermediate layers (200-3000 m) (Flato et al., 2013). These biases are also present in CMIP6, albeit
42   with slightly smaller magnitudes (Figure 3.25). Here we expand the global zonal mean bias assessment to
43   consider the three independent ocean basins individually, which allows for an assessment as to which basin
44   biases are dominating the global zonal mean. The basin with the most pronounced biases is the Atlantic, with
45   a strong upper ocean (<300 m) fresh bias, of order 0.3 PSS-78 just like the global zonal mean, and a marked
46   subsurface salinity bias that exceeds 0.5 PSS-78 in equatorial waters between 400–1000 m.
48   The Pacific Ocean shares the strongest similarity to the global bias, with a similar upper ocean (<300 m)
49   fresh bias, of a smaller magnitude than the Atlantic, with a pronounced off-equator feature at intermediate
50   depths (500-1500 m). Lower magnitude positive salinity biases (~0.3 PSS-78) are also present in both
51   hemispheres between 200-3000 m, and deeper in the SH (Figure 3.25). The Indian Ocean shows similar
52   features to the SH Pacific, with a marked upper ocean (<500 m) fresh bias of order 0.3 PSS-78, and a strong
53   near-surface positive bias of order 0.4 PSS-78 associated with the Arabian Sea (Figure 3.25).
55   For the Southern Ocean in CMIP5, considerable fresh biases exist through the water column, and are most
     Do Not Cite, Quote or Distribute                     3-61                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                     IPCC AR6 WGI

 1   pronounced in the ventilated layers representing the subtropical mode and intermediate water masses (Sallée
 2   et al., 2013). A fresh bias in upper and intermediate layers of comparable magnitude is also seen in CMIP6
 3   (Figure 3.25). The structure of the biases in the CMIP6 multi-model mean (which averages across many
 4   simulations with differing subsurface geographies and differing Southern Ocean salinity biases (Beadling et
 5   al., 2020)) is similar to that evident in the CMIP5 multi-model mean, but with slightly smaller magnitudes.
 6   The Arctic Ocean also on average exhibits a surface-enhanced fresh bias in the upper ocean (Figure 3.25),
 7   which is much larger than its SH counterpart.
 9   In summary, the structure of the salinity biases in the multi-model mean has not changed substantially
10   between CMIP5 and CMIP6 (medium confidence), though there is limited evidence that the magnitude of
11   subsurface biases has been reduced. This provides confidence in the utility of CMIP-class models for
12   detection and attribution of ocean salinity.
15 Salinity Change Attribution
17   AR5 concluded that it was very likely that anthropogenic forcings had made a discernible contribution to
18   surface and subsurface ocean salinity changes since the 1960s (Bindoff et al., 2013; Rhein et al., 2013). They
19   highlighted that the spatial patterns of salinity trends, and the mean fields of salinity and evaporation minus
20   precipitation are all similar, with an enhancement to Atlantic Ocean salinity and freshening in the Pacific and
21   Southern Oceans. Since AR5 all subsequent work on assessing observed and modelled salinity changes has
22   confirmed these results.
24   Considerable changes to observed broad- or basin-scale ocean near-surface salinity fields have been reported
25   (see Section, and these have been linked to changes in the evaporation minus precipitation patterns
26   at the ocean surface through model simulations, typically expressing a pattern of change where
27   climatological mean fresh regions become fresher and corresponding salty regions becoming saltier (Durack
28   et al., 2012, 2013; Zika et al., 2015; Lago et al., 2016; Skliris et al., 2016, 2018; Cheng et al., 2020), also
29   broadly present in the CMIP6 multi-model mean (Figure 3.27). At basin-scales, the depth-integrated effect of
30   mean salinity changes as captured in halosteric sea level for the top 0 to 2000 m has also been assessed based
31   on observational products, and these results mirror near-surface patterns in the CMIP5 and CMIP6 models,
32   with most areas that are becoming fresher at the surface exhibiting increases in halosteric sea level, and areas
33   becoming saltier exhibiting decreases (Durack et al., 2014b; Figure 3.28). Further investigations using
34   observations and models together have tied the long-term patterns of surface and subsurface salinity changes
35   to coincident changes to the evaporation minus precipitation field over the ocean (Durack et al., 2012, 2013;
36   Durack, 2015; Levang and Schmitt, 2015; Zika et al., 2015, 2018; Grist et al., 2016; Lago et al., 2016; Cheng
37   et al., 2020), however the rate of these changes through time continues to be an active area of active research
38   (Skliris et al., 2014; Zika et al., 2015, 2018; Cheng et al., 2020; Sallée et al., 2021).
40   Climate change detection and attribution assessments have considered salinity, with the first of these
41   assessed in AR5 (Bindoff et al., 2013). Since this time, the positive detection conclusions (Stott et al., 2008;
42   Pierce et al., 2012; Terray et al., 2012) have been supported by a number of more recent and independent
43   assessments which have reproduced the multi-decadal basin-scale patterns of change in observations and
44   models (Figures 3.27 and 3.28; Durack, 2015; Durack et al., 2014b; Levang and Schmitt, 2015; Skliris et al.,
45   2016). Observed depth-integrated basin responses, contrasting the Pacific and Atlantic basins (freshening
46   Pacific and enhanced salinity Atlantic) were also shown to be replicated in most historical (natural and
47   anthropogenically forced) simulations, with this basin contrast absent in CMIP5 and CMIP6 natural-only
48   simulations that exclude anthropogenic forcing (Durack et al., 2014b; Figure 3.28).
50   While observational sparsity considerably limits quantification of regional changes, a recent study by
51   Friedman et al. (2017) assessed salinity changes in the Atlantic Ocean from 1896 to 2013 and confirmed the
52   pattern of mid-to-low latitude enhanced salinity and high latitude North Atlantic freshening over this period
53   exists even after accounting for the effects of the NAO and AMO.
55   Considering the bulk of evidence, it is extremely likely that human influence has contributed to observed
     Do Not Cite, Quote or Distribute                      3-62                                       Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 1   near-surface and subsurface salinity changes across the globe since the mid-20th century. All available multi-
 2   decadal assessments have confirmed that the associated pattern of change corresponds to fresh regions
 3   becoming fresher and salty regions becoming saltier (high confidence). CMIP5 and CMIP6 models are only
 4   able to reproduce these patterns in simulations that include greenhouse gas increases (medium confidence).
 5   Changes to the coincident atmospheric water cycle and ocean-atmosphere fluxes (evaporation and
 6   precipitation) are the primary drivers of the basin-scale observed salinity changes (high confidence). This
 7   result is supported by all available observational assessments, along with a growing number of climate
 8   modelling studies targeted at assessing ocean and water cycle changes. The basin-scale changes are
 9   consistent across models and intensify on centennial scales from the historical period through to the
10   projections of future climate (high confidence).
15   Figure 3.27: Maps of multi-decadal salinity trends for the near-surface ocean. Units are Practical Salinity Scale
16                1978 [PSS-78] decade-1. (top) The best estimate (Section observed trend (Durack and Wijffels,
17                2010 updated, 1950-2019). (bottom) Simulated trend from the CMIP6 historical experiment multi-model
18                mean (1950-2014). Black contours show the climatological mean salinity in increments of 0.5 PSS-78
19                (thick lines 1 PSS-78). Further details on data sources and processing are available in the chapter data
20                table (Table 3.SM.1).
22   [END FIGURE 3.27 HERE]
27   Figure 3.28: Long-term trends in halosteric and thermosteric sea level in CMIP6 models and observations. Units
28                are mm year-1. In the right column, three observed maps of 0 to 2000 m halosteric sea level trends (right
29                column) top from (Durack and Wijffels, 2010, 1950-2019 updated - D&W), middle from (Good et al.,
30                2013, 1950-2019 updated- EN4), and lower from (Ishii et al., 2017, 1955-2019 updated – Ishii), and
31                bottom, the CMIP6 historical multi-model mean (1950-2014). Red and orange colours show a halosteric
32                contraction (enhanced salinity) and blue and green a halosteric expansion (reduced salinity). In the left
33                column, basin-integrated halosteric (top) and thermosteric (bottom) trends for the Atlantic and Pacific, the
34                two largest ocean basins, where Pacific anomalies are presented on the x-axis and Atlantic on the y-axis.
35                Observational estimates are presented in black, CMIP6 historical (all forcings) simulations are shown in
36                orange squares, with the multi-model mean shown as a dark orange diamond with a black bounding box.
37                CMIP6 hist-nat (historical natural forcings only) simulations are shown in green squares with the multi-
38                model mean as a dark green diamond with a black bounding box. Further details on data sources and
39                processing are available in the chapter data table (Table 3.SM.1).
41   [END FIGURE 3.28 HERE]
44   3.5.3    Sea Level
46   In keeping with the scope of this chapter, this section addresses global and basin-scale sea level changes,
47   whereas regional and local sea level changes are assessed in Section 9.6. In the AR5, the observed sea level
48   budget was closed by considering all contributing factors including ocean warming, mass contributions from
49   terrestrial storage, glaciers, and the Antarctic and Greenland ice sheets (Church et al., 2013a). The SROCC
50   found that the observed global mean sea level (GMSL) rise is consistent within uncertainties with the sum of
51   the estimated observed contributions for 1993–2015 and 2006–2015.
54 Sea Level Evaluation
56   The current generation of climate models do not fully resolve many of the components required to close the
57   observed sea level budget, such as glaciers, ice sheets and land water storage (see Section 9.6 and Box 9.1).
     Do Not Cite, Quote or Distribute                          3-63                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   Consequently, most CMIP-based analyses of sea level change have focused on thermosteric sea level
 2   changes (i.e., thermal expansion due to warming) and ocean dynamic sea level change, both of which are
 3   simulated in the CMIP5-generation of models. The improved agreement between modelled thermal
 4   expansion and observed estimates during the historical period led the SROCC to assess a high confidence
 5   level in the simulated thermal expansion using climate models and high confidence in their ability to project
 6   future thermal expansion.
 8   Since CMIP5 models do not include all necessary components of sea level change, this gap has been bridged
 9   by using offline models (for glacier melt and ice-sheet surface mass balance) driven by reanalyses and model
10   output. Some studies have used offline mass inputs to account for dynamic ice sheet and terrestrial
11   contributions. Slangen et al. (2017) and Meyssignac et al. (2017) suggested including corrections to several
12   contributions to sea level changes including to the Greenland surface mass balance and glacier contributions,
13   based on differences between CMIP5-driven model results and reanalysis-driven results. This helps close the
14   gap between models and observations for the 20th century globally, as well as providing better agreement
15   with tide gauge observations in terms of interannual and multi-decadal variability at the regional scale.
17   In CMIP6, ice sheets (see Section and 9.4) are included for the first time in ISMIP6 (Nowicki et al.,
18   2016). There is also scope for new insights into terrestrial water contributions from land surface (and sub-
19   surface) modelling in the Land Surface, Snow and Soil moisture Model Intercomparison Project (LS3MIP;
20   van den Hurk et al., 2016). In parallel, the GlacierMIP project (Hock et al., 2019a; Marzeion et al., 2020; see
21   Section and 9.5) is also underway, and has provided more quantitative guidance and a comprehensive
22   assessment of the uncertainties and best estimates of the current and future contributions of glaciers to the
23   sea level budget.
26 Sea Level Change Attribution
28   The SROCC concluded with high confidence that the dominant cause of global mean sea level rise since
29   1970 is anthropogenic forcing. Prior to that, the AR5 had concluded that it is very likely that there has been a
30   substantial contribution from anthropogenic forcings to global mean sea level rise since the 1970s. Since the
31   AR5, several studies have identified a human contribution to observed sea level change resulting from a
32   warming climate as manifest in thermosteric sea level change and the contribution from melting glaciers and
33   ice sheets.
35   For the global mean thermosteric sea level change, Slangen et al. (2014) showed the importance of
36   anthropogenic forcings (combined greenhouse gas and aerosol forcings) for explaining the magnitude of the
37   observed changes between 1957-2005, considering the full depth of the ocean and natural forcings in order
38   to capture the variability (see also Figure 3.29). Over the 1950-2005 period, Marcos and Amores (2014)
39   found that human influence explains 87% of the 0-700 m global thermosteric sea level rise. Both
40   thermosteric and regional dynamic patterns of sea level change in individual forcing experiments from
41   CMIP5 were considered by Slangen et al. (2015) who showed that responses to anthropogenic forcings are
42   significantly different from both internal variability and inter-model differences and that although
43   greenhouse gas and anthropogenic aerosol forcings produce opposite global mean sea level responses, there
44   are differences in response on regional scales. Based on these studies, we conclude that it is very likely that
45   anthropogenic forcing was the main driver of the observed global mean thermosteric sea level change since
46   1970.
48   In an attribution study of the sea-level contributions of glaciers, Marzeion et al. (2014) found that between
49   1991 and 2010, the anthropogenic fraction of global glacier mass loss was 69 ±24% (see also Section
50 Slangen et al. (2016) considered all quantifiable components of the global mean sea level budget
51   and showed that anthropogenically forced changes account for 69 ± 31% of the observed sea level rise over
52   the period 1970 to 2005, whereas natural forcings combined with internal variability have a much smaller
53   effect - only contributing 9 ± 18% of the change over the same period. These studies indicate that about 70%
54   of the combined change in glaciers, ice sheet surface mass balance and thermal expansion since 1970 can be
55   attributed to anthropogenic forcing, and that this percentage has increased over the course of the 20th century.
     Do Not Cite, Quote or Distribute                      3-64                                      Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 1   Detection studies on GMSL change in the 20th century (Becker et al., 2014; Dangendorf et al., 2015) found
 2   that observed total GMSL change in the 20th century was inconsistent with internal variability. Dangendorf et
 3   al. (2015) determined that for 1900 to 2011 at least 45% of GMSL change is human-induced. A study that
 4   developed a semi-empirical model to link sea-level change to observed GMST change concluded that at least
 5   41% of the 20th century sea-level rise would not have happened in the absence of the century’s increasing
 6   GMST and that there was a 95% probability that by 1970 GMSL was higher than that which would have
 7   occurred in the absence of increasing GMST (Kopp et al., 2016). Richter et al. (2020) compared modelled
 8   sea level change with the satellite altimeter observations from 1993 to 2015; a period short enough that
 9   internal variability can dominate the spatial pattern of change. They found that when GMSL is not removed,
10   model simulated zonally averaged sea level trends are consistent with altimeter observations globally as well
11   as in each ocean basin and much larger than might be expected from internal variability. Using spatial
12   correlation, Fasullo and Nerem (2018) showed that the satellite altimeter trend pattern is already detectable.
14   We note that current detection and attribution studies do not yet include all processes that are important for
15   sea-level change (see Section 9.6). However, based on the body of literature available, we conclude that the
16   main driver of the observed GMSL rise since at least 1970 is very likely anthropogenic forcing.
21   Figure 3.29: Simulated and observed global mean sea level change due to thermal expansion for CMIP6 models
22                and observations relative to the baseline period 1850-1900. Historical simulations are shown in brown,
23                natural only in green, greenhouse gas only in grey, and aerosol only in blue (multi-model means shown as
24                thick lines, and shaded ranges between the 5th and 95th percentile). The best estimate observations (black
25                solid line) for the period of 1971-2018, along with very likely ranges (black shading) are from Section
26       and are shifted to match the multi-model mean of the historical simulations for the 1995-2014
27                period. Further details on data sources and processing are available in the chapter data table (Table
28                3.SM.1).
30   [END FIGURE 3.29 HERE]
33   3.5.4   Ocean Circulation
35   Circulation of the ocean, whether it be wind or density driven, plays a prominent role in the heat and
36   freshwater transport of the Earth system (Buckley and Marshall, 2016). Thus, its accurate representation is
37   crucial for the realistic representation of water mass properties, and replication of observed changes driven
38   by atmosphere-land-ocean coupling. Here, we assess the ability of CMIP models to reproduce the observed
39   large-scale ocean circulation, along with assessment of the detection and attribution of any
40   anthropogenically-driven changes. We also note that the process-based understanding of these circulation
41   changes and circulation changes occurring at smaller scales are assessed in Section 9.2.3.
44 Atlantic Meridional Overturning Circulation (AMOC)
46   The Atlantic Meridional Overturning Circulation (AMOC) represents a large-scale flow of warm salty water
47   northward at the surface and a return flow of colder water southward at depth. As such, its mean state plays
48   an important role in transporting heat in the climate system, while its variability can act to redistribute heat
49   (see Sections and for more details). Paleo-climatic and model evidence suggest that changes
50   in AMOC strength have played a prominent role in past transitions between warm and cool climatic phases
51   (e.g., Dansgaard et al., 1993; Ritz et al., 2013)
53   AR5 concluded that while climate models suggested that an AMOC slowdown would occur in response to
54   anthropogenic forcing, the short direct observational AMOC record precluded it from being used to support
55   this model finding. Chapter 2 reports with high confidence, a weakening of the AMOC was observed in the
56   mid-2000s to the mid-2010s, while again also noting that the observational record was too short to determine
     Do Not Cite, Quote or Distribute                         3-65                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   whether this is a significant trend or a manifestation of decadal and multi-decadal variability (Section
 2 Indirect evidence of AMOC weakening since at least the 1950s is also presented, but confidence
 3   in this longer-term decrease was low (Section
 5   Despite the additional six years or so of observations since the AR5, the evaluation of the AMOC in models
 6   continues to be severely hampered by the geographically sparse and temporally short observational record.
 7   The longest continuous observational estimates of the AMOC are based on measurements taken at 26°N by
 8   the RAPID-MOCHA array (Smeed et al., 2018). Basic evaluation of the AMOC at 26°N shows that the
 9   CMIP5 and CMIP6 multi-model mean overturning strength is comparable with RAPID (Reintges et al.,
10   2017; Weijer et al., 2020), but the model range is large (12-29 Sv for CMIP5, Zhang and Wang (2013); and
11   10-31 Sv for CMIP6, Weijer et al. (2020)) (Figure 3.30a). It is noted that deviations of AMOC strength in
12   CMIP5 models have been related to global-scale sea surface temperature biases (Wang et al., 2014a). Both
13   coupled and ocean-only models also underestimate the depth of the AMOC cell (Danabasoglu et al., 2014;
14   Weijer et al., 2020; Figure 3.30a). Paleo-climatic evidence has also raised questions regarding the accuracy
15   of the representation of the strength and depth of the modelled AMOC during past periods (Otto-Bliesner et
16   al., 2007; Muglia and Schmittner, 2015). Overall, however, both the CMIP5 and CMIP6 model ensembles
17   simulate the general features of the AMOC mean state reasonably well, but there is a large spread in the
18   latitude and depth of the maximum overturning, and the maximum AMOC strength (Figure 3.30a).
20   The short length of the observed time-series (RAPID has measured the AMOC since 2004), sparse
21   observations, observational uncertainties (Sinha et al., 2018), as well as significant observed variability on
22   interannual and longer time scales, makes comparison with modelled AMOC variability challenging. RAPID
23   observations show that the overturning at 26°N was 2.9 Sv weaker in the multi-year average of 2008-2012
24   relative to 2004-2008 and 2.5 Sv weaker in 2012-2017 relative to 2004-2008 (Smeed et al., 2014, 2018) (see
25   also Section As expected, this weakening was accompanied by a significant reduction in
26   northward heat transport (Bryden et al., 2020). CMIP5 and CMIP6 models produce a forced weakening of
27   the AMOC over the 2012-2017 period relative to 2004-2008, but at 26°N the multi-model mean response is
28   substantially weaker than the observed AMOC decline over the same period. The discrepancy between the
29   modelled multi-model mean (i.e. the forced response) and the RAPID observed AMOC changes has led
30   studies to suggest that the observed weakening over 2004-2017 is largely due to internal variability (Yan et
31   al., 2018). However, comparison of observed RAPID AMOC variability with modelled variability also
32   reveals that most CMIP5 models appear to underestimate the interannual and decadal timescale AMOC
33   variability (Roberts et al., 2014; Yan et al., 2018), and, although the overall variance is larger in CMIP6 than
34   in CMIP5, similar results are found analysing the CMIP6 models (Figure 3.30b,c). It is currently unknown
35   why most models underestimate this AMOC variability, or whether they are underestimating the internal or
36   externally forced components. This underestimation of AMOC variability may also have potential
37   implications for detection and attribution, the relationship between AMOC and AMV (see Section 3.7.7),
38   and near-term predictions. There is also emerging evidence, based on analysis of freshwater transports, that
39   the AMOC in CMIP5-era models is too stable, largely due to systematic biases in ocean salinity (Liu et al.,
40   2017a; Mecking et al., 2017). Such a systematic bias may potentially be linked with the underestimation of
41   both simulated AMOC internal variability through eddy-mean flow interactions that are poorly represented
42   in standard CMIP-class model resolution (Leroux et al., 2018), and externally forced change.
44   As reported in Section, estimates of AMOC since at least 1950, which are generated from observed
45   surface temperatures or sea surface height, suggest the AMOC weakened through the 20th century (low
46   confidence) (Ezer et al., 2013; Caesar et al., 2018). Over the same period, the CMIP5 multi-model mean
47   showed no significant net forced response in AMOC (Cheng et al., 2013). However, a significant forced
48   change is simulated in the CMIP6 multi-model mean, where a clear increase of the AMOC is seen over the
49   1940-1985 period (Menary et al., 2020) (Figure 3.30e). Although there is general agreement that the
50   influence of greenhouse gases acts to a weaken the modelled AMOC (Delworth and Dixon, 2006; Caesar et
51   al., 2018), changes in solar, volcanic and anthropogenic aerosol emissions can lead to temporary changes in
52   AMOC on decadal-to-multidecadal timescales (Delworth and Dixon, 2006; Menary et al., 2013; Menary and
53   Scaife, 2014; Swingedouw et al., 2017; Undorf et al., 2018b). As such, the simulated net forced response in
54   AMOC is often a balance between the different forcing factors (Delworth and Dixon, 2006; Menary et al.,
55   2020) (Section The differing AMOC response of CMIP5 and CMIP6 models during the historical
     Do Not Cite, Quote or Distribute                      3-66                                      Total pages: 202
     Final Government Distribution                            Chapter 3                                         IPCC AR6 WGI

 1   period has been associated with stronger aerosol effective radiative forcing in the CMIP6 models (Menary et
 2   al., 2020), such that the aerosol-induced AMOC increase during the 1940-1985 period overcomes the
 3   greenhouse gas induced decline (Figure 3.30e). However, models simulate a range of anthropogenic aerosol
 4   effective radiative forcing and a range of historical AMOC trends in CMIP6 (Menary et al., 2020) and there
 5   remains considerable uncertainty over the realism of the CMIP6 AMOC response during the 20th century
 6   (Figure 3.30d-f) due to disagreement among the differing lines of evidence. For example, ocean reanalysis
 7   (Jackson et al., 2019) and forced ocean model simulations (Robson et al., 2012; Danabasoglu et al., 2016),
 8   which show AMOC changes that are broadly consistent with the CMIP6 response, appear to disagree with
 9   observational estimates of AMOC over the historical period (Ezer et al., 2013; Caesar et al., 2018). It is
10   noted, however, that the relatively short length of the forced ocean simulations and ocean reanalysis
11   precludes a comparable assessment of 20th century trends. Furthermore, despite the similar AMOC evolution
12   seen in forced ocean model simulations and the CMIP6 models, it is unclear whether the same underlying
13   mechanisms are responsible for the changes.
15   In summary, models do not support robust assessment of the role of anthropogenic forcing in the observed
16   AMOC weakening between the mid-2000s and the mid-2010s, which is assessed to have occurred with high
17   confidence in Section, as the changes are outside of the range of modelled AMOC trends
18   (regardless of whether they are forced or internally generated) in most models. Thus, we have low confidence
19   that anthropogenic forcing has influenced the observed changes in AMOC strength in the post-2004 period.
20   In addition to this, there remains considerable uncertainty over the realism of the CMIP6 AMOC response
21   during the 20th century due to disagreement among the differing lines of observational and modelled
22   evidence (i.e., historical AMOC estimates, ocean reanalysis, forced ocean simulations and historical CMIP6
23   simulations). Thus, we have low confidence that anthropogenic forcing has had a significant influence on
24   changes in AMOC strength during the 1860-2014 period.
29   Figure 3.30: Observed and CMIP6 simulated AMOC mean state, variability and long-term trends. (a) AMOC
30                meridional streamfunction profiles at 26.5°N from the historical CMIP5 (1860-2004) and CMIP6 (1860-
31                2014) simulations compared with the mean maximum overturning depth (horizontal grey line) and
32                magnitude (vertical grey line) from the RAPID observations (2004-2018). The distributions of model
33                ranges of AMOC maximum magnitude and depth are respectively displayed on the x- and y-axis. (b)
34                Distributions of overlapping 8-year AMOC trends from individual CMIP6 historical simulations (pink
35                box plots) are plotted along with the combined distributions of all available CMIP5 (blue boxplot) and
36                CMIP6 (red boxplot) models. For reference, the observed 8-year trend calculated between 2004-2012 is
37                also shown as a horizontal grey line (following Roberts et al., 2014) (c) Distributions of interannual
38                AMOC variability from individual CMIP6 model historical simulations, along with the combined
39                distributions of all available CMIP5 and CMIP6 models. Interannual variability in models and
40                observations are estimated as annual mean (April-March) differences, and the horizontal grey line is the
41                observed value for 2009/2010 minus 2008/2009 (following Roberts et al., 2014). (d-f) Distributions of
42                linear AMOC trends calculated over various time periods (see panel titles) in CMIP6 simulations forced
43                with: greenhouse gas forcing only (GHG), natural forcing only (NAT), anthropogenic aerosol forcing
44                only (AER) and all forcing combined (Historical; HIST). (a-f) Boxes indicate 25th to 75th percentile,
45                whiskers indicate 1st and 99th percentiles, and dots indicate outliers, while the horizontal black line is the
46                multi-model mean trend. In (d-f) the multi-model mean trend is also written above each distribution. The
47                multi-model distributions in (a-c) were produced with one historical ensemble member per model for
48                which the AMOC variable was available (listed), while those in (d-f) were produced with the AMOC
49                detection and attribution simulation data sets utilised by Menary et al. (2020). Further details on data
50                sources and processing are available in the chapter data table (Table 3.SM.1).
52   [END FIGURE 3.30 HERE]
55 Southern Ocean Circulation
57   The Southern Ocean circulation provides the principal connections between the world’s major ocean basins
     Do Not Cite, Quote or Distribute                           3-67                                           Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   through the circulation of the Antarctic Circumpolar Current (ACC), while also largely controlling the
 2   connection between the deep and upper layers of the global ocean circulation, through its upper and lower
 3   overturning cells.
 5   The assessment of observations presented in Sections and reports that there is no evidence
 6   of an ACC transport change, and it is unlikely that the mean meridional position of the ACC has moved
 7   southward in recent decades (Section and This is despite observations of surface wind
 8   displaying an intensification and southward shift (Section There is low confidence in an observed
 9   intensification of overturning in the Southern Ocean’s upper ocean and there is medium confidence for a
10   slowdown of the Antarctic Bottom Water circulation and commensurate Antarctic Bottom Water volume
11   decrease since the 1990s (Section Section presents new evidence, since the SROCC, which
12   assessed with medium confidence that the lower cell can episodically increase as a response to climatic
13   anomalies, temporally counteracting the forced tendency for reduced bottom water formation.
15   The modelled strength of the ACC clearly improved from CMIP3, in which the models tended to
16   underestimate the strength of the ACC, to CMIP5 (Meijers et al., 2012). This improvement in the realism of
17   ACC strength continues from CMIP5 to CMIP6, with the modelled ACC strength converging toward the
18   magnitude of observed estimates of net flow through the Drake Passage (Beadling et al., 2020). There is,
19   however, a small number of models that still display an ACC that is much weaker than that observed, while
20   several models also display much more pronounced ACC decadal variability than that observed (Beadling et
21   al., 2020). The increased realism of the ACC was at least partly related to noted improvements in all metrics
22   of the Southern Ocean’s surface wind stress forcing (Beadling et al., 2020). The most notable wind stress
23   forcing improvements were found in the strength and the latitudinal position of the zonally-averaged
24   westerly wind stress maximum (Beadling et al., 2020; Bracegirdle et al., 2020a). While the two-cell structure
25   of the overturning circulation appears to be well captured by CMIP5 models (Sallée et al., 2013; Russell et
26   al., 2018), they tend to underestimate the intensity of the lower cell overturning, and overestimate the
27   intensity of the upper cell overturning (Sallée et al., 2013). As the lower overturning cell is closely related to
28   Antarctic Bottom Water formation and deep convection, both fields also display substantial errors in CMIP5
29   models (Heuzé et al., 2013, 2015). CMIP6 climate models show clear improvements compared to CMIP5 in
30   their representation of Antarctic Bottom Water, which suggests an improved representation of the lower
31   overturning cell (Heuzé, 2021).
33   Despite notable improvements of CMIP6 models compared to CMIP5 models, inherent limitations in the
34   representation of important processes at play in the Southern Ocean’s horizontal and vertical circulation
35   remain (Section For instance, Southern Ocean mesoscale eddies are largely parameterised in the
36   current generation of climate models and despite their small spatial scales, they are a key element for
37   establishing the ACC and upper overturning cell, as well as for their future evolution under changing
38   atmospheric forcing (Kuhlbrodt et al., 2012; Downes and Hogg, 2013; Gent, 2016; Downes et al., 2018;
39   Poulsen et al., 2018). The absence of ice-sheet coupling in the CMIP6 model suite is another important
40   limitation, as basal meltwater and calving can influence the circulation, particularly the lower cell of the
41   Southern Ocean (Bronselaer et al., 2018; Golledge et al., 2019; Lago and England, 2019; Jeong et al., 2020a;
42   Moorman et al., 2020). We note that early development of global climate models with interactive ice-shelves
43   cavities has begun and is showing potential to be developed (Jeong et al., 2020b).
45   In summary, while there have been improvements across successive CMIP phases (from CMIP3 to CMIP6)
46   in the representation of the Southern Ocean circulation, such that the mean zonal and overturning
47   circulations of the Southern Ocean are now broadly reproduced, substantial observational uncertainty and
48   climate model challenges preclude attribution of Southern Ocean circulation changes (high confidence).
51   3.6     Human Influence on the Biosphere
53   3.6.1    Terrestrial Carbon Cycle
55   The AR5 did not make attribution statements on changes in global carbon sinks. The IPCC Special Report on
     Do Not Cite, Quote or Distribute                       3-68                                       Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   Climate Change and Land (SRCCL) assessed with high confidence that global vegetation photosynthetic
 2   activity has increased over the last 2-3 decades (Jia et al., 2019). That increase was attributed to direct land
 3   use and management changes, as well as to CO2 fertilisation, nitrogen deposition, increased diffuse radiation
 4   and climate change (high confidence). The AR5 assessed with high confidence that CMIP5 Earth System
 5   Models (ESMs) simulate the global mean land and ocean carbon sinks within the range of observation-based
 6   estimates (Flato et al., 2013). The IPCC SRCCL, however, noted the remaining shortcomings of carbon
 7   cycle schemes in ESMs (Jia et al., 2019), which for example do not properly incorporate thermal responses
 8   of respiration and photosynthesis, and frequently omit representations of permafrost thaw (Comyn-Platt et
 9   al., 2018), the nitrogen cycle (Thomas et al., 2015b) and its influence on vegetation dynamics (Jeffers et al.,
10   2015), the phosphorus cycle (Fleischer et al., 2019), and accurate implications of carbon store changes for a
11   range of land use and land management options (Erb et al., 2018; Harper et al., 2018) (see Sections
12   and 5.4, Figure 5.24 and Table 5.4 for details).
14   This section considers three main large-scale indicators of climate change relevant to the terrestrial carbon
15   cycle: atmospheric CO2 concentration, atmosphere-land CO2 fluxes, and leaf area index. These indicators
16   were chosen because they have been the target of attribution studies. Other indicators, like land use and
17   management, and wildfires, relate to human influence but are discussed in Chapter 5. Chapter 7 discusses
18   energetic consequences of changes in the terrestrial carbon cycle in Section CMIP5 and CMIP6
19   ESMs are most often run with prescribed observed historical changes in atmospheric CO2 concentration and
20   diagnose CO2 emissions consistent with these. Such calculations require that the models simulate realistic
21   changes in the terrestrial carbon cycle over the historical period, as changes to land carbon stores will
22   influence the size of CO2 emissions consistent with prescribed CO2 pathways, and associated remaining
23   carbon budgets (Section 5.5). Such testing of existing models is needed while also recognising there are
24   process representations still requiring inclusion.
26   Since the AR5, atmospheric inversion studies have further tested or constrained models, while new datasets
27   have been used to constrain specific parts of the terrestrial carbon cycle such as plant respiration
28   (Huntingford et al., 2017). Figure 3.31 compares historical emissions-driven CMIP6 simulations of global
29   mean atmospheric CO2 concentration and net ocean and land carbon fluxes to the assessed CO2
30   concentration and fluxes from the Global Carbon Project (Friedlingstein et al., 2019). For 2014, the CMIP6
31   models simulate a range of CO2 concentrations centred around the observed value of 397 ppmv, with a range
32   of 381 to 412 ppmv. GSAT anomalies simulated over the historical period are very similar in models that
33   simulate or prescribe changes in atmospheric CO2 concentrations (Figure 3.31b and Figure 3.4a). Most
34   models simulate realistic temporal evolution of the global net ocean and land carbon fluxes, although model
35   spread is larger over land (Figure 3.31c,d, see also Sections 3.6.2 and, Figure 5.24). Although
36   literature published soon after the AR5 highlighted the importance of representing nitrogen limitation on
37   plant growth (Peng and Dan, 2015; Thomas et al., 2015b), more recent studies note that models without
38   nitrogen limitation can still be consistent with the latest estimates of historical carbon cycle changes (Arora
39   et al., 2020; Meyerholt et al., 2020). Uncertainties in the photosynthetic response to atmospheric CO2
40   concentrations at global scales, shifts in carbon allocation and turnover, land-use change (Hoffman et al.,
41   2014; Wieder et al., 2019), and water limitation are also important influences on land carbon fluxes.
43   All models and observational estimates agree that interannual variability in net CO2 uptake is much larger
44   over land than over the ocean. Studies demonstrate that regional variations in both the trends and the yearly
45   strength of the terrestrial carbon sink are considerable. Land carbon uptake is dominated by the extratropical
46   northern latitudes (Ciais et al., 2019) (see also Section and Figure 5.25) because the tropics may have
47   become a net source of carbon (Baccini et al., 2017). At local to regional scales, the dominant driver of
48   yearly sink strength variations is water availability, but at continental to global scales, temperature anomalies
49   are the dominant driver (Section; Jung et al., 2017). The major role of levels of water stored in the
50   ground in influencing land-atmosphere CO2 exchange has also been confirmed through simultaneous
51   analysis of satellite gravimetry and atmospheric CO2 levels (Humphrey et al., 2018). When considered
52   globally, simulated land and ocean carbon sinks fall within the range of observation-based estimates with
53   high confidence. But there is also high confidence that that apparent success arises for the wrong reasons, as
54   models underestimate the NH carbon sink, as discussed in Section
     Do Not Cite, Quote or Distribute                      3-69                                       Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 4   Figure 3.31: Evaluation of historical emission-driven CMIP6 simulations for 1850-2014. Observations (black) are
 5                compared to simulations of global mean (a) atmospheric CO2 concentration (ppmv), with observations
 6                from the National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA
 7                ESRL) (Dlugokencky and Tans, 2020), (b) air surface temperature anomaly (°C) with respect to the 1850-
 8                1900 mean, with observations from HadCRUT4 (Morice et al., 2012) (c) land carbon uptake (PgC yr-1),
 9                (d) ocean carbon uptake (PgC yr-1), both with observations from the Global Carbon Project (GCP)
10                (Friedlingstein et al., 2019) and grey shading indicating the observational uncertainty. Land and ocean
11                carbon uptakes are plotted using a 10-year running mean for better visibility. The ocean uptake is offset
12                to 0 in 1850 to correct for pre-industrial riverine-induced carbon fluxes. Further details on data sources
13                and processing are available in the chapter data table (Table 3.SM.1).
15   [END FIGURE 3.31 HERE]
18   The seasonal cycle in atmospheric CO2, which is driven by the drawdown of carbon by photosynthesis on
19   land during the summer and release by respiration during the winter, has increased in amplitude since the
20   start of systematic monitoring (Figure 3.32, see also Section This trend, which is larger at higher
21   latitudes of the NH, was first reported by Keeling et al. (1996) and has continued. Changes in vegetation
22   productivity have also been observed, as well as longer growing seasons (Park et al., 2016). However, a
23   slowdown of the increasing trend has been noted, linked to a slowdown of both vegetation greening and
24   growing-season length increases (Buermann et al., 2018; Li et al., 2018b; Wang et al., 2020b). Figure 3.32
25   shows that CMIP6 terrestrial carbon cycle models partially capture the increasing amplitude of the seasonal
26   cycle of the land carbon sink, also seen in observational reconstructions. However, the identification of the
27   human influence that contributes most strongly to these changes in the seasonal cycle is debated.
29   Proposed causes of the trend in the amplitude of the seasonal cycle of CO2, and its amplification at higher
30   latitudes, include increases in the summer productivity and/or increases in the magnitude of winter
31   respiration of northern ecosystems (Barichivich et al., 2013; Graven et al., 2013a; Forkel et al., 2016; Wenzel
32   et al., 2016), increases in productivity throughout the NH by CO2 fertilisation, and increases in the
33   productivity of agricultural crops in northern mid-latitudes (Gray et al., 2014; Zeng et al., 2014). Recent
34   studies have attempted to quantify the different contributions by comparing atmospheric CO2 observations
35   with ensembles of land surface model simulations. Piao et al. (2017) found that CO2 fertilization of
36   photosynthesis is the main driver of the increase in the amplitude of the seasonal cycle of atmospheric CO2
37   but noted that climate change drives the latitudinal differences in that increase. North of 40°N, Bastos et al.
38   (2019) also found CO2 fertilization to be the most likely driver, with warming at northern high latitudes
39   contributing a decrease in amplitude, in contrast to earlier conclusions (Graven et al., 2013b; Forkel et al.,
40   2016), and agricultural and land use changes making only a small contribution. For temperate regions of the
41   NH, Wang et al. (2020b) found that the importance of CO2 fertilisation is decreased by drought stress, but
42   also found only a small contribution from agricultural and land use changes. However, many global models
43   do not include nitrogen fertilisation, changes to crop cultivars or irrigation effects, with the latter associated
44   with deficiencies in simulated terrestrial water cycling (Yang et al., 2018a). All these factors influence the
45   capability of models to simulate accurately the seasonal cycle in atmosphere-land CO2 exchanges. Model
46   comparisons to the atmospheric CO2 concentration record for Barrow, Alaska, suggest that models
47   underestimate current levels of carbon fixation (Winkler et al., 2019) and have deficiencies in their
48   phenological representation of greenness levels, particularly for autumn (Li et al., 2018c). Based on these
49   studies and noting the uncertainty in the processes ultimately driving changes in atmospheric CO2 seasonal
50   cycles (Section, we assess with medium confidence that fertilisation by anthropogenic increases in
51   atmospheric CO2 concentrations is the main driver of the increase in the amplitude of the seasonal cycle of
52   atmospheric CO2.
54   Detection and attribution methods have been applied to leaf area index, which represents “greenness” and
55   general photosynthetic productivity (see Section Nitrogen deposition and land cover change trends
56   remain small compared to variability, so attributing changes in leaf area index to those processes is difficult.
     Do Not Cite, Quote or Distribute                         3-70                                         Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1   Using three satellite products and ten land models, Zhu et al. (2016) found increases in leaf area index
 2   (greening) over 25-50% of global vegetated areas, and they attributed 70% of this greening to CO2
 3   fertilisation, although they found that land use change can dominate regionally. This is consistent with the
 4   attribution study of observed greening of Mao et al. (2016), and with Mao et al. (2013) who found that CO2
 5   fertilisation was the dominant cause of enhanced vegetation growth, with latitudinal changes in leaf area
 6   index explained by the larger land surface warming in the NH. These conclusions are also consistent with
 7   those of Zhu et al. (2017), who found a dominant role for CO2 fertilisation in driving leaf area index changes
 8   in an attribution study in which land models were first weighted by performance. However, Chen et al.
 9   (2019) has challenged these results by showing that greening in India and China was driven by land-use
10   change.
12   Leaf area index increases attributed to CO2 fertilisation are due to a direct raised physiological response.
13   However, for drylands, CO2-induced stomatal closure may act to conserve soil moisture and thereby
14   indirectly drive higher photosynthesis through higher water use efficiency (Lu et al., 2016). In models with
15   nitrogen deposition, there is evidence that this simulated effect also influences leaf area index trends,
16   however because of a lack of literature based on large-scale land simulations including both nutrient
17   limitation and crop intensification, it is not yet possible to make an attribution statement about their
18   individual roles in leaf area index changes.
20   In summary, Earth system models simulate globally averaged land carbon sinks within the range of
21   observation-based estimates (high confidence), but global-scale agreement masks large regional
22   disagreements. Based on new studies that attribute changes in atmospheric CO2 seasonal cycle to CO2
23   fertilisation, albeit counteracted by other factors, combined with the medium confidence that models
24   represent the processes driving changes in the seasonal cycle, we assess that there is medium confidence that
25   CO2 fertilisation is the main driver of the increase in the amplitude of the seasonal cycle of atmospheric CO2.
26   Based on available literature, CO2 fertilisation has been the main driver of the observed greening trend, but
27   there is only low confidence in this assessment because of ongoing debate about the relative roles of CO2
28   fertilisation, high latitude warming, and land management, and the low number of models that represent the
29   whole suite of processes involved.
34   Figure 3.32: Relative change in the amplitude of the seasonal cycle of global land carbon uptake in the historical
35                CMIP6 simulations from 1961-2014. Net biosphere production estimates from 19 CMIP6 models (red),
36                the data-led reconstruction JMA-TRANSCOM (Maki et al., 2010; dotted) and atmospheric CO2 seasonal
37                cycle amplitude changes from observations (global as dashed line, Mauna Loa Observatory (MLO)
38                (Dlugokencky et al., 2020) in bold black). Seasonal cycle amplitude is calculated using the curve fit
39                algorithm package from the National Oceanic and Atmospheric Administration Earth System Research
40                Laboratory (NOAA ESRL). Relative changes are referenced to the 1961-1970 mean and for short time
41                series adjusted to have the same mean as the model ensemble in the last 10 years. Interannual variation
42                was removed with a 9-year Gaussian smoothing. Shaded areas show the one sigma model spread (grey)
43                for the CMIP6 ensemble and the one sigma standard deviation of the smoothing (red) for the CO2 MLO
44                observations. Inset: average seasonal cycle of ensemble mean net biosphere production and its one sigma
45                model spread for 1961-1970 (orange dashed line, light orange shading) and 2005-2014 (solid green line,
46                green shading). Further details on data sources and processing are available in the chapter data table
47                (Table 3.SM.1).
49   [END FIGURE 3.32 HERE]
52   3.6.2   Ocean Biogeochemical Variables
54   Since CMIP5, there has been a general increase in ocean horizontal and vertical grid resolution in ocean
55   model components (Arora et al., 2020; Séférian et al., 2020). The latter of these developments is particularly
56   significant for projections of ocean stressors as it directly affects the representation of stratification. Updates
     Do Not Cite, Quote or Distribute                        3-71                                        Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   in the representation of ocean biogeochemical processes between CMIP5 and CMIP6 have typically
 2   involved an increase in model complexity. Specific developments have been the more widespread inclusion
 3   of micronutrients, such as iron, variable stoichiometric ratios, more detailed representation of lower trophic
 4   levels including bacteria and the cycling and sinking of organic matter. CMIP6 biogeochemical model
 5   performance is generally an improvement on that of the parent CMIP5 generation of models (Séférian et al.,
 6   2020). The global representation of present-day air-sea carbon fluxes and surface chlorophyll concentrations
 7   show moderate improvements between CMIP5 and CMIP6. Similar improvements are seen in the
 8   representation of subsurface oxygen concentrations in most ocean basins, while the representation of surface
 9   macronutrient concentrations in CMIP6 is shown to have improved with respect to silicic acid but declined
10   slightly with respect to nitrate. Model representation of the micronutrient iron has not improved substantially
11   since CMIP5, but many more models are capable of representing iron. In addition, a comparison of the
12   carbon concentration and carbon climate feedbacks shows no significant change between CMIP5 and CMIP6
13   (Arora et al., 2020).
15   Since AR5, research has also focused on the detection and attribution of regional patterns in ocean
16   biogeochemical change relating to interior de-oxygenation, air-sea CO2 flux, and ocean carbon uptake and
17   associated acidification. Characterization of flux variability requires understanding of the suite of physical
18   and biological processes including transport, heat fluxes, interior ventilation, biological production and gas
19   exchange which can have very different controls on seasonal versus interannual timescales in both the North
20   Pacific (Ayers and Lozier, 2012) and North Atlantic (Breeden and McKinley, 2016). In the Southern Ocean,
21   models have difficulty reproducing the observed seasonal cycle and interannual variability, making
22   attribution particularly challenging (Lovenduski et al., 2016; Mongwe et al., 2016, 2018).
24   The AR5 concluded that oxygen concentrations have decreased in the open ocean since 1960 and such
25   decreases can be attributed in part to human influences with medium confidence. The decrease in ocean
26   oxygen content in the upper 1000 m, between 1970-2010, is further confirmed in SROCC (medium
27   confidence), with the oxygen minimum zone expanding in volume (see also Section Observed
28   oxygen declines over the last several decades (Stendardo and Gruber, 2012; Stramma et al., 2012; Schmidtko
29   et al., 2017) match model estimates in the surface ocean (Oschlies et al., 2017) but are much larger than
30   model derived estimates in the interior (Bopp et al., 2013; Cocco et al., 2013). Some of this difference has
31   been interpreted as due to a lack of representation of coastal eutrophication in these models (Breitburg et al.,
32   2018), but much of it remains unexplained. This disparity is particularly true in the Eastern Pacific oxygen
33   minimum zone, where some CMIP5 models showed increasing trends whereas observations show a strong
34   decrease (Cabré et al., 2015). However, proxy reconstructions suggest that over the last century the ocean
35   may have in fact undergone increases in oxygen in the most oxygen poor regions (Deutsch et al., 2014). As
36   discussed in Section 5.3.1 ocean oxygen went through wide oscillations on multi-centennial timescales
37   through the last deglaciation, with abrupt warming resulting in loss of oxygen in subsurface waters of the
38   North Pacific (Praetorius et al., 2015). The global upper ocean oxygen inventoryis negatively correlated with
39   ocean heat content with a regression coefficient comparable to that found in ocean models (Ito et al., 2017).
40   Variability and trends in the observed upper ocean oxygen concentration are mainly driven by the apparent
41   oxygen utilization component with small contributions from oxygen solubility, suggesting that changing
42   ocean circulation, mixing, and/or biochemical processes, rather than thermally induced solubility effects may
43   be the main drivers of observed de-oxygenation. The spatial distribution of the ocean de-oxygenation in the
44   interior of the ocean as well as over coastal areas are further assessed in Section 5.3.
46   As one of the most commonly observed surface parameters, the partial pressure of CO2 has been the topic of
47   considerable detection and attribution work. In North Atlantic subtropical and equatorial biomes, warming
48   has been shown to be a significant and persistent contributor to the observed increase in the partial pressure
49   of CO2 since the mid‐2000s with long‐term warming leading to a reduction in ocean carbon uptake (Fay and
50   McKinley, 2013), and with both the partial pressure of CO2 and associated carbon uptake demonstrating
51   strong predictability as a function of interannual to decadal climate state (Li et al., 2016a; Li and Ilyina,
52   2018). In the Southern Ocean however, detection and attribution of surface trends in the partial pressure of
53   CO2 has proven more elusive and dependent on methodology, with some studies suggesting that Southern
54   Ocean carbon uptake slowed from about 1990 to 2006 and subsequently strengthened from 2007 to 2010
55   (Lovenduski et al., 2008; Fay et al., 2014; Ritter et al., 2017). Other studies have suggested that poor
     Do Not Cite, Quote or Distribute                      3-72                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   representation of the seasonal cycle in the Southern Ocean may confound models’ ability to represent
 2   changes in the partial pressure of CO2 in the Southern Ocean (Nevison et al., 2016; Mongwe et al., 2018).
 4   Section assesses that both observational reconstructions based on the partial pressure of CO2 and
 5   ocean biogeochemical models show a quasi-linear increase in the ocean sink of anthropogenic CO2 from 1.0
 6   ± 0.3 PgC yr-1 to 2.5 ± 0.6 PgC yr-1 between 1960–1969 and 2010–2019 in response to global CO2 emissions
 7   (high confidence). During the 1990s, the global net flux of CO2 into the ocean is estimated to have weakened
 8   to 0.8 ± 0.5 PgC yr-1 while in 2000 and thereafter, it is estimated to have strengthened considerably to rates
 9   of 2.0 ± 0.5 PgC yr-1, associated with changes in SST, the surface concentration of dissolved inorganic
10   carbon and alkalinity, and decadal variations in atmospheric forcing (Landschützer et al., 2016, see also
11   Section 5.2).
13   Ocean acidification is one of the most detectible metrics of environmental change and was well covered in
14   the AR5, in which it was assessed that the uptake of anthropogenic CO2 had very likely resulted in
15   acidification of surface waters (Bindoff et al., 2013). Since then, observations and simulations of
16   multidecadal trends in surface carbon chemistry have increased in robustness. The evidence on ocean pH
17   decline had further strengthened in SROCC with a good agreement found between CMIP5 models and
18   observations and an assessment that the ocean was continuing to acidify in response to ongoing carbon
19   uptake (Bindoff et al., 2019). An observed decrease in global ocean pH is assessed in Section to be
20   virtually certain to have occurred with a rate of 0.003-0.026 decade-1 at the surface for the past 40 years. The
21   ocean acidification has occurred not only in the surface layer but also in the interior of the ocean (Sections
22 and 5.3.3). Rates have been observed to be nearly as high (between −0.015 and −0.020 decade-1) in
23   mode and intermediate waters of the North Atlantic through the combined effect of increased anthropogenic
24   and remineralized carbon (Ríos et al., 2015) and acidification has been observed down to 3000 m in the deep
25   water formation regions (Perez et al., 2018). There has also been considerable improvement in detection and
26   attribution of anthropogenic CO2 versus eutrophication based acidification in coastal waters (Wallace et al.,
27   2014).
29   The increased evidence in recent studies supports an assessment that it is virtually certain that the uptake of
30   anthropogenic CO2 was the main driver of the observed acidification of the global ocean. The observed
31   increase in acidification over the North Atlantic subtropical and equatorial regions since 2000 is likely
32   associated in part with an increase in ocean temperature, a response which corresponds to the expected
33   weakening of the ocean carbon sink with warming. Due to strong internal variability, systematic changes in
34   carbon uptake in response to climate warming have not been observed in most other ocean basins at present.
35   We further assess, consistent with AR5 and SROCC, that deoxygenation in the upper ocean is due in part to
36   anthropogenic forcing, with medium confidence. There is high confidence that Earth system models simulate
37   a realistic time evolution of the global mean ocean carbon sink.
40   3.7     Human Influence on Modes of Climate Variability
42   This section assesses model evaluation and attribution of changes in the modes of climate variability listed in
43   Cross-Chapter Box 2.2, Table 2. The structure of the modes is described in Annex IV, observed changes in
44   the modes and associated teleconnections are assessed in Sections 2.4, and the role of the modes in shaping
45   regional climate is assessed in Section
48   3.7.1    North Atlantic Oscillation and Northern Annular Mode
50   The Northern Annular Mode (NAM; also known as Arctic Oscillation) is an oscillation of atmospheric mass
51   between the Arctic and northern mid-latitudes, analogous to the Southern Annular Mode (SAM; Section
52   3.7.2). It is the leading mode of variability of sea-level pressure in the northern extratropics but also has a
53   clear fingerprint through the troposphere up to the lower stratosphere, with maximum expression in boreal
54   winter (Kidston et al., 2015). The North Atlantic Oscillation (NAO) can be interpreted as the regional
55   expression of the NAM and captures most of the related variance in the troposphere over a broad North
     Do Not Cite, Quote or Distribute                      3-73                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   Atlantic/Europe domain. Indices measuring the state of the NAO correlate highly with those of the NAM,
 2   and teleconnection patterns for both modes are rather similar (Feldstein and Franzke, 2006). A detailed
 3   description of the NAM and the NAO as well as their associated teleconnection over land is given in Annex
 4   IV.2.1.
 6   AR5 found that while models simulated correctly most of the spatial properties of the NAM, substantial
 7   inter-model differences remained in the details of the associated teleconnection patterns over land (Flato et
 8   al., 2013). AR5 reported that most models did not reproduce the observed positive trend of the NAO/NAM
 9   indices during the second half of the 20th century. It was unclear to what extent this failure reflected model
10   shortcomings and/or if the observed trend could be simply related to pronounced internal climate variability.
11   AR5 accordingly did not make an attribution assessment for the NAO/NAM.
13   New studies since the AR5 continue to find that CMIP5 models reproduce the spatial structure and
14   magnitude of the NAM reasonably well (Lee and Black, 2013; Zuo et al., 2013; Davini and Cagnazzo, 2014;
15   Ying et al., 2014; Ning and Bradley, 2016; Deser et al., 2017b; Gong et al., 2017) although the North Pacific
16   SLP anomalies remain generally too strong (Zuo et al., 2013; Gong et al., 2017) and the subtropical North
17   Atlantic lobe of SLP anomalies conversely too weak (Ning and Bradley, 2016) in many models. Such overall
18   biases noted in both CMIP3 and CMIP5 (Davini and Cagnazzo, 2014) persist in CMIP6 historical
19   simulations, even though the multi-model multi-member ensemble mean spatial correlation between
20   modelled and observed NAM is slightly higher (Figure 3.33a,d,g). Regarding the NAO, the majority of
21   CMIP5 models very successfully simulate its spatial structure (Lee et al., 2019) and its associations with
22   extratropical jet, storm track and blocking variations over a broad North-Atlantic/Europe domain (Davini and
23   Cagnazzo, 2014) and over land through teleconnections (Volpi et al., 2020). The good performance of the
24   models is confirmed in CMIP6 with a marginal improvement of the averaged observation-model spatial
25   correlation (Figure 3.33b,e,h) and better skill based on other evaluation metrics (Fasullo et al., 2020). The
26   slight underestimation of the SLP anomalies related to the NAO centres of actions over the Azores and
27   Greenland-Iceland-Norwegian Seas remain unchanged compared to CMIP5.
29   CMIP5 models with a model top within the stratosphere seriously underestimate the amplitude of the
30   variability of the wintertime NAM expression in the stratosphere, in contrast to CMIP5 models which extend
31   well above the stratopause (Lee and Black, 2015). However, even in the latter models, the stratospheric
32   NAM events, and their downward influence on the troposphere, are insufficiently persistent (Charlton-Perez
33   et al., 2013; Lee and Black, 2015). Increased vertical resolution does not show any significant added value in
34   reproducing the structure and magnitude of the tropospheric NAM (Lee and Black, 2013) nor in the NAO
35   predictability as assessed in a seasonal prediction context with a multi-model approach (Butler et al., 2016).
36   On the other hand, there is mounting evidence that a correct representation of the Quasi Biennal Oscillation,
37   extratropical stratospheric dynamics (the polar vortex and sudden stratospheric warmings), and related
38   troposphere-stratosphere coupling, as well as their interplay with ENSO, are important for NAO/NAM
39   timing (Scaife et al., 2016; Karpechko et al., 2017; Domeisen, 2019; Domeisen et al., 2019), in spite of
40   underestimated troposphere-stratosphere coupling found in models compared to observations (O’Reilly et
41   al., 2019a).
43   The observed trend of the NAM and NAO indices is positive in winter when calculated from the 1960s
44   (Section but it includes large multidecadal variability, which means that the nature of the trend
45   should be interpreted with caution (Gillett et al., 2013). The multi-model multi-member ensemble mean of
46   the trend estimated from historical simulations over that period is very close to zero for both CMIP5 and
47   CMIP6 (Figure 3.33j,k and Figure 3.34a). Even if one cannot rule out that 1958-2014 was an exceptional
48   period of variability, the observational estimates of the wintertime NAO trend lie outside the 5th-95th
49   percentile range of the distribution of trends in the CMIP6 historical simulations, and the observed NAM
50   trends over the same period lie above the 90th percentile. There is a tendency for the CMIP5 models to
51   systematically underestimate the level of multidecadal versus interannual variability of the winter NAO and
52   jet stream compared to observations (Wang et al., 2017c; Bracegirdle et al., 2018; Simpson et al., 2018).
53   Results from CMIP6 (Figure 3.33j,k) and over the 1958-2019 period (Figure 3.34a) confirm this conclusion
54   and seriously questions the ability of the models to simulate long-term fluctuations of the NAO/NAM,
55   independently of its forced or internal origins.
     Do Not Cite, Quote or Distribute                     3-74                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 2   Dedicated SST-forced stand-alone atmospheric model experiments (AMIP) suggest that ocean forcing
 3   appears to play a role in decadal variability of the NAO and associated fluctuations in the strength of the jet
 4   (Woollings et al., 2015). In particular, Atlantic and Indian Ocean SST anomalies (Fletcher and Cassou, 2015;
 5   Baker et al., 2019; Douville et al., 2019; Dhame et al., 2020) may have contributed to the long-term positive
 6   trend of the winter NAO/NAM over the 20th century, but there is only low confidence in such a causal
 7   relationship because of the limitation of the imposed SST approach in AMIP and the uncertainties in
 8   observed SST trends among datasets used as forcing of the atmospheric model. The representation of the
 9   NAM and NAO spatial structure is slightly improved in AMIP ensembles (Figure 3.33g,h), which also
10   produce slightly larger trends than the historical simulations for the NAO, but not for the NAM.
12   When calculated over the most recent two decades, the wintertime NAM/NAO trend is weakly negative
13   since the mid-1990s (Hanna et al., 2015). Recent studies based on observations (Gastineau and Frankignoul,
14   2015) and dedicated modelling experiments (Davini et al., 2015; Peings and Magnusdottir, 2016) suggest
15   that the recent dominance of negative NAM/NAO could be partly related to the latest shift of the Atlantic
16   Multidecadal Variability (AMV) to a warm phase (Sections 2.4.4 and 3.7.7). Some recent modelling studies
17   also find that the Arctic sea ice decline might be partly responsible for more recurrent negative NAM/NAO
18   (Peings and Magnusdottir, 2013; Kim et al., 2014a; Nakamura et al., 2015), while other studies do not
19   robustly identify such responses in models (see also Cross-chapter Box 10.1).
21   In contrast to winter, the observed trend of the NAO index over 1958-2014 is overall negative in summer
22   and is associated with more recurrent blocking conditions over Greenland, in particular since the mid-1990s,
23   thus contributing to the acceleration of melting of the Arctic sea ice (Section and Greenland ice sheet
24   (Section (Fettweis et al., 2013; Hanna et al., 2015; Ding et al., 2017b). The origin of the negative
25   trend of the summer NAO has not been clearly identified, and is hypothesized to be the result of combined
26   influences (Lim et al., 2019), though trends in summertime NAO should also be interpreted with caution
27   because of the presence of strong multidecadal variability. The recent observed negative NAO prevalence
28   and related blocking over Greenland is not present in any of the CMIP5 models (Hanna et al., 2018).
30   Regarding the influence of external forcings since pre-industrial time, AR5 noted that CMIP5 models tend to
31   show an increase in the NAM in response to greenhouse gas increases (Bindoff et al., 2013). Based on the
32   CMIP5 historical ensemble, Gillett and Fyfe (2013) however showed that such a trend is not significant in all
33   seasons. A multi-model assessment of eight CMIP5 models found a NAM increase in response to
34   greenhouse gases, but no robust influence of aerosol changes (Gillett et al., 2013). As for ozone depletion,
35   there is no robust detectable influence on long-term trends of the NAO/NAM (Karpechko et al., 2018a) in
36   contrast to the SAM (Section 3.7.2), but there are indications that extreme Arctic ozone depletion events and
37   their surface expression are linked to an anomalously strong NAM episodes (Calvo et al., 2015; Ivy et al.,
38   2017). However, the direction of causality here is not clear.
40   Conclusions on external forcing influences on the NAM are supported by CMIP6 results based on single
41   forcing ensembles (Figure 3.34a). Positive trends are found in historical simulations over 1958-2019 in
42   boreal winter and are mainly driven by greenhouse gas increases. No significant trends are simulated in
43   response to anthropogenic aerosols, stratospheric ozone or natural forcing. Albeit weak and not statistically
44   significant, the sign of the multi-model mean forced response due to natural forcing is consistent with the
45   observed reduction of solar activity since the 1980s (Section 2.2.1) whose influence would have favoured the
46   negative phase of wintertime NAM/NAO based on the fingerprint of the nearly periodical 11-year solar
47   cycle extracted from models (Scaife et al., 2013; Andrews et al., 2015; Thiéblemont et al., 2015) or
48   observations (Gray et al., 2016; Lüdecke et al., 2020). But such an NAO response to solar forcing remains
49   highly uncertain and controversial, being contradicted by longer proxy records over the last millenium
50   (Sjolte et al., 2018) and modelling evidence (Gillett and Fyfe, 2013; Chiodo et al., 2019). For all seasons and
51   for all individual forcings, uncertainties remain in the estimation of the forced response in the NAM trend as
52   evidenced by considerable model spread (Figure 3.34a) and because the simulated forced component has
53   small amplitude compared to internal variability.
55   Despite new effort since AR5 to reconstruct the NAO beyond the instrumental record, it is still very
     Do Not Cite, Quote or Distribute                      3-75                                      Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 1   challenging to assess the role of external forcings in the apparent multidecadal-to-centennial variability
 2   present throughout the last millenium. Large uncertainties remain in the reconstructed NAO index that are
 3   sensitive to the types of proxies and statistical methods (Trouet et al., 2012; Ortega et al., 2015; Anchukaitis
 4   et al., 2019; Cook et al., 2019; Hernández et al., 2020; Michel et al., 2020) and reconstructed NAO variations
 5   are often not reproduced using pseudo-proxy approaches in models (Lehner et al., 2012; Landrum et al.,
 6   2013). At low frequency, it remains challenging to evaluate if the observed or reconstructed signal
 7   corresponds to an actual change in the NAO intraseasonal to interannual intrinsic properties or rather to a
 8   change in the mean background atmospheric circulation changes projecting on a specific phase of the mode.
 9   Consequently, conflicting results emerge in the attribution of reconstructed long-term variations in the NAO
10   to solar forcing, whose influence thus remains controversial (Gómez-Navarro and Zorita, 2013; Moffa-
11   Sánchez et al., 2014; Ortega et al., 2015; Ait Brahim et al., 2018; Sjolte et al., 2018; Xu et al., 2018).
12   Influences from major volcanic eruptions appear to be more robust (Ortega et al., 2015; Swingedouw et al.,
13   2017) even if some modelling experiments question the amplitude of the response, which mostly projects on
14   the positive phase of the NAM/NAO (Bittner et al., 2016). The forced response is dependent on the strength,
15   seasonal timing and location of the eruption but may also depend on the mean climate background state
16   (Zanchettin et al., 2013) and/or the phases of the main modes of decadal variability such as the AMV
17   (Section 3.7.7) (Ménégoz et al., 2018).
19   Finally, there is some evidence of an apparent signal-to-noise problem referred to as “paradox” in seasonal
20   and decadal hindcasts of the NAO run over 1979-2018 (Scaife and Smith, 2018), which suggests that the
21   NAO response to external forcing, SST or sea ice anomalies could be too weak in models. The weakness of
22   the signal has been related to troposphere-stratosphere coupling which is too intermittent (O’Reilly et al.,
23   2019a) and to chronic model biases in the persistence of NAO/NAM daily regimes, which is critically
24   underestimated in coupled models (Strommen and Palmer, 2019; Zhang and Kirtman, 2019), and which does
25   not exhibit significant improvement when model resolution is increased (Fabiano et al., 2020). Note however
26   that the apparent signal-to-noise problem may be dependent on the period analysed over the 20th century,
27   which questions its interpretation as a general characteristic of coupled models (Weisheimer et al., 2020).
29   In summary, CMIP5 and CMIP6 models are skilful in simulating the spatial features and the variance of the
30   NAM/NAO and associated teleconnections (high confidence). There is limited evidence for a significant role
31   for anthropogenic forcings in driving the observed multidecadal variations of the NAM/NAO from the mid
32   20th century. Confidence in attribution is low (i) because there is a large spread in the modelled forced
33   responses which is overwhelmed anyway by internal variability, (ii) because of the apparent signal-to-noise
34   problem and (iii) because of the chronic inability of models to produce a range of trends which encompasses
35   the observed estimates over the last 60 years.
40   Figure 3.33: Model evaluation of NAM, NAO and SAM in boreal winter. Regression of Mean Sea Level Pressure
41                (MSLP) anomalies (in hPa) onto the normalized principal component (PC) of the leading mode of
42                variability obtained from empirical orthogonal decomposition of the boreal winter (Dec.-Feb) MSLP
43                poleward of 20ºN for the observed Northern Annular Mode (NAM, a), over 20ºN-80°N, 90°W-40°E for
44                the North Atlantic Oscillation as shown by the black sector (NAO, b), and poleward of 20ºS for the
45                Southern Annular Mode (SAM, c) for the JRA-55 reanalysis. Cross marks indicate regions where the
46                anomalies are not significant at the 10% level based on t-test. The period used to calculate the
47                NAO/NAM is 1958-2014 but 1979-2014 for the SAM. (d-f) Same but for the multi-model ensemble
48                (MME) mean from CMIP6 historical simulations. Models are weighted in compositing to account for
49                differences in their respective ensemble size. Diagonal lines stand for regions where less than 80% of the
50                runs agree in sign. (g-i) Taylor diagram summarizing the representation of the modes in models and
51                observations following Lee et al. (2019) for CMIP5 (light blue) and CMIP6 (red) historical runs. The
52                reference pattern is taken from JRA-55 (a-c). The ratio of standard deviation (radial distance), spatial
53                correlation (radial angle) and resulting root-mean-squared-errors (solid isolines) are given for individual
54                ensemble members (crosses) and for other observational products (ERA5 and NOAA 20CRv3, black
55                dots). Coloured dots stand for weighted multi-model mean statistics for CMIP5 (blue) and CMIP6 (light
56                red) as well as for AMIP simulations from CMIP6 (orange). (j-l) Histograms of the trends built from all
57                individual ensemble members and all the models (brown bars). Vertical lines in black show all the
     Do Not Cite, Quote or Distribute                           3-76                                          Total pages: 202
     Final Government Distribution                           Chapter 3                                        IPCC AR6 WGI

 1                 observational estimates. The orange, light-red, and light blue lines indicate the weighted multi-model
 2                 mean of CMIP6 AMIP, CMIP6 and CMIP5 historical simulations, respectively. Further details on data
 3                 sources and processing are available in the chapter data table (Table 3.SM.1).
 5   [END FIGURE 3.33 HERE]
10   Figure 3.34: Attribution of observed seasonal trends in the annular modes to forcings. Simulated and observed
11                trends in NAM indices over 1958-2019 (a) and in SAM indices over 1979-2019 (b) and over 2000-2019
12                (c) for boreal winter (December-February average; DJF) and summer (June-August average; JJA). The
13                indices are based on the difference of the normalized zonally averaged monthly mean sea level pressure
14                between 35ºN and 65ºN for the NAM and between 40ºS and 65ºS for the SAM as defined in Jianping and
15                Wang (2003) and Gong and Wang (1999), respectively: the unit is decade–1. Ensemble mean, interquartile
16                ranges and 5th and 95th percentiles are represented by empty boxes and whiskers for pre-industrial
17                control simulations and historical simulations. The number of ensemble members and models used for
18                computing the distribution is given in the upper-left legend. Grey lines show observed trends from the
19                ERA5 and JRA-55 reanalyses. Multi-model multi-member ensemble means of the forced component of
20                the trends as well as their 5- 95% confidence intervals assessed from t-statistics, are represented by filled
21                boxes, based on CMIP6 individual forcing simulations from DAMIP ensembles; greenhouse gases in
22                brown, aerosols in light blue, stratospheric ozone in purple and natural forcing in green. Models with at
23                least 3 ensemble members are used for the filled boxes, with black dots representing the ensemble means
24                of individual models. Further details on data sources and processing are available in the chapter data table
25                (Table 3.SM.1).
27   [END FIGURE 3.34 HERE]
30   3.7.2    Southern Annular Mode
32   The Southern Annular Mode (SAM) consists of a meridional redistribution of atmospheric mass around
33   Antarctica (Figure 3.33c,f), associated with a meridional shift of the jet and surface westerlies over the
34   Southern Ocean. SAM indices are variously defined as the difference in zonal-mean sea level pressure or
35   geopotential height between middle and high latitudes or via a principal-component analysis (Annex IV.2.2).
36   Observational aspects of the SAM are assessed in Section
38   AR5 assessed that CMIP5 models have medium performance in reproducing the SAM with biases in pattern
39   (Flato et al., 2013). It also concluded that the trend of the SAM toward its positive phase in austral summer
40   since the mid-20th century is likely to be due in part to stratospheric ozone depletion, and there was medium
41   confidence that greenhouse gases have also played a role (Bindoff et al., 2013). Based on proxy
42   reconstructions, AR5 found with medium confidence that the positive SAM trend since 1950 was anomalous
43   compared to the last 400 years (Masson-Delmotte et al., 2013b).
45   Additional research has shown that CMIP5 models reproduce the spatial structure of the SAM well, but tend
46   to overestimate its variability in austral summer at interannual time scales, albeit within the observational
47   uncertainty (Zheng et al. 2013; Schenzinger and Osprey 2015; Figure 3.33c,f,i). This is related to the
48   models’ tendency to simulate slightly more persistent SAM anomalies in summer compared to reanalyses
49   (Schenzinger and Osprey, 2015; Bracegirdle et al., 2020b). This may be due in part to too weak a negative
50   feedback from tropospheric planetary waves (Simpson et al., 2013). CMIP6 models show improved
51   performance in reproducing the spatial structure and interannual variance of the SAM in summer based on
52   Lee et al. (2019) diagnostics (Figure 3.33i), with a better match of its trend with reanalyses over 1979-2014
53   (Figure 3.33l), more realistic persistence and improved positioning of the westerly jet, which in CMIP5
54   models on average is located too far equatorward (Bracegirdle et al., 2020a; Grose et al., 2020). In CMIP5, it
55   is also found that models which extend above the stratopause tend to simulate stronger summertime trends in
56   the late 20th century than their counterparts with tops within the stratosphere (Rea et al., 2018a; Son et al.,
57   2018b), though other differences between these sets of models, such as additional physical processes
     Do Not Cite, Quote or Distribute                          3-77                                          Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   operating in the stratosphere or interactive ozone chemistry, may have also affected these results (Gillett et
 2   al., 2003a; Sigmond et al., 2008; Rea et al., 2018b). At the surface, Ogawa et al. (2015) demonstrate with an
 3   atmospheric model the importance of sharp midlatitude SST gradients for stratospheric ozone depletion to
 4   affect the SAM in summer. These studies imply that the well resolved stratosphere combined with finer
 5   ocean horizontal resolution has contributed to the stronger simulated trends in CMIP6 than in CMIP5.
 7   CMIP6 historical simulations capture the observed positive trend of summertime SAM when calculated from
 8   the 1970s to the 2010s (Figure 3.34b). Thomas et al. (2015) found that the chance for the observed 1980-
 9   2004 trend to occur only due to internal variability is less than 10% in many of the CMIP5 models, and
10   results from CMIP6 models suggest that the chance of the 1979-2019 trend being due to internal variability
11   could be even lower (Figure 3.34b). Although paleo-reconstructions of the SAM index are uncertain and
12   vary in terms of long-term trends (Section, new reconstructions show that the 60-year summertime
13   SAM trend since the mid-20th century is outside the 5th-95th percentile range of the trends in the pre-
14   industrial variability, which matches the trend range of CMIP5 pre-industrial control simulations well
15   (Dätwyler et al., 2018).
17   In general agreement with AR5, new research continues to indicate that both stratospheric ozone depletion
18   and increasing greenhouse gases have contributed to the trend of the SAM during austral summer toward its
19   positive phase in recent decades (Solomon and Polvani, 2016), with the ozone depletion influence
20   dominating (Gerber and Son, 2014a; Son et al., 2018a). In CMIP6 historical simulations there are significant
21   positive SAM trends over the 1979-2019 period in austral summer, although the contribution from ozone
22   forcing evaluated with the four available models is not significant (Figure 3.34b). Three of these models
23   share the same standard prescribed ozone forcing and produce significantly positive SAM trends over an
24   extended period (1957-2019). The fourth model, MRI-ESM2-0, has the option of interactive ozone
25   chemistry. Its ozone-only experiment is forced by prescribed ozone derived from its own historical
26   simulations and produces a negative SAM trend associated with weak ozone depletion (Morgenstern et al.,
27   2020). Morgenstern et al. (2014) and Morgenstern et al. (2020) find an indirect influence of greenhouse
28   gases on the SAM via induced ozone changes in coupled chemistry-climate simulations, which differ from
29   the prescribed ozone simulations shown in Figure 3.34b. Since ~1997, the effective abundance of ozone-
30   depleting halogen has been decreasing in the stratosphere (WMO, 2018), leading to a stabilization or even a
31   reversal of stratospheric ozone depletion (Sections and The ozone stabilisation and slight
32   recovery since ~2000 may have caused a pause in the summertime SAM trend (Figure 3.34c; Saggioro and
33   Shepherd, 2019; Banerjee et al., 2020), although some influence from internal variability cannot be ruled out.
34   While some studies find an anthropogenic aerosol influence on the summertime SAM (Gillett et al., 2013;
35   Rotstayn, 2013), recent studies with larger multi-model ensembles find that this effect is not robust (Steptoe
36   et al., 2016; Choi et al., 2019), consistent with CMIP6 single forcing ensembles (Figure 3.34). In the CMIP5
37   simulations, volcanic stratospheric aerosol has a significant weakening effect on the SAM in autumn and
38   winter (Cross-Chapter Box 4.1; Gillett and Fyfe, 2013), but there is no evidence that this effect leads to a
39   significant multidecadal trend since the late 20th century. Beyond external forcing, Fogt et al. (2017) show a
40   significant association of tropical SST variability with the summertime SAM trend since the mid-20th century
41   in agreement with (Lim et al., 2016) who however demonstrate that such a teleconnection between the
42   summertime SAM and El Niño-Southern Oscillation (Annex IV.2.3), found in observations, is missing in
43   many CMIP5 models.
45   On longer time scales, last Millennium experiments from CMIP5 models fail to capture multicentennial
46   variability evident in the reconstructions for the pre-industrial era (Abram et al., 2014; Dätwyler et al., 2018),
47   which is also the case in those from available CMIP6 models (Figure 3.35). However there is large
48   uncertainty among reconstructions (Section It is therefore unclear whether this disagreement reflects
49   this observational uncertainty, whether forcings such as variations in the imposed insolation may be too
50   weak, whether models are insufficiently sensitive to such variations, or whether internal variability including
51   that associated with tropical Pacific variability is underrepresented (Abram et al., 2014). The explanation
52   could be a combination of all these factors. However, despite the aforementioned limitations of the
53   reconstructions, Section assesses that the recent positive trend in the SAM is likely unprecedented in
54   at least the past millenium (medium confidence). CMIP5 and CMIP6 last-millennium simulations only
55   capture the present anomalous state during in the final decades of the simulations which are dominated by
     Do Not Cite, Quote or Distribute                      3-78                                        Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   human influence; this state is also outside the range of simulated variability characteristic of pre-industrial
 2   times.
 4   In summary, it is very likely that anthropogenic forcings have contributed to the observed trend of the
 5   summer SAM toward its positive phase since the 1970s. This assessment is supported by further model
 6   studies that confirm the human influence on the summertime SAM with improved models since AR5. While
 7   ozone depletion contributed to the trend from the 1970s to the 1990s (medium confidence), its influence has
 8   been small since 2000, leading to a weaker summertime SAM trend over 2000-2019 (medium confidence).
 9   Climate models reproduce the spatial structure of the summertime SAM observed since the late 1970s well
10   (high confidence). CMIP6 models reproduce the spatiotemporal features and recent multidecadal trend of the
11   summertime SAM better than CMIP5 models (medium confidence). However, there is a large spread in the
12   intensity of the SAM response to ozone and greenhouse gas changes in both CMIP5 and CMIP6 models
13   (high confidence), which limits the confidence in the assessment of the ozone contribution to the observed
14   trends. CMIP5 and CMIP6 models do not capture multicentennial variability of the SAM found in proxy
15   reconstructions (low confidence). This confidence level reflects that it is unclear whether this is due to a
16   model or an observational shortcoming.
21   Figure 3.35: Southern Annular Mode (SAM) indices in the last millennium. (a) Annual SAM reconstructions by
22                Abram et al. (2014) and Dätwyler et al. (2018). (b) The annual-mean SAM index defined by Gong and
23                Wang (1999) in CMIP5 and CMIP6 Last Millennium simulations extended by historical simulations. All
24                indices are normalized with respect to 1961-1990 means and standard deviations. Thin lines and thick
25                lines show 7-year and 70-year moving averages, respectively. Further details on data sources and
26                processing are available in the chapter data table (Table 3.SM.1).
28   [END FIGURE 3.35 HERE]
31   3.7.3   El Niño-Southern Oscillation
33   The El Niño-Southern Oscillation (ENSO), which is generated via seasonally modulated interactions
34   between the tropical Pacific ocean and atmosphere, influences severe weather, rainfall, river flow and
35   agricultural production over large parts of the world (McPhaden et al., 2006). In fact, the remote climate
36   influence of ENSO is so large that knowledge of its current phase and forecasts of its future phase largely
37   underpin many seasonal rainfall and temperature forecasts worldwide (Annex IV.2.3).
39   AR5 noted that there have been clear improvements in the simulation of ENSO through previous generations
40   of CMIP models (Flato et al., 2013), such that many CMIP5 models displayed behaviour that was
41   qualitatively similar to that of the observed ENSO (Guilyardi et al., 2012). However, systematic errors were
42   identified in the models’ representation of the Tropical Pacific mean state and aspects of their interannual
43   variability that affect quantitative comparisons. The AR5 assessment of ENSO concluded that the
44   considerable observed inter-decadal modulations in ENSO amplitude and spatial pattern were largely
45   consistent with unforced model simulations. Thus, there was low confidence in the role of a human-induced
46   influence in these (Bindoff et al., 2013).
48   Observed ENSO amplitude, which is measured by the standard deviation of SST anomalies in a central
49   equatorial Pacific region often referred to as the Nino 3.4 region, along with the lifecycle of events, are both
50   reasonably well reproduced by most CMIP5 and CMIP6 models (Figure 3.36) (Bellenger et al., 2014;
51   Planton et al., 2020). The average CMIP5 model ENSO amplitude is slightly lower than that observed, while
52   the average CMIP6 model ENSO amplitude is slightly higher than observed (Figure 3.36). The ENSO
53   amplitude of the individual models, however, is highly variable across CMIP5 and CMIP6 models with
54   many displaying either more or less variability than observed (Stevenson, 2012; Grose et al., 2020; Planton
55   et al., 2020).

     Do Not Cite, Quote or Distribute                       3-79                                       Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 5   Figure 3.36: Life cycle of (left) El Niño and (right) La Niña events in observations (black) and historical
 6                simulations from CMIP5 (blue; extended with RCP4.5) and CMIP6 (red). An event is detected when
 7                the December ENSO index value in year zero exceeds 0.75 times its standard deviation for 1951-2010. (a,
 8                b) Composites of the ENSO index (ºC). The horizontal axis represents month relative to the reference
 9                December (the grey vertical bar), with numbers in parentheses indicating relative years. Shading and lines
10                represent 5th-95th percentiles and multi-model ensemble means, respectively. (c, d) Mean durations
11                (months) of El Niño and La Niña events defined as number of months in individual events for which the
12                ENSO index exceeds 0.5 times its December standard deviation. Each dot represents an ensemble
13                member from the model indicated on the vertical axis. The boxes and whiskers represent multi-model
14                ensemble mean, interquartile ranges and 5th and 95th percentiles of CMIP5 and CMIP6. The CMIP5 and
15                CMIP6 multi-model ensemble means and observational values are indicated at top right of each panel.
16                The multi-model ensemble means and percentile values are evaluated after weighting individual members
17                with the inverse of the ensemble size of the same model, so that individual models are equally weighted
18                irrespective of their ensemble sizes. The ENSO index is defined as the SST anomaly averaged over the
19                Niño 3.4 region (5ºS-5ºN, 170ºW-120ºW). All results are based on 5-month running mean SST anomalies
20                with triangular-weights after linear detrending. Further details on data sources and processing are
21                available in the chapter data table (Table 3.SM.1).
23   [END FIGURE 3.36 HERE]
26   ENSO events are often synchronized to the seasonal cycle in the observations, as the associated SST
27   anomalies tend to peak in boreal winter (November-January) and be at their weakest in the boreal spring
28   (March-April) (Harrison and Larkin, 1998; Larkin and Harrison, 2002). The majority of CMIP5 and CMIP6
29   models broadly reproduce the seasonality of ENSO SST variability in the central equatorial Pacific
30   (Taschetto et al., 2014a; Abellán et al., 2017; Grose et al., 2020; Planton et al., 2020) (Figure 3.37).
31   However, CMIP5 models, while displaying an improvement on CMIP3 models, appear to underrepresent the
32   magnitude of the seasonal variance modulation (Bellenger et al., 2014). This under-representation of
33   seasonal variance modulation continues in CMIP6 models, which display no statistically significant
34   difference in this behaviour when compared to CMIP5 models (Planton et al., 2020) (Figure 3.37).
39   Figure 3.37: ENSO seasonality in observations (black) and historical simulations from CMIP5 (blue; extended
40                with RCP4.5) and CMIP6 (red) for 1951-2010. (a) Climatological standard deviation of the monthly
41                ENSO index (SST anomaly averaged over the Niño 3.4 region; °C). Shading and lines represent 5th-95th
42                percentiles and multi-model ensemble means, respectively. (b) Seasonality metric, which is defined for
43                each model and each ensemble member as the ratio of the ENSO index climatological standard deviation
44                in November-January (NDJ) to that in March-May (MAM). Each dot represents an ensemble member
45                from the model indicated on the vertical axis. The boxes and whiskers represent the multi-model
46                ensemble mean, interquartile ranges and 5th and 95th percentiles of CMIP5 and CMIP6 individually. The
47                CMIP5 and CMIP6 multi-model ensemble means and observational values are indicated at the top right
48                of the panel. The multi-model ensemble means and percentile values are evaluated after weighting
49                individual members with the inverse of the ensemble size of the same model, so that individual models
50                are equally weighted irrespective of their ensemble sizes. All results are based on 5-month running mean
51                SST anomalies with triangular-weights after linear detrending. Further details on data sources and
52                processing are available in the chapter data table (Table 3.SM.1).
54   [END FIGURE 3.37 HERE]
57   Observations show strong multi-decadal modulation of ENSO variance throughout the 20th century, with the
58   most recent period displaying larger variability while the mid-century displayed relatively low ENSO
     Do Not Cite, Quote or Distribute                         3-80                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   variability (e.g., Li et al., 2013; McGregor et al., 2013; Hope et al., 2017; Figure 2.36). As assessed in
 2   Section 2.4.2, ENSO amplitude since 1950 is higher than over the pre-industrial period from 1850 as far
 3   back as 1400 (medium confidence), but there is low confidence that it is higher than the variability over
 4   periods prior to 1400. This reported variance increase suggests that external forcing plays a role in the ENSO
 5   variance changes (Hope et al., 2017b). However, large ensembles of single model or multiple model
 6   simulations do not find strong trends in ENSO variability over the historical period, suggesting that external
 7   forcing has not yet modulated ENSO variability with a magnitude that exceeds the range of internal
 8   variability (Hope et al., 2017b; Stevenson et al., 2017; Maher et al., 2018b). This is consistent with the
 9   Chapter 2 assessment, that there is no clear evidence for a recent sustained shift in ENSO beyond the range
10   of variability on decadal to millennial timescales (Section 2.4.2). CMIP5 and CMIP6 models show a
11   decrease in ENSO variance in the mid-Holocene (Brown et al., 2020), though not to the extent seen in paleo-
12   proxy records (Emile-Geay et al., 2016). This suggests that both modelled and observed ENSO respond to
13   changes in external forcing, but not necessarily in the same manner.
15   Most CMIP5 and CMIP6 models are found to represent the general structure of observed SST anomalies
16   during ENSO events well (Kim and Yu, 2012; Taschetto et al., 2014b; Brown et al., 2020; Grose et al.,
17   2020). However, the majority of CMIP5 models display SST anomalies that: i) extend too far to the west
18   (Taschetto et al., 2014b; Capotondi et al., 2015); and ii) have meridional widths that are too narrow (Zhang
19   and Jin, 2012) compared to the observations. CMIP6 models display a statistically significant improvement
20   in the longitudinal representation of ENSO SST anomalies relative to CMIP5 models (Planton et al., 2020),
21   however, systematic biases in the zonal extent and meridional width remain in CMIP6 models (Fasullo et al.,
22   2020; Planton et al., 2020). The ENSO phase asymmetry, where observed strong El Niño events are larger
23   and have a shorter duration than strong La Niña events (Ohba and Ueda, 2009; Frauen and Dommenget,
24   2010), is also underrepresented in both CMIP5 and CMIP6 models (Zhang and Sun, 2014; Planton et al.,
25   2020). In this instance, both CMIP5 and CMIP6 models typically display El Niño events that have a longer
26   duration than those observed, La Niña events that have a similar duration to those observed, and there is very
27   little asymmetry in the duration of El Niño and La Niña phases (Figure 3.36). Roberts et al. (2018) find an
28   improvement in amplitude asymmetry in a HighResMIP model, but the underrepresentation remains.
30   The continuum of El Niño events are typically stratified into two types (often termed “flavours”), Central
31   Pacific and East Pacific, where the name denotes the location of the events’ largest SST anomalies (Annex
32   IV.2.3; Capotondi et al., 2015). As discussed in Section 2.4.2, the different types of events tend to produce
33   distinct teleconnections and climatic impacts (e.g., Taschetto et al., 2020). The characteristics of El Niño
34   events of these two flavours in CMIP5 were generally comparable to the observations (Taschetto et al.,
35   2014b). CMIP6 models, however, display a statistically significant improvement in the representation of this
36   ENSO event-to-event SST anomaly diversity when compared with CMIP5 models (Planton et al., 2020). In
37   addition to this ENSO event diversity, the short observational record also displays an increase in the number
38   of the Central Pacific-type events in recent decades (Ashok et al., 2007; McPhaden et al., 2011), which has
39   also been identified as unusual in the context of the last 500-800 years based on recent paleo-climatic
40   reconstructions (Section 2.4.2; Liu et al., 2017; Freund et al., 2019). However, the short observational record
41   combined with observational (L’Heureux et al., 2013) and paleo-climatic reconstruction uncertainties
42   preclude firm conclusions being made about the long-term changes in the occurrence of different El Niño
43   event types. Initial analysis with a selected number of CMIP3 models suggested that there may be a forced
44   component to this recent prominence of Central Pacific-type events (Yeh et al., 2009), but analysis since
45   then suggests that this behaviour is i) consistent with that expected from internal variability (Newman et al.,
46   2011); and ii) not apparent across the full CMIP5 ensemble of historical simulations (Taschetto et al.,
47   2014b). Analysis of single-model large ensembles suggests that changes to ENSO event type in response to
48   historical radiative forcing are not significant (e.g., Stevenson et al., 2017). These same results, however,
49   also suggest that multiple forcings can have significant influences on ENSO type and that the net response
50   will depend on the accurate representation of the balance of these forcings (Stevenson et al., 2017).
52   The climatic effects of ENSO outside the tropical Pacific largely arise through atmospheric teleconnections
53   that are induced by ENSO-driven changes in deep tropical atmospheric convection and heating (Yeh et al.,
54   2018). The teleconnections to higher latitudes are forced by waves that propagate into the extratropics
55   (Hoskins and Karoly, 1981) and respectively excite the Pacific-North American pattern (Horel and Wallace,
     Do Not Cite, Quote or Distribute                     3-81                                       Total pages: 202
     Final Government Distribution                        Chapter 3                                      IPCC AR6 WGI

 1   1981) and Pacific-South American pattern (Karoly, 1989; Irving and Simmonds, 2016) in the Northern and
 2   Southern Hemispheres. Given the influence of these teleconnections on climate and extremes around the
 3   globe, it is important to understand how well they are reproduced in CMIP models. What has also become
 4   clear is that spatial correlations of ENSO’s teleconnections calculated over relatively short periods (< 100
 5   years) may not be the most effective way to assess these relationships (Langenbrunner and Neelin, 2013;
 6   Perry et al., 2019). This is because the spatial patterns are significantly affected by internal atmospheric
 7   variability on relatively short time scales (< 100 years) (Batehup et al., 2015; Perry et al., 2019). However,
 8   looking at simplified metrics like the agreement in the sign of the teleconnections (Langenbrunner and
 9   Neelin, 2013), regional average teleconnection strength over land (Perry et al., 2019), or a combination of
10   both (Power and Delage, 2018a) provides a more robust depiction of the teleconnection representation.
11   Examining sign agreement for the teleconnection patterns, ensembles of CMIP5 AMIP simulations display
12   broad spatial regions with high sign agreement with the observations, suggesting that the model ensemble is
13   producing useful information regarding the teleconnected precipitation signal (Langenbrunner and Neelin,
14   2013). Looking at regional averages of CMIP5 historical simulations, Power and Delage (2018) show that
15   the average coupled model teleconnection pattern reproduces the sign of the observed teleconnections in the
16   majority of the 25 regions analysed. The sign agreement between the observed teleconnection and the multi-
17   model mean teleconnection remains strong in CMIP6 (18 out of 20 displayed regions) (Figure 3.38), and the
18   observed DJF teleconnection strength falls within the modelled range in all of the displayed regions for
19   temperature and precipitation. It is noted, however, that while there is broad agreement in ENSO
20   teleconnections between CMIP6 models and observations during DJF (e.g., Fasullo et al., 2020), there are
21   regions and seasons where the modelled teleconnection strength is outside the observed range (Chen et al.,
22   2020).
27   Figure 3.38: Model evaluation of ENSO teleconnection for 2m-temperature and precipitation in boreal winter
28                (December-January-February). Teleconnections are identified by linear regression with the Niño 3.4
29                SST index based on ERSSTv5 during the period 1958-2014. Maps show observed patterns for
30                temperature from the Berkeley Earth dataset over land and from ERSSTv5 over ocean (ºC, top) and for
31                precipitation from GPCC over land (shading, mm day–1) and GPCP worldwide (contours, period: 1979-
32                2014). Distributions of regression coefficients (grey histograms) are provided for a subset of AR6
33                reference regions defined in Atlas (Section A1.3) for temperature (top) and precipitation (bottom). All
34                fields are linearly detrended prior to computation. Multi-model multi-member ensemble means are
35                indicated by thick vertical black lines. Blue vertical lines show three observational estimates of
36                temperature, based on Berkeley Earth, GISTEMP and CRUTS datasets, and two observational estimates
37                of precipitation, based on GPCC and CRUTS datasets. Further details on data sources and processing are
38                available in the chapter data table (Table 3.SM.1).
40   [END FIGURE 3.38 HERE]
43   Most CMIP5 and CMIP6 models exhibit ENSO behaviour during the historical period that, to first order, is
44   qualitatively similar to that of the observed ENSO. Many studies are now delving deeper into the models to
45   understand if they are accurately producing the dynamics driving ENSO and its initiation (Jin et al., 2006;
46   Bellenger et al., 2014; Vijayeta and Dommenget, 2017a; Bayr et al., 2019; Planton et al., 2020). For both
47   CMIP3 and CMIP5, diagnostics of ENSO event growth appear to show that the models, while producing
48   ENSO variability that is qualitatively similar to that observed, do not represent the balance of the underlying
49   dynamics well. The atmospheric Bjerknes feedback is too weak in the majority of models, while the surface
50   heat flux feedback is also too weak in the majority of models. The former restricts event growth, while the
51   latter restricts event damping, which when combined allow most models to produce variability in a range
52   that is consistent with the observations (Bellenger et al., 2014; Kim et al., 2014b; Vijayeta and Dommenget,
53   2017b; Bayr et al., 2019). Analysis of ENSO representation in a subset of CMIP6 models by Planton et al.
54   (2020) suggests that these issues remain.
56   To conclude, ENSO representation in CMIP5 models displayed a significant improvement from the
     Do Not Cite, Quote or Distribute                        3-82                                        Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   representation of ENSO variability in CMIP3 models, which displayed much more intermodel spread in
 2   standard deviation, and stronger biennial periodicity (Guilyardi et al., 2012; Flato et al., 2013). In general,
 3   there has been no large step change in the representation of ENSO between CMIP5 and CMIP6, however,
 4   CMIP6 models appear to better represent some key ENSO characteristics (e.g., Brown et al., 2020; Planton
 5   et al., 2020). The instrumental record and paleo-proxy evidence through the Holocene all suggest that ENSO
 6   can display considerable modulations in amplitude, pattern and period (see also Section 2.4.2). For the
 7   period since 1850, there is also no clear evidence for a sustained shift in ENSO beyond the range of internal
 8   variability (Section 2.4.2). However, paleo-proxy evidence indicates with medium confidence that ENSO
 9   variability since 1950 is greater than at any time between 1400 and 1850 (Section 2.4.2). Coupled models
10   display large changes of ENSO behaviour in the absence of external forcing changes, and little-to-no
11   variance sensitivity to historical anthropogenic forcing. Thus, there is low confidence that anthropogenic
12   forcing has led to the changes of ENSO variability inferred from paleo-proxy evidence.
14   Chapter 2 reports low confidence that the apparent change from East Pacific- to Central Pacific-type El Niño
15   events that occurred in the last 20-30 years was representative of a long term change. While some climate
16   models do suggest external forcing may affect the El Niño event type, most climate models suggest that what
17   has been observed is well within the range of natural variability. Thus, there is low confidence that
18   anthropogenic forcing has had an influence on the observed changes in El Niño event type.
21   3.7.4   Indian Ocean Basin and Dipole Modes
23   The Indian Ocean Basin (IOB) and Dipole (IOD) modes are the two leading modes of interannual SST
24   variability over the tropical Indian Ocean, featuring basin-wide warming/cooling and an east-west dipole of
25   SST anomalies, respectively (Annex IV.2.4). The IOD mode is anchored to boreal summer to autumn by the
26   air-sea feedback, and often develops in concert with ENSO. Driven by matured ENSO events, the IOB mode
27   peaks in boreal spring and often persists into the subsequent summer. Similar patterns of Indian Ocean SST
28   variability also dominate its decadal and longer time scale variability (Han et al., 2014b).
30   AR5 concluded that models show high and medium performance in reproducing the IOB and IOD modes,
31   respectively (medium confidence), with difficulty in reproducing the persistence of the IOB and the pattern
32   and magnitude of the IOD (Flato et al., 2013). There was low confidence that changes in the IOD were
33   detectable or attributable to human influence (Bindoff et al., 2013).
35   Since the AR5, CMIP5 model representation of these modes has been analysed in detail, finding that most of
36   the models qualitatively reproduce the spatial and seasonal features of the IOB and IOD modes (Chu et al.,
37   2014; Liu et al., 2014; Tao et al., 2016b). Improvements in simulating the IOB mode since CMIP3 have been
38   identified in reduced multi-model mean bias and inter-model spread (Tao et al., 2016b). CMIP5 models
39   overall capture the transition from the IOD to IOB modes during an ENSO event (Tao et al., 2016b). The
40   IOB mode is forced in part through a cross-equatorial wind-evaporation-SST feedback triggered by ENSO-
41   forced anomalous ocean Rossby waves that propagate to the shallow climatological thermocline dome in the
42   tropical southwestern Indian Ocean (Du et al., 2009). Consistently, models with a deeper climatological
43   thermocline dome produce a weaker and less persistent IOB mode (Li et al., 2015b; Zheng et al., 2016). The
44   deep thermocline bias remains in the ensemble mean of CMIP5 models due to a common surface easterly
45   wind bias over the equatorial Indian Ocean (Lee et al., 2013) associated with weaker South Asian summer
46   monsoon circulation (Li et al., 2015c). However, the influence of this systematic bias may be compensated
47   by other biases, resulting in a realistic IOB magnitude (Tao et al., 2016b). Halder et al. (2020) found that
48   CMIP6 models reproduce the IOB mode reasonably well, but did not evaluate the progress since CMIP5.
50   By contrast, the IOD magnitude is overestimated by CMIP5 models on average, though with noticeable
51   improvements from CMIP3 models (Liu et al., 2014). The overestimation of the IOD magnitude remains in
52   most of 34 CMIP6 models examined in McKenna et al. (2020) with worsening on average in July and
53   August. A too steep climatological thermocline slope along the equator due to the surface easterly wind bias
54   in boreal summer and autumn contributes to this IOD magnitude bias through an excessively strong Bjerknes
55   feedback in CMIP5 (Liu et al., 2014; Li et al., 2015c; Hirons and Turner, 2018). The surface easterly bias
     Do Not Cite, Quote or Distribute                     3-83                                       Total pages: 202
     Final Government Distribution                     Chapter 3                                    IPCC AR6 WGI

 1   and associated east-west SST gradient bias are not improved in CMIP6 (Long et al., 2020), suggesting that
 2   the thermocline bias also remains. McKenna et al. (2020) additionally find degradation in the positive-
 3   negative asymmetry of the IOD but an improvement in IOD frequency in a subset of CMIP6 models
 4   compared to CMIP5. In terms of teleconnections, the equatorial surface easterly wind bias also affects the
 5   IOD-associated moisture transport anomalies toward tropical eastern Africa (Hirons and Turner, 2018)
 6   where the IOD is associated with strong precipitation anomalies in boreal autumn (Annex IV.2.4). CMIP5
 7   and CMIP6 models capture the IOD teleconnection to Southern and Central Australian precipitation
 8   although it is weaker on average than observed, with no clear improvements from CMIP5 to CMIP6 (Grose
 9   et al., 2020). Strong IOD events could also influence the NH extratropical circulation in winter and in
10   particular the NAM (Section 3.7.1), based on interference between forced Rossby waves emerging from the
11   Indian Ocean and climatological stationary waves (Fletcher and Cassou, 2015). The record positive phase of
12   the NAO/NAM in winter 2019-2020 assessed over the instrumental era has been accordingly linked to the
13   record IOD event of autumn 2019 (Hardiman et al., 2020), which has been associated with the devastating
14   record fire season in Australia (Wang and Cai, 2020).
16   The observed Indian Ocean basin-average SST increase on multidecadal and centennial time scales is well
17   represented by CMIP5 historical simulations, and has been attributed to the effects of greenhouse gases
18   offset in part by the effects of anthropogenic aerosols mainly through aerosol-cloud interactions (Dong and
19   Zhou, 2014; Dong et al., 2014b). The observed SST trend is larger in the western than eastern tropical Indian
20   Ocean, which leads to an apparent upward trend of the IOD index, but this trend is statistically insignificant
21   (Section 2.4.3). CMIP5 models capture this warming pattern, which may be associated with Walker
22   circulation weakening over the Indian Ocean due to greenhouse gas forcing (Dong and Zhou, 2014).
23   However, strong internal decadal IOD-like variability and observational uncertainty preclude attribution (Cai
24   et al., 2013; Han et al., 2014c; Gopika et al., 2020). Such a positive IOD-like change in equatorial zonal SST
25   gradient suggests an increase in the frequency of extreme positive events (Cai et al., 2014) and skewness
26   (Cowan et al., 2015) of the IOD mode. While there is some evidence of an increase in frequency of positive
27   IOD events during the second half of the 20th century, the current level of IOD variability is not
28   unprecedented in a proxy reconstruction for the last millenium (Section 2.4.3, Abram et al., 2020). Besides,
29   the IOD magnitude in the late 20th century is not significantly different between CMIP5 simulations forced
30   by historical and natural-only forcings, though this conclusion is based on only five selected ensemble
31   members that realistically reproduce statistical features of the IOD (Blau and Ha, 2020). While selected
32   CMIP5 models show weakening (Thielke and Mölg, 2019) and seasonality changes (Blau and Ha, 2020) in
33   IOD-induced rainfall anomalies in tropical eastern Africa, no comparison with observational records has
34   been made. Likewise, while a strengthening tendency of the ENSO-IOB mode correlation and resultant
35   intensification of the IOB mode are found in historical or future simulations in selected CMIP5 models (Hu
36   et al., 2014; Tao et al., 2015), such a change has not been detected in observational records.
38   After linear detrending, Pacific decadal variability (PDV; Annex IV.2.6; Section 3.7.6) has been suggested
39   as a driver of decadal-to-multidecadal variations in the IOB mode (Dong et al., 2016). However, correlation
40   between the PDV and a decadal IOB index, defined from linearly detrended SST, changed from positive to
41   negative during the 1980s (Han et al., 2014a). The increase in anthropogenic forcing and recovery from the
42   eruptions of El Chichón in 1982 and Pinatubo in 1991 may have overwhelmed the PDV influence, and
43   explain this change (Dong and McPhaden, 2017; Zhang et al., 2018a). However, the low statistical degrees
44   of freedom hamper clear detection of human influence in this correlation change.
46   To summarize, there is medium confidence that changes in the interannual IOD variability in the late 20th
47   century inferred from observations and proxy records are within the range of internal variability. There is no
48   evidence of anthropogenic influence on the interannual IOB. On decadal-to-multidecadal time scales, there is
49   low confidence that human influence has caused a reversal of the correlation between PDV and decadal
50   variations in the IOB mode. The low confidence in this assessment is due to the short observational record,
51   limited number of models used for the attribution, lack of model evaluation of the decadal IOB mode, and
52   uncertainty in the contribution from volcanic aerosols. Nevertheless, CMIP5 models have medium overall
53   performance in reproducing both the interannual IOB and IOD modes, with an apparently good performance
54   in reproducing the IOB magnitude arising from compensation of biases in the formation process, and overly
55   high IOD magnitude due to the mean state bias (high confidence). There is no clear improvement in the
     Do Not Cite, Quote or Distribute                     3-84                                     Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   simulation of the IOD from CMIP5 to CMIP6 models, though there is only medium confidence in this
 2   assessment, since only a subset of CMIP6 models have been examined. There is no evidence for
 3   performance changes in simulating the IOB from CMIP5 to CMIP6 models.
 6   3.7.5   Atlantic Meridional and Zonal Modes
 8   The Atlantic Zonal Mode (AZM), often referred to as the Atlantic Equatorial Mode or Atlantic Niño, and the
 9   Atlantic Meridional Mode (AMM) are the two leading basin wide patterns of interannual-to-decadal
10   variability in the tropical Atlantic. Akin to ENSO in the Pacific, the term Atlantic Niño is broadly used to
11   refer to years when the SSTs in the tropical eastern Atlantic basin along the cold tongue are significantly
12   warmer than the climatological average. The AMM is characterized by anomalous cross-equatorial gradients
13   in SST. Both modes are associated with altered strength of the ITCZ and/or latitudinal shifts in the ITCZ,
14   which locally affect African and American monsoon systems and remotely affect Tropical Pacific and Indian
15   Ocean variability through inter-basins teleconnections. A detailed description of both AZM and AMM, as
16   well as their associated teleconnection over land, is given in Annex IV.2.5
18   AR5 mentioned the considerable difficulty in simulating both Atlantic Niño and AMM despite some
19   improvements in CMIP5 for some models (Flato et al., 2013). Severe biases in mean state and variance for
20   both SST and atmospheric dynamics including rainfall (e.g. a double ITCZ) as well as teleconnections were
21   reported. AR5 highlighted the complexity of the Tropical Atlantic biases, which were explained by multiple
22   factors both in the ocean and atmosphere.
24   Since AR5, further analysis of the major persistent biases in models has been reported (Xu et al., 2014;
25   Jouanno et al., 2017; Yang et al., 2017b; Dippe et al., 2018; Lübbecke et al., 2018; Voldoire et al., 2019a).
26   Errors in equatorial and basin wide trade winds, cloud cover and ocean vertical mixing and dynamics both
27   locally and in remote subtropical upwelling regions, key thermodynamic ocean-atmosphere feedbacks, and
28   tropical land-atmosphere interaction have been shown to be detrimental to the representation of both the
29   Atlantic Niño and AMM leading to poor teleconnectivity over land (Rodríguez-Fonseca et al., 2015;
30   Wainwright et al., 2019) and between tropical basins (Ott et al., 2015).
32   Despite some improvements (Richter et al., 2014; Nnamchi et al., 2015), biases in the mean state are so large
33   that the mean east-west temperature gradient at the equator along the thermocline remains opposite to
34   observed in two thirds of the CMIP5 models (Section, which clearly affects the simulation of the
35   Atlantic Niño and associated dynamics (Muñoz et al., 2012; Ding et al., 2015; Deppenmeier et al., 2016).
36   The interhemispheric SST gradient is also systematically underestimated in models with a too cold mean
37   state in the northern part of the Tropical Atlantic ocean and too warm conditions in the Southern Atlantic
38   basin. The seasonality is poorly reproduced and the wind-SST coupling is weaker than observed so that
39   altogether, and despite AMM-like variability in 20th century climate simulations, AMM is not the dominant
40   Atlantic mode in all CMIP5 models (Liu et al., 2013; Amaya et al., 2017). These biases in mean state
41   translate into biases in modelling the mean ITCZ (Flato et al., 2013). Similar biases were found in
42   experiments using CMIP5 models but with different climate background states, such as Last Glacial
43   Maximum, Mid-Holocene and future scenario simulations (Brierley and Wainer, 2018)⁠. Analyses from
44   CMIP6 show encouraging results in the representation of Atlantic Niño and AMM modes of variability in
45   terms of amplitude and seasonality. Some models now display reduced biases in the spatial structure of the
46   modes and related explained variance but persistent errors still remain on average in the timing of the modes
47   and in the coupled nature of the modes, i.e. the strength of the link between ocean (SST, mixed layer depth)
48   and atmospheric (wind) anomalies (Richter and Tokinaga, 2020b), as well as in the Indian Ocean equatorial
49   east-west temperature gradient (Section, Figure 3.24).
51   There are some recent indications that increasing model resolution both vertically and horizontally, in the
52   ocean and atmospheric component (Richter, 2015; Small et al., 2015; Harlaß et al., 2018), could partly
53   alleviate some tropical Atlantic biases in mean state (Section, seasonality, interannual-to-decadal
54   variability and associated teleconnectivity over land, such as with the West African monsoon (Steinig et al.,
55   2018). Results from CMIP6 tend to confirm that increasing resolution is not the unique way to address the
     Do Not Cite, Quote or Distribute                     3-85                                      Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   biases in the Tropical Atlantic (Richter and Tokinaga, 2020b). For instance, the inclusion of a stochastic
 2   physics scheme has a nearly equivalent effect in the improvement of the mean number and the strength
 3   distribution of tropical Atlantic cyclones (Vidale et al., 2021).
 5   Section 2.4.4 assess that there is low confidence in any sustained changes to the AZM and AMM variability
 6   in instrumental observations. Moreover any attribution of possible human influence on the Atlantic Modes
 7   and associated teleconnections is limited by the poor fidelity of CMIP5 and CMIP6 models in reproducing
 8   the mean tropical Atlantic climate, its seasonality and variability, despite hints of some improvement in
 9   CMIP6, as well as other sources of uncertainties related to limited process understanding in the observations
10   (Foltz et al., 2019), the response of the tropical Atlantic climate to anthropogenic aerosol forcing (Booth et
11   al., 2012; Zhang et al., 2013a) and the presence of strong multidecadal fluctuations related to AMV (Section
12   3.7.7) and cross-tropical basin interactions (Martín-Rey et al., 2018; Cai et al., 2019). The fact that most
13   models poorly represent the climatology and variability of the tropical Atlantic combined with the short
14   observational record makes it difficult to place the recent observed changes in the context of internal multi-
15   annual variability versus anthropogenic forcing.
17   In summary, based on CMIP5 and CMIP6 results, there is no robust evidence that the observed changes in
18   either the Atlantic Niño or AMM modes and associated teleconnections over the second half of the 20th
19   century are beyond the range of internal variability or have been influenced by natural or anthropogenic
20   forcing. Considering the physical processes responsible for model biases in these modes, increasing
21   resolution in both ocean and atmosphere components may be an opportunity for progress in the simulation of
22   the tropical Atlantic changes as evidenced by some individual model studies (Roberts et al., 2018), but this
23   needs confirmation from a multi-model perspective.
26   3.7.6   Pacific Decadal Variability
28   Pacific decadal variability (PDV) is the generic term for the modes of variability in the Pacific Ocean that
29   vary on decadal to interdecadal timescales. PDV and its related teleconnections encompass the Pacific
30   Decadal Oscillation (PDO; Mantua et al. 1997; Mantua and Hare 2002; Zhang et al. 1997), and an
31   anomalous SST pattern in the North Pacific, as well as a broader structure associated with Pacific-wide SSTs
32   termed the Interdecadal Pacific Oscillation (IPO; Power et al. 1999; Folland et al. 2002; Henley et al. 2015).
33   Since the PDO and IPO indices are highly correlated, this section assesses them together as the PDV (Annex
34   IV.2.6).
36   AR5 mentioned an overall limited level of evidence for both CMIP3 and CMIP5 evaluation of the Pacific
37   modes at interdecadal timescales, leading to low confidence statements about the models’ performance in
38   reproducing PDV (Flato et al., 2013) and similarly low confidence in the attribution of observed PDV
39   changes to human influence (Bindoff et al., 2013).
41   The implication of PDV in the observed slowdown of the GMST warming rate in the early 2000s (Cross-
42   Chapter Box 3.1) has triggered considerable research on decadal climate variability and predictability since
43   the AR5 (Meehl et al., 2013, 2016b; England et al., 2014; Dai et al., 2015; Steinman et al., 2015; Kosaka and
44   Xie, 2016; Cassou et al., 2018). Many studies find that the broad spatial characteristics of PDV are
45   reasonably well represented in unforced climate models (Henley 2017; Newman et al. 2016) and in historical
46   simulations in CMIP5 and CMIP6 (Figure 3.39), although there is sensitivity to methodology used to remove
47   the externally-forced component of the SST (Bonfils and Santer, 2011; Xu and Hu, 2018). Compared with
48   CMIP3 models, CMIP5 models exhibit overall slightly better performance in reproducing PDV and
49   associated teleconnections (Polade et al., 2013; Joshi and Kucharski, 2017), and also smaller inter-model
50   spread (Lyu et al., 2016). CMIP6 models on average show slightly improved reproduction of the PDV spatial
51   structure than CMIP5 models (Figure 3.39a-c; Fasullo et al., 2020). SST anomalies in the subtropical South
52   Pacific lobe are however too weak relative to the equatorial and North Pacific lobes in CMIP5 pre-industrial
53   control and historical simulations (Henley et al., 2017), a bias that remains in CMIP6 (Figure 3.39b).
55   Biases in the PDV temporal properties and amplitude are present in CMIP5 (Cheung et al., 2017; Henley,
     Do Not Cite, Quote or Distribute                     3-86                                      Total pages: 202
     Final Government Distribution                        Chapter 3                                     IPCC AR6 WGI

 1   2017). While model evaluation is severely hampered by short observational records and incomplete
 2   observational coverage before satellite measurements started, the duration of PDV phases appears to be
 3   shorter in coupled models than in observations, and correspondingly the ratio of decadal to interannual
 4   variance is underestimated (Henley et al. 2017; Figure 3.39e,f). This apparent bias may be associated with
 5   overly biennial behaviour of Pacific trade wind variability and related ENSO activity, leaving too weak
 6   variability on decadal timescales (Kociuba and Power, 2015). ENSO influence on the extratropical North
 7   Pacific Ocean at decadal timescales is also very diverse among both CMIP3 and CMIP5 models, being
 8   controlled by multiple factors (Nidheesh et al., 2017). In terms of amplitude, the variance of the PDV index
 9   after decadal filtering is significantly weaker in the concatenated CMIP5 ensemble than the three
10   observational estimates used in Figure 3.39e (p < 0.1 with an F-test). Consequently, the observed PDV
11   fluctuations over the historical period often lie in the tails of the model distributions (Figure 3.39e,f). Even if
12   one cannot rule out that the observed PDV over the instrumental era represents an exceptional period of
13   variability, it is plausible that the tendency of the CMIP5 models to systematically underestimate the low
14   frequency variance is due to an incomplete representation of decadal-scale mechanisms in these models. This
15   situation is slightly improved in CMIP6 historical simulations but remains a concern (Fasullo et al., 2020).
16   The results of McGregor et al. (2018) suggest that the underrepresentation of the magnitude stems from
17   Atlantic mean SST bias (Section through inter-basin coupling.
19   While PDV is primarily understood as an internal mode of variability (Si and Hu, 2017), there are some
20   indications that anthropogenically induced SST changes project onto PDV and have contributed to its past
21   evolution (Bonfils and Santer, 2011; Dong et al., 2014a; Boo et al., 2015; Xu and Hu, 2018). However, the
22   level of evidence is limited because of the difficulty in correctly separating internal versus externally forced
23   components in the observed SST variations, and because it is unclear whether the dynamics of the PDV are
24   operative in this forced SST change pattern. Over the last two to three decades which encompass the period
25   of slower GMST increase (Cross-Chapter Box 3.1), Smith et al. (2016) found that anthropogenic aerosols
26   have driven part of the PDV change toward its negative phase. A similar result is shown in Takahashi and
27   Watanabe (2016) who found intensification of the Pacific Walker circulation in response to aerosol forcing
28   (Section Indeed, CMIP6 models simulate a negative PDV trend since the 1980s (Figure 3.39f),
29   which is much weaker than internal variability. However, a response to anthropogenic aerosols is not
30   robustly identified in a large ensemble of a model (Oudar et al., 2018), across CMIP5 models (Hua et al.,
31   2018), or in idealized model simulations (Kuntz and Schrag, 2016). Alternatively, inter-basin teleconnections
32   associated with the warming of the North Atlantic Ocean related to the mid-1990s phase shift of the AMV
33   (McGregor et al., 2014; Li et al., 2015d; Chikamoto et al., 2016; Kucharski et al., 2016; Ruprich-Robert et
34   al., 2017), and also in the Indian Ocean (Luo et al., 2012; Mochizuki et al., 2016), could have favoured a
35   delayed PDV transition to its negative phase in the 2000s. Considering the possible influence of external
36   forcing on Indian Ocean decadal variability (Section 3.7.4) and AMV (Section 3.7.7), any such human
37   influence on PDV would be indirect through changes in these ocean basins, and then imported to the Pacific
38   via inter-basin coupling. However, this human influence on AMV, and how consistently the inter-basin
39   processes affect PDV phase shifts, are uncertain. Other modelling studies find that anthropogenic aerosols
40   can influence the PDV (Verma et al., 2019; Amiri-Farahani et al., 2020; Dow et al., 2020). It is however
41   unclear whether and how much those forcings contributed to the observed variations of PDV. In CMIP6
42   models, the temporal correlation of the multi-model ensemble mean PDV index with its observational
43   counterpart is insignificant and negligible (Figure 3.39f), suggesting that any externally-driven component in
44   historical PDV variations was weak. Lastly, the multi-model ensemble mean computed from CMIP6
45   historical simulations shows slightly stronger variation than the CMIP5 counterpart, suggesting a greater
46   simulated influence from external forcings in CMIP6. Still, the fraction of the forced signal to the total PDV
47   is very low (Figure 3.39f), in contrast to AMV (Section 3.7.7). Consistently, Liguori et al. (2020) estimate
48   that the variance fraction of the externally-driven to total PDV is up to only 15% in a multi-model large
49   ensemble of historical simulations. These findings support an assessment that PDV is mostly driven by
50   internal variability since the pre-industrial era. The sensitivity of ensemble-mean PDV trends to the
51   ensemble size (Oudar et al., 2018), and the dominance of the ensemble spread over the ensemble mean in the
52   60-year trend of the equatorial Pacific zonal SST gradient in large ensemble simulations (Watanabe et al.,
53   2021), also support this statement.
55   In CMIP5 last millennium simulations, there is no consistency in temporal variations of PDV across the
     Do Not Cite, Quote or Distribute                       3-87                                        Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 1   ensemble (Fleming and Anchukaitis, 2016). This supports the notion that PDV is internal in nature.
 2   However, this issue remains difficult to firmly conclude because paleoclimate reconstructions of PDV have
 3   too poor a level of agreement for a rigorous model evaluation in past millennia (Henley, 2017).
 5   To conclude, there is high confidence that internal variability has been the main driver of the PDV since pre-
 6   industrial times, despite some modelling evidence for potential external influence. This assessment is
 7   supported by studies based on large ensemble simulations that found the dominance of internally-driven
 8   PDV, and the CMIP6-based assessment (Figure 3.39). As such, PDV is an important driver of decadal
 9   internal climate variability which limits detection of human influence on various aspects of decadal climate
10   change on global to regional scales (high confidence). Model evaluation of PDV is hampered by short
11   observational records, spatial incompleteness of observations before the satellite observation era, and poor
12   agreement among paleoclimate reconstructions. Despite the limitations of these model-observation
13   comparisons, CMIP5 models, on average, simulate broadly realistic spatial structures of the PDV, but with a
14   clear bias in the South Pacific (medium confidence). CMIP5 models also very likely underestimate PDV
15   magnitude. CMIP6 models tend to show better overall performance in spatial structure and magnitude of
16   PDV, but there is low confidence in this assessment due to the lack of literature.
21   Figure 3.39: Model evaluation of the Pacific Decadal Variability (PDV). (a, b) Sea surface temperature (SST)
22                anomalies (ºC) regressed onto the Tripole Index (TPI; Henley et al., 2015) for 1900-2014 in (a) ERSSTv5
23                and (b) CMIP6 multi-model ensemble (MME) mean composite obtained by weighting ensemble
24                members by the inverse of the model ensemble size. A 10-year low-pass filter was applied beforehand.
25                Cross marks in (a) represent regions where the anomalies are not significant at the 10% level based on t-
26                test. Diagonal lines in (b) indicate regions where less than 80% of the runs agree in sign. (c) A Taylor
27                diagram summarizing the representation of the PDV pattern in CMIP5 (each a cross in light blue, and the
28                weighted multi-model mean as a dot in dark blue), CMIP6 (each ensemble member is shown as a cross in
29                red, weighted multi-model mean as a dot in orange) and observations over [40ºS-60ºN, 110ºE-70ºW]. The
30                reference pattern is taken from ERSSTv5 and black dots indicate other observational products
31                (HadISSTv1 and COBE-SST2). (d) Autocorrelation of unfiltered annual TPI at lag 1 year and 10-year
32                low-pass filtered TPI at lag 10 years for observations over 1900-2014 (horizontal lines) and 115-year
33                chunks of pre-industrial control simulations (open boxes) and individual historical simulations over 1900-
34                2014 (filled boxes) from CMIP5 (blue) and CMIP6 (red). (e) As in (d), but standard deviation of the
35                unfiltered and filtered TPI (ºC). Boxes and whiskers show weighted multi-model mean, interquartile
36                ranges and 5th and 95th percentiles. (f) Time series of the 10-year low-pass filtered TPI (ºC) in ERSSTv5,
37                HadISSTv1 and COBE-SST2 observational estimates (black) and CMIP5 and CMIP6 historical
38                simulations. The thick red and light blue lines are the weighted multi-model mean for the historical
39                simulations in CMIP5 and CMIP6, respectively, and the envelopes represent the 5th-95th percentile range
40                across ensemble members. The 5-95% confidence interval for the CMIP6 multi-model mean is given in
41                thin dashed lines. Further details on data sources and processing are available in the chapter data table
42                (Table 3.SM.1).
44   [END FIGURE 3.39 HERE]
47   3.7.7   Atlantic Multidecadal Variability
49   Atlantic Multidecadal Variability (AMV) refers to a climate mode representing basin-wide multidecadal
50   fluctuations in surface temperatures in the North Atlantic (Figure 3.40a,f), with teleconnections particularly
51   pronounced over the adjacent continents and the Arctic. The AMV phenomenon is usually assessed through
52   SST anomalies averaged over the entire North Atlantic basin, hereafter the AMV index, but it is associated
53   with many physical processes including three-dimensional ocean circulation, such as AMOC fluctuations
54   (Section 3.5.4), gyre adjustments, and salt and heat transport in the entire North Atlantic and subarctic
55   basins. The AMV, together with the PDV, has been shown to have modulated GSAT on multi-decadal
56   timescales since pre-industrial time (Cross Chapter-Box 3.1, Wu et al., 2019; Li et al., 2020). A detailed
57   description of the AMV as well as its associated teleconnection over land is given in Annex IV.2.7.
     Do Not Cite, Quote or Distribute                         3-88                                         Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 2   In AR5, climate models suggested that the AMV was primarily internally-driven alongside some
 3   contribution from external forcings (mainly anthropogenic aerosols) over the late 20th century. But AR5 also
 4   concluded that models show medium performance in reproducing the observed AMV, with difficulties in
 5   simulating the timescale, the spatial structure and the coherency between all the physical processes involved.
 7   Climate models analyzed since AR5 continue to simulate AMV-like variability as part of their internal
 8   variability. This statement is mostly based on CMIP5 pre-industrial control and historical simulations
 9   (Wouters et al., 2012; Schmith et al., 2014; Menary et al., 2015; Ruprich-Robert and Cassou, 2015; Brown et
10   al., 2016b; Chen et al., 2016; Kim et al., 2018b) and is also true for the CMIP6 models (Menary et al., 2018;
11   Voldoire et al., 2019b). Models also continue to support links to a wide array of remote climate influences
12   through atmospheric teleconnections (Martin et al., 2013; Ruprich-Robert et al., 2017, 2018; Monerie et al.,
13   2019; Qasmi et al., 2020; Ruggieri et al., 2020). Even if debate remains (Clement et al., 2015; Cane et al.,
14   2017; Mann et al., 2020), there is now stronger evidence for a crucial role of oceanic dynamics in internal
15   AMV that is primarily linked to the AMOC and its interplay with the NAO (Zhang et al., 2013a; Müller et
16   al., 2015; O’Reilly et al., 2016a, 2019b; Delworth et al., 2017; Zhang, 2017; Kim et al., 2019; Sun et al.,
17   2019). However, considerable diversity in the spatio-temporal properties of the simulated AMV is found in
18   both pre-industrial control and historical CMIP5 experiments (Zhang and Wang, 2013; Wills et al., 2019).
19   Such model diversity is presumably associated with the wide range of coupled processes associated with
20   AMV (Baker et al., 2017; Woollings et al., 2018a) including large-scale atmospheric teleconnections and
21   regional feedbacks relating to tropical clouds, Arctic sea ice in the subarctic basins and Saharan dust, whose
22   relative importance and interactions across timescales are specific to each model (Brown et al., 2016; Martin
23   et al, 2014).
25   Additional studies since AR5 corroborate that CMIP5-era models tend to underestimate many aspects of
26   observed AMV and its SST fingerprint. On average, the duration of the modelled AMV episodes is too short,
27   its magnitude is too weak and its basin-wide SST spatial structure is limited by the poor representation of the
28   link between the tropical North Atlantic and the subpolar North Atlantic/Nordic seas (Martin et al., 2013;
29   Qasmi et al., 2017). Such mismatches between observed and simulated AMV (Figure 3.40c-e) have been
30   associated with intrinsic model biases in both mean state (Menary et al., 2015; Drews and Greatbatch, 2016)
31   and variability in the ocean and overlying atmosphere. For instance, compared to available observational
32   data CMIP5 models underestimate the ratio of decadal to interannual variability of the main drivers of AMV,
33   namely the AMOC, NAO and related North Atlantic jet variations (Bracegirdle et al., 2018; Kim et al.,
34   2018a; Simpson et al., 2018; Yan et al., 2018) (Section 3.7.1), which has strong implications for the
35   simulated temporal statistics of AMV, AMV-induced teleconnections (Ault et al., 2012; Menary et al., 2015)
36   and AMV predictability.
38   The increase of AMV variance in CMIP6 models (stronger magnitude and longer duration) seems to be
39   explained by the enhanced variability in the subpolar North Atlantic SST (Figure 3.40b,c), which is
40   particularly pronounced in some models, associated with greater variability in the AMOC (Voldoire et al.,
41   2019a; Boucher et al., 2020; Section 3.5.4) and greater GMST multidecadal variability (Voldoire et al.,
42   2019b; Parsons et al., 2020; Section 3.3.1) (Figure 3.40c-f). The decadal variance in SST in the subpolar
43   North Atlantic seems now to be slightly overestimated in CMIP6 compared to observational estimates, while
44   the AMV-related tropical SST anomalies remains weaker in line with CMIP5 (Figure 3.40b). The
45   mechanisms producing the tropical-extratropical relationship at decadal timescales remain poorly understood
46   despite stronger evidence since AR5 for the importance of the subpolar gyre SST anomalies in generating
47   tropical changes through atmospheric teleconnection (Caron et al., 2015; Ruprich-Robert et al., 2017; Kim et
48   al., 2020). Significant discrepancies remain in the simulated AMV spatial pattern when historical simulations
49   are compared to multivariate observations (Yan et al., 2018; Robson et al., 2020).
51   There is additional evidence since AR5 that external forcing has been playing an important role in shaping
52   the timing and intensity of the observed AMV since pre-industrial times (Bellomo et al., 2018; Andrews et
53   al., 2020). The time synchronisation between observed and multi-model mean AMV SST indices is
54   significant in both CMIP5 and CMIP6 historical simulations, while the explained variance of the forced
55   response in CMIP6 appears stronger (Figure 3.40d-f). The competition between greenhouse gas warming
     Do Not Cite, Quote or Distribute                     3-89                                      Total pages: 202
     Final Government Distribution                         Chapter 3                                      IPCC AR6 WGI

 1   and anthropogenic sulphate aerosol cooling has been proposed to be particularly important over the latter
 2   half of the 20th century (Booth et al., 2012; Steinman et al., 2015; Murphy et al., 2017; Undorf et al., 2018a;
 3   Haustein et al., 2019). The latest observed AMV shift from the cold to the warm phase in the mid-1990s at
 4   the surface ocean is well captured in the CMIP6 forced component and may be associated with the lagged
 5   response to increased AMOC due to strong anthropogenic aerosol forcing over 1955-1985 (Menary et al.,
 6   2020) in combination with the rapid response through surface flux processes to declining aerosol forcing and
 7   increasing greenhouse gas influence since then. However, natural forcings may have also played a significant
 8   role. For instance, volcanic forcing has been shown to contribute in part to the cold phases of the AMV-
 9   related SST anomalies observed in the 20th century (Terray, 2012; Bellucci et al., 2017; Swingedouw et al.,
10   2017; Birkel et al., 2018). Over the last millenium, natural forcings including major volcanic eruptions and
11   fluctuations in solar activity are thought to have driven a larger fraction of the multidecadal variations in the
12   AMV, with some interplay with internal processes (Otterå et al., 2010; Knudsen et al., 2014; Moffa-Sánchez
13   et al., 2014; Wang et al., 2017b; Malik et al., 2018; Mann et al., 2021), but other studies question the role of
14   natural forcings over this period (Zanchettin et al., 2014; Lapointe et al., 2020).
16   Model evaluation of the AMV phenomenon remains difficult because of short observational records
17   (especially of detailed process-based observations), the lack of stationarity in the variance, spatial patterns
18   and frequency of the AMV assessed from modelled SST (Qasmi et al., 2017), difficulties in estimating the
19   forced signals in both historical simulations and observations (Tandon and Kushner, 2015), and because of
20   probable interplay between internally and externally-driven processes (Watanabe and Tatebe, 2019).
21   Furthermore, models simulate a large range of historical anthropogenic aerosol forcing (Smith et al., 2020)
22   and questions often referred to as signal-to-noise paradox have been raised concerning the models’ ability to
23   correctly simulate the magnitude of the response of AMV-related atmospheric circulation phenomena, such
24   as the NAO (Section 3.7.1), to both internally and externally generated changes (Scaife and Smith, 2018).
25   Related methodological and epistemological uncertainties also call into question the relevance of the
26   traditional basin-average SST index to assessing the AMV phenomenon (Zanchettin et al., 2014; Frajka-
27   Williams et al., 2017; Haustein et al., 2019; Wills et al., 2019).
29   To summarize, results from CMIP5 and CMIP6 models together with new statistical techniques to evaluate
30   the forced component of modelled and observed AMV, provide robust evidence that external forcings have
31   modulated AMV over the historical period. In particular, anthropogenic and volcanic aerosols are thought to
32   have played a role in the timing and intensity of the negative (cold) phase of AMV recorded from the mid-
33   1960s to mid-1990s and subsequent warming (medium confidence). However, there is low confidence in the
34   estimated magnitude of the human influence. The limited level of confidence is primarily explained by
35   difficulties in accurately evaluating model performance in simulating AMV. The evaluation is severely
36   hampered by short instrumental records but also, equally importantly, by the lack of detailed and coherent
37   long-term process-based observations (for example of the AMOC, aerosol optical depth, surface fluxes and
38   cloud changes), which limit our process understanding. In addition, studies often rely solely on simplistic
39   SST indices that may be hard to interpret (Zhang et al., 2016) and may mask critical physical inconsistencies
40   in simulations of the AMV compared to observations (Zhang, 2017).
45   Figure 3.40: Model evaluation of Atlantic Multi-decadal Variability (AMV). (a, b) Sea surface temperature (SST)
46                anomalies (ºC) regressed onto the AMV index defined as the 10-year low-pass filtered North Atlantic (0º-
47                60°N, 80°W-0°E) area-weighted SST* anomalies over 1900-2014 in (a) ERSSTv5 and (b) the CMIP6
48                multi-model ensemble (MME) mean composite obtained by weighting ensemble members by the inverse
49                of each model’s ensemble size. The asterisk denotes that the global mean SST anomaly has been removed
50                at each time step of the computation. Cross marks in (a) represent regions where the anomalies are not
51                significant at the 10% level based on a t-test. Diagonal lines in (b) show regions where less than 80% of
52                the runs agree in sign. (c) A Taylor diagram summarizing the representation of the AMV pattern in
53                CMIP5 (each member is shown as a cross in light blue, and the weighted multi-model mean is shown as a
54                dot in dark blue), CMIP6 (each member is shown as a cross in red, and the weighted multi-model mean is
55                shown as a dot in orange) and observations over [0º-60°N, 80°W-0°E]. The reference pattern is taken
56                from ERSSTv5 and black dots indicate other observational products (HadISSTv1 and COBE-SST2). (d)
     Do Not Cite, Quote or Distribute                        3-90                                         Total pages: 202
     Final Government Distribution                        Chapter 3                                       IPCC AR6 WGI

 1                Autocorrelation of unfiltered annual AMV index at lag 1 year and 10-year low-pass filtered AMV index
 2                at lag 10 years for observations over 1900-2014 (horizontal lines) and 115-year chunks of pre-industrial
 3                control simulations (open boxes) and individual historical simulations over 1900-2014 (filled boxes) from
 4                CMIP5 (blue) and CMIP6 (red). (e) As in (d), but showing standard deviation of the unfiltered and
 5                filtered AMV indices (ºC). Boxes and whiskers show the weighted multi-model mean, interquartile
 6                ranges and 5th and 95th percentiles. (f) Time series of the AMV index (ºC) in ERSSTv5, HadISSTv1 and
 7                COBE-SST2 observational estimates (black) and CMIP5 and CMIP6 historical simulations. The thick red
 8                and light blue lines are the weighted multi-model mean for the historical simulations in CMIP5 and
 9                CMIP6, respectively, and the envelopes represent the 5th-95th percentile range obtained from all
10                ensemble members. The 5- 95% confidence interval for the CMIP6 multi-model mean is shown by the
11                thin dashed line. Further details on data sources and processing are available in the chapter data table
12                (Table 3.SM.1).
14   [END FIGURE 3.40 HERE]
17   3.8     Synthesis across Earth System Components
19   3.8.1    Multivariate Attribution of Climate Change
21   AR5 concluded that human influence on the climate system is clear (IPCC, 2013), based on observed
22   increasing greenhouse gas concentrations in the atmosphere, positive radiative forcing, observed warming,
23   and physical understanding of the climate system. AR5 also assessed that it was virtually certain that internal
24   variability alone could not account for observed warming since 1951 (Bindoff et al., 2013). Evidence has
25   grown since AR5 that observed changes since the 1950s in many parts of the climate system are attributable
26   to anthropogenic influence. So far, this chapter has focused on examining individual aspects of the climate
27   system in separate sections. The results presented in Sections 3.3 to 3.7 substantially strengthen our
28   assessment of the role of human influence on climate since pre-industrial times. In this section we look
29   across the whole climate system to assess to what extent a physically consistent picture of human induced
30   change emerges across the climate system (Figure 3.41).
32   The observed global surface air temperature warming of 0.9-1.2°C in 2010-2019 is much greater than can be
33   explained by internal variability (likely -0.2°C-0.2°C) or natural forcings (likely -0.1°C-0.1°C) alone, but
34   consistent with the assessed anthropogenic warming (likely 0.8°C-1.3°C) (Section Moreover it is
35   very likely that human-induced greenhouse gas increases were the main driver of tropospheric warming since
36   1979 (Section and virtually certain that greenhouse gas forcing was the main driver of the observed
37   changes in hot and cold extremes on the global scale (Cross-Chapter Box 3.2).
39   As might be expected from a warming atmosphere, moisture in the troposphere has increased and
40   precipitation patterns have changed. Human influence has likely contributed to the observed changes in
41   humidity and precipitation (Section 3.3.2). It is likely that human influence, in particular due to greenhouse
42   gas forcing, is the main driver of the observed intensification of heavy precipitation in global land regions
43   during recent decades (Cross-Chapter Box 3.2). The pattern of ocean salinity changes indicate that fresh
44   regions are becoming fresher and that salty regions are becoming saltier as a result of changes in ocean-
45   atmosphere fluxes through evaporation and precipitation (high confidence) making it extremely likely that
46   human influence has contributed to observed near-surface and subsurface salinity changes since the mid-20th
47   century (Section Taken together this evidence indicates a human influence on the water cycle.
49   It is very likely that human influence was the main driver of Arctic sea ice loss since the late 1970s (3.4.1),
50   and very likely that it contributed to the observed reductions in Northern Hemisphere springtime snow cover
51   since 1950 (3.4.2). Human influence was very likely the main driver of the recent global, near-universal
52   retreat of glaciers and it is very likely that it contributed to the observed surface melting of the Greenland Ice
53   Sheet over the past two decades (Sections 3.4.3). It is extremely likely that human influence was the main
54   driver of the ocean heat content increase observed since the 1970s (Section 3.5.1), and very likely that human
55   influence was the main driver of the observed global mean sea level rise since at least 1970 (Section 3.5.3).
     Do Not Cite, Quote or Distribute                        3-91                                         Total pages: 202
     Final Government Distribution                           Chapter 3                                       IPCC AR6 WGI

 1   Combining the evidence from across the climate system (Sections 3.3-3.7) increases the level of confidence
 2   in the attribution of observed climate change to human influence and reduces the uncertainties associated
 3   with assessments based on a single variable. From this combined evidence, it is unequivocal that human
 4   influence has warmed the global climate system.
 9   Figure 3.41: Summary figure showing simulated and observed changes in key large-scale indicators of climate
10                change across the climate system, for continental, ocean basin and larger scales. Black lines show
11                observations, brown lines and shading show the multi-model mean and 5th-95th percentile ranges for
12                CMIP6 historical simulations including anthropogenic and natural forcing, and blue lines and shading
13                show corresponding ensemble means and 5th-95th percentile ranges for CMIP6 natural-only simulations.
14                Temperature time series are as in Figure 3.9, but with smoothing using a low pass filter. Precipitation
15                time series are as in Figure 3.15 and ocean heat content as in Figure 3.26. Further details on data sources
16                and processing are available in the chapter data table (Table 3.SM.1).
18   [END FIGURE 3.41 HERE]
21   3.8.2     Multivariate Model Evaluation
23   Similar to the assessment of multivariate attribution of climate change in the previous section, this section
24   covers the performance of the models across different variables (Sections and different classes of
25   models (Section Here the focus is on a system-wide assessment using integrative measures of model
26   performance that characterize model performance using multiple diagnostic fields derived from multi-model
27   ensembles.
30    Integrative Measures of Model Performance
32   The purpose of this section is to use multivariate analyses to address how well models simulate present-day
33   and historical climate. For every diagnostic field considered, model performance is compared to one or
34   multiple observational references, and the quality of the simulation is expressed as a single number, for
35   example a correlation coefficient or a root mean square difference versus the observational reference. By
36   simultaneously assessing different performance indices, model improvements can be quantified, similarities
37   in behaviour between different models become apparent, and dependencies between various indices become
38   evident (Gleckler et al., 2008; Waugh and Eyring, 2008).
40   AR5 found significant differences between models in the simulation of mean climate in the CMIP5 ensemble
41   when measured against meteorological reanalyses and observations (Flato et al., 2013), see also Stouffer et
42   al. (2017). AR5 determined that for the diagnostic fields analysed, the models usually compared similarly
43   against two different reference datasets, suggesting that model errors were generally larger than
44   observational uncertainties or other differences between the observational references. In agreement with
45   previous assessments, the CMIP5 multi-model mean generally performed better than individual models
46   (Annan and Hargreaves, 2011; Rougier, 2016). AR5 considered 13 atmospheric fields in its assessment for
47   the instrumental period but did not assess multi-variate model performance in other climate domains (e.g.,
48   ocean, land, and sea ice). AR5 found only modest improvement regarding the simulation of climate for two
49   periods of the Earth’s history (the Last Glacial Maximum and the mid-Holocene) between CMIP5 and
50   previous paleoclimate simulations. Similarly, for the modern period only modest, incremental progress was
51   found between CMIP3 and CMIP5 regarding the simulation of precipitation and radiation. The
52   representation of clouds also showed improvement, but remained a key challenge in climate modelling (Flato
53   et al., 2013).
55   The type of multi-variate analysis of models presented in AR5 remains critical to building confidence for
56   example in projections of climate change. It is expanded here to the previous-generation CMIP3 and present-
     Do Not Cite, Quote or Distribute                          3-92                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   generation CMIP6 models and also to more variables and more climate domains, covering land and ocean as
 2   well as sea ice. The multi-variate evaluation of these three generations of models is performed relative to the
 3   observational datasets listed in Annex I, Table AI.1. For many of these datasets, a rigorous characterization
 4   of the observational uncertainty is not available, see discussion in Chapter 2. Here, as much as possible,
 5   multiple independent observational datasets are used. Disagreements among them would cause differences in
 6   model scoring, indicating that observational uncertainties may be substantial compared to model errors.
 7   Conversely, similar scores against different observational datasets would suggest model biases may be larger
 8   than the observational uncertainty.
10   An analysis of a basket of 16 atmospheric variables (Figure 3.42a) assessed across CMIP3, CMIP5, and
11   CMIP6 models but excluding high-resolution models participating in HighResMIP, reveals the progress
12   made between these three generations of models (Bock et al., 2020). Progress is evidenced by the increasing
13   prevalence of blue colours (indicating a performance better than the median) for the more recent model
14   versions. Additionally, a few CMIP6 models outperform the best-performing CMIP5 models. Progress is
15   evident across all 16 variables. As noted in AR5, the models typically score similarly against both
16   observational reference datasets, indicating that indeed uncertainties in these references are smaller than
17   model biases. Several models and model families perform better compared to observational references than
18   the median, across a majority of the climate variables assessed, and conversely some other models or model
19   families compare more poorly against these reference datasets. Such a good correspondence across a range
20   of diagnostic fields probing different aspects of climate enhances confidence that the improved performances
21   reflect progress in the physical realism of these simulations. An alternative explanation, that progress is due
22   to a cancellation of errors achieved by model tuning, appears improbable given the large number of
23   diagnostic fields involved here. However, several instances of poor model performance (red colours in
24   Figure 3.42) still exist in the CMIP6 ensemble. Family relationships (i.e. various degrees of shared
25   formulations; Knutti et al., 2013) between the models are apparent, for example, the GISS, GFDL, CESM,
26   CNRM, and HadGEM/UKESM1/ACCESS families score similarly across all atmospheric variables, both for
27   the CMIP5 and CMIP6 generations. In the cases of CESM2 / CESM2(WACCM), CNRM-CM6-1 / CNRM-
28   ESM2-1, NorCPM1 / NorESM2-LM, and HadGEM3-GC31-LL / UKESM1-0-LL, the high-complexity
29   model scores as well or better than its lower-complexity counterpart, indicating that increasing complexity
30   by adding Earth system features, which by removing constraints could be expected to degrade a model’s
31   performance, does not necessarily do so. Several high climate-sensitivity models (Section 7.5; Meehl et al.,
32   2020), in particular CanESM5, CESM2, CESM2-WACCM, HadGEM3-GC31-LL, and UKESM1-0-LL,
33   score well against the benchmarks. In accordance with AR5 and earlier assessments, the multi-model mean,
34   with some notable exceptions, is better than any individual model (Annan and Hargreaves, 2011; Rougier,
35   2016).
37   Regarding model performance for the ocean and the cryosphere (Figure 3.42b), it is apparent that for many
38   models there are substantial differences between the scores for Arctic and Antarctic sea ice concentration.
39   This might suggest that it is not sea ice physics directly that is driving such differences in performance but
40   rather other influences, such as differences in geography, the role of large ice shelves (which are absent in
41   the Arctic), or large-scale ocean dynamics. As for atmospheric variables, progress is evident also across all 4
42   ocean and 10 land variables from CMIP5 to CMIP6.
44   In summary, CMIP6 models perform generally better for a basket of variables covering mean historical
45   climate across the atmosphere, ocean, and land domains than previous-generation and older models (high
46   confidence). Earth System models characterized by additional biogeochemical feedbacks often perform at
47   least as well as related more constrained, lower-complexity models lacking these feedbacks (medium
48   confidence). In many cases, the models score similarly against both observational references, indicating that
49   model errors are usually larger than observational uncertainties (high confidence). Moreover, synthesizing
50   across Sections 3.3-3.7, we assess that the CMIP6 multi-model mean captures most aspects of observed
51   climate change well (high confidence).
     Do Not Cite, Quote or Distribute                     3-93                                       Total pages: 202
     Final Government Distribution                          Chapter 3                                        IPCC AR6 WGI

 1   Figure 3.42: Relative space-time root-mean-square deviation (RMSD) calculated from the climatological
 2                seasonal cycle of the CMIP simulations (1980-1999) compared to observational datasets. (a) CMIP3,
 3                CMIP5, and CMIP6 for 16 atmospheric variables (b) CMIP5 and CMIP6 for 10 land variables and four
 4                ocean/sea-ice variables. A relative performance measure is displayed, with blue shading indicating better
 5                and red shading indicating worse performance than the median of all model results. A diagonal split of a
 6                grid square shows the relative error with respect to the reference data set (lower right triangle) and an
 7                additional data set (upper left triangle). Reference/additional datasets are from top to bottom in (a):
10                ERA5/NCEP, ERA5/NCEP, AIRS/ERA5, ERA5/NCEP and in (b): CERES-EBAF/-, CERES-EBAF/-,
11                CERES-EBAF/-, CERES-EBAF/-, LandFlux-EVAL/-, Landschuetzer2016/ JMA-TRANSCOM;
13                HadISST/-, HadISST/-, ERA-Interim/-. White boxes are used when data are not available for a given
14                model and variable. Figure is updated and expanded from Bock et al. (2020), their Figure 5 CC
15                BY4.0 Further details on data sources and processing are
16                available in the chapter data table (Table 3.SM.1).
17   [END FIGURE 3.42 HERE]
20   Using centred pattern correlations (quantifying pattern similarity on a scale of -1 to 1, with 1 expressing
21   perfect similarity and 0 no relationship) for selected fields, the AR5 documented improvements between
22   CMIP3 and CMIP5 in surface air temperature, outgoing longwave radiation, and precipitation (Figure 9.6 of
23   Flato et al., 2013). Little further progress between CMIP3 and CMIP5 was found for fields that were already
24   quite well simulated in CMIP3 (such as surface air temperature and outgoing longwave). For precipitation,
25   the spread reduced because the worst-performing models improved. The short-wave cloud radiative effect
26   remained relatively poorly simulated with significant inter-model spread (e.g., Calisto et al., 2014). This
27   comparison of centred pattern correlations is designed to help determine the quality of simulation of different
28   diagnostics relative to each other, and also to examine progress between generations of models. Figure 3.43
29   shows the centred pattern correlations for 16 variables for CMIP3, CMIP5 and CMIP6 models. In the
30   ensemble averages, CMIP6 performs better than CMIP5 and CMIP3 for near-surface temperature,
31   precipitation, mean sea-level pressure, and many other variables. For the variables shown, the uncertainties
32   in observational datasets, in particular for precipitation and northward wind at 850 hPa, remain substantial
33   relative to mean model errors (see grey dots in Figure 3.43).
38   Figure 3.43: Centred pattern correlations between models and observations for the annual mean climatology
39                over the period 1980–1999. Results are shown for individual CMIP3 (cyan), CMIP5 (blue) and CMIP6
40                (red) models (one ensemble member from each model is used) as short lines, along with the
41                corresponding ensemble averages (long lines). Correlations are shown between the models and the
42                primary reference observational data set (from left to right: ERA5, GPCP-SG, CERES-EBAF, CERES-
44                ERA5, AIRS, ERA5). In addition, the correlation between the primary reference and additional
45                observational data sets (from left to right: NCEP, GHCN, -, -, -, -, ERA5, HadISST, NCEP, NCEP,
46                NCEP, NCEP, NCEP, NCEP, ERA5, NCEP) are shown (solid grey circles) if available. To ensure a fair
47                comparison across a range of model resolutions, the pattern correlations are computed after regridding all
48                datasets to a resolution of 4º in longitude and 5º in latitude. Figure is updated and expanded from Bock et
49                al. (2020), their Figure 7 CC BY4.0 Further details on data
50                sources and processing are available in the chapter data table (Table 3.SM.1).
52   [END FIGURE 3.43 HERE]
55   In addition to the multivariate assessments of simulations of the recent historical period, simulations of
56   selected periods of the Earth’s more distant history can be used to benchmark climate models by exposing
57   them to climate forcings that are radically different from the present and recent past (Harrison et al., 2015,
58   2016; Kageyama et al., 2018b; Tierney et al., 2020a). These time periods provide an out-of-sample test of
     Do Not Cite, Quote or Distribute                          3-94                                         Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   models because they are not in general used in the process of model development. They encompass a range
 2   of climate drivers, such as volcanic and solar forcing for the Last Millennium, orbital forcing for the mid-
 3   Holocene and Last Interglacial, and changes in greenhouse gases and ice sheets for the LGM, mid-Pliocene
 4   Warm Period, and early Eocene (Sections 2.2 and 2.3). These drivers led to climate changes, including in
 5   surface temperature (Section and the hydrological cycle (Section, which are described by
 6   paleoclimate proxies that have been synthesised to support evaluations of models on a global and regional
 7   scale. However, the more sparse, indirect, and regionally incomplete climate information available from
 8   paleo-archives motivates a different form for the multivariate analysis of simulations available covering
 9   these periods versus the equivalent for the historical period, as described below.
11   AR5 found that reconstructions and simulations of past climates both show similar responses in terms of
12   large-scale patterns of climate change, such as polar amplification (Flato et al., 2013). However, for several
13   regional signals (e.g., the north-south temperature gradient in Europe and regional precipitation changes), the
14   magnitude of change seen in the proxies relative to the pre-industrial period was underestimated by the
15   models. When benchmarking CMIP5/PMIP3 models against reconstructions of the mid-Holocene and LGM,
16   AR5 found only a slight improvement compared with earlier model versions across a range of variables. For
17   the Last Interglacial, it was noted that the magnitude of observed annual mean warming in the NH was only
18   reached in summer in the models. For the mid-Pliocene Warm Period, it was noted that both proxies and
19   models showed a polar amplification of temperature compared with the pre-industrial period, but a formal
20   model evaluation was not carried out.
22   Since AR5, new simulation protocols have been developed in the framework of PMIP4 (Kageyama et al.,
23   2018b) which is further described for the mid-Holocene and the Last Interglacial by Otto-Bliesner et al.
24   (2017), for the LGM by Kageyama et al. (2017), for the Pliocene by Haywood et al. (2016), and for the early
25   Eocene by Lunt et al. (2017). These have resulted in new model simulations for these time periods (Brierley
26   et al., 2020; Haywood et al., 2020; Kageyama et al., 2021a; Lunt et al., 2021; Otto-Bliesner et al., 2021).
27   These time periods span an assessed temperature range of 20°C (Section, and for all periods the
28   PMIP4 model ensemble mean is within 0.5°C of the assessed range of GSAT (Figure 3.44a). Those time
29   periods for which the model ensemble mean is outside the assessed range of GSAT, the mid-Holocene and
30   the Last Interglacial, are primarily forced by orbital changes not greenhouse gas forcing, and as a result the
31   forcing, as well as the assessed and modelled response, are relatively close to zero in the global annual mean.
32   During these periods, climate change therefore is a consequence of more poorly understood Earth System
33   feedbacks acting on the response to orbital differences versus the present-day, affecting the seasonality of
34   insolation.
36   Polar amplification in the LGM, mid-Pliocene Warm Period, and Early Eocene climatic optimum (EECO)
37   simulations are assessed in Section 7.4.4. Here we focus on the mid-Holocene and the LGM, which have
38   been a part of AMIP or CMIP through several assessment cycles, and as such serve as a reference to quantify
39   regional model-data agreement from one IPCC assessment to another. We compare the results from 15
40   CMIP6 models using the PMIP4 protocol (CMIP6-PMIP4), with non-CMIP6 models using the PMIP4
41   protocol, with PMIP3 models, and with regional temperature and precipitation changes from proxies for the
42   mid-Holocene (Figure 3.44b). For 6 out of 7 variables shown, the CMIP6 multi-model mean captures the
43   correct sign of the change. For 5 out of 7 of them the CMIP6 ensemble mean is within the reconstructed
44   range. For the other two variables (changes in the mean temperature of the warmest month over North
45   America and in the mean annual precipitation over West Africa) nearly all PMIP4 and PMIP3 models are
46   outside the reconstructed range. CMIP6 models show regional patterns of temperature changes similar to the
47   PMIP3 ensemble (Brierley et al., 2020), but the slight mid-Holocene cooling in PMIP4 compared with
48   PMIP3, probably associated with lower imposed mid-Holocene carbon dioxide concentrations (Otto-Bliesner
49   et al., 2017), improves the regional model performance for summer and winter temperatures (Figure 3.44b).
50   However, this cooling also results in a CMIP6 mid-Holocene GSAT that lies further from the assessed range
51   (Figure 3.44a). All models show an expansion of the monsoon areas from the pre-industrial to the mid-
52   Holocene simulations in the NH, but this expansion in some cases is only large enough to cancel out the bias
53   in the pre-industrial control simulations (Brierley et al., 2020; Section There is a slight improvement
54   in representing the northward expansion of the West African monsoon region in PMIP4 compared with
55   PMIP3 (Figure 3.11, Figure 3.44b).
     Do Not Cite, Quote or Distribute                      3-95                                       Total pages: 202
     Final Government Distribution                          Chapter 3                                       IPCC AR6 WGI

 2   Fourteen simulations of the LGM climate have been produced following the CMIP6-PMIP4 protocol using
 3   11 models, 5 of which are from the latest CMIP6 generation. The multi-model-mean global cooling
 4   simulated by these models is close to that simulated by the CMIP5-PMIP3 ensemble, but the range of results
 5   is larger. The increase in the range is largely due to the inclusion of CESM2 which simulates a much larger
 6   cooling than the other PMIP4 models (Figure 3.44a). This is consistent with its larger climate sensitivity
 7   (Zhu et al., 2021) (see also Section The other models on average also simulate slightly larger
 8   cooling, possibly a consequence of the lower ice sheet assumed in PMIP4 versus PMIP3 (Kageyama et al.,
 9   2017, 2021a). The PMIP4 multi-model mean is within the range of reconstructed regional averages for 4 out
10   of 7 regional variables; this is unchanged from PMIP3 but for different variables (Figure 3.44c). For all
11   fields, the results of many individual models are outside the reconstructed range. For 2 variables out of 7
12   (changes in the mean temperature of the warmest month and mean annual precipitation over Western
13   Europe) no model is within the range of the reconstructions. This analysis is strengthened compared with the
14   equivalent analysis in AR5 because it is based on larger and improved reconstructions (Cleator et al., 2020).
15   Most CMIP6-PMIP4 models simulate slightly stronger AMOC, but no strong deepening of the AMOC
16   (Kageyama et al., 2021a), while most other PMIP4 models simulate a strengthening and strong deepening of
17   the AMOC, as was the case for the PMIP3 models (Muglia and Schmittner, 2015; Sherriff-Tadano et al.,
18   2018). Only one model (CESM1.2) shows a shoaling of LGM AMOC which is consistent with
19   reconstructions (Marzocchi and Jansen, 2017; Sherriff-Tadano et al., 2018).
21   17 PMIP4 models completed Last Interglacial simulations (Otto-Bliesner et al., 2021). The comparison to
22   reconstructions is generally good, except for some discrepancies, such as for upwelling systems such as in
23   the South East Atlantic or resulting from local melting of remnant ice sheets absent in the Last Interglacial
24   simulation protocol. All models simulate a decrease in Arctic sea ice in summer, commensurate with
25   increased summer insolation, while some models even simulate a large or complete loss (Guarino et al.,
26   2020; Kageyama et al., 2021b). Sea ice reconstructions for the Central Arctic are however too uncertain to
27   evaluate this behaviour. The Last Interglacial simulations indicate a clear relationship between simulated sea
28   ice loss and model responses to increased greenhouse gas forcing (Kageyama et al., 2021b; Otto-Bliesner et
29   al., 2021).
31   Overall, the PMIP multi-model means agree very well (within 0.5°C of the assessed range) with GSAT
32   reconstructed from proxies across multiple time periods, spanning a range from 6°C colder than pre-
33   industrial (Last Glacial Maximum) to 14°C warmer than pre-industrial (Early Eocene Climate Optimum)
34   (high confidence). During the orbitally-forced mid-Holocene, the CMIP6 multi-model mean captures the
35   sign of the regional changes in temperature and precipitation in most regions assessed, and there have been
36   some regional improvements compared to AR5 (medium confidence). The limited number of CMIP6
37   simulations of the LGM hinders model evaluation of the multi-model mean, but for both LGM and mid-
38   Holocene, models tend to underestimate the magnitude of large changes (high confidence). Some long-
39   standing model-data discrepancies, such as a dry bias in North Africa in the mid-Holocene, have not
40   improved in CMIP6 compared with PMIP3 (high confidence).
45   Figure 3.44: Multivariate synopsis of paleoclimate model results compared to observational references. Data-
46                model comparisons for GSAT anomalies for five PMIP4 periods and for regional features for the (b) mid-
47                Holocene and (c) LGM periods, for PMIP3 and PMIP4 models. The results from the CMIP6 models are
48                specified as coloured dots. In (a) the light orange shading shows the very likely assessed ranges presented
49                in Section In (b) and (c), the regions and variables are defined as follows: North America (20-
50                50°N, 140-60°W), Western Europe (35-70°N, 10°W-30°E) and West Africa (0°-30°N, 10°W-30°E);
51                MTCO: Mean Temperature of the Coldest Month (°C), MTWA: Mean Temperature of the Warmest
52                Month (°C), MAP: Mean Annual Precipitation (mm/year). In (b) and (c) the ranges shown for the
53                reconstructions (Bartlein et al. (2011) for MH and Cleator et al. (2020) for LGM) are based on the
54                standard error given at each site: the average and associated standard deviation over each area is obtained
55                by computing 1000 times the average of randomly drawn values from the Gaussian distributions defined
56                at each site by the reconstruction mean and standard error; the light orange colour shows the ± 1 standard
     Do Not Cite, Quote or Distribute                         3-96                                          Total pages: 202
     Final Government Distribution                         Chapter 3                                       IPCC AR6 WGI

 1                deviation of these 1000 estimates. The dots on (b) and (c) show the average of the model output for grid
 2                points for which there are reconstructions. The ranges for the model results are based on an ensemble of
 3                1000 averages over 50 years randomly picked in the model output time series for each region and each
 4                variable: the mean ± one standard deviation is plotted for each model. Figure is adapted from Brierley et
 5                al. (2020), their Figure S3 for the mid-Holocene and from Kageyama et al. (2021), their Figure 12 for the
 6                LGM. Further details on data sources and processing are available in the chapter data table (Table
 7                3.SM.1).
 9   [END FIGURE 3.44 HERE]
12   Process Representation in Different Classes of Models
14   Based on new scientific insights and newly available observations, many improvements have been made to
15   models from CMIP5 to CMIP6, including changes in the representation of physics of the atmosphere, ocean,
16   sea ice, and land surface. In many cases, changes in the detailed representation of cloud and aerosol
17   processes have been implemented. The new generation of CMIP6 climate models also features increases in
18   spatial resolution, as well as inclusion of additional Earth system processes and new components (see further
19   details in Section and in Tables AII.5 and AII.6). Such changes to model physics and resolution are
20   often designed to improve the fitness-for-purpose of a model such as projecting regional aspects of climate
21   (Section 10.3) or to more fully represent feedbacks to make the models more fit for long-term climate
22   projections affected for example by carbon cycle feedbacks (see also Section
24   Factors affecting model performance include resolution, the type of dynamical core (spectral,. finite
25   difference or finite volume), physics parameters and parameterisations, model structure (e.g. many of the
26   coupled HighResMIP models (Haarsma et al., 2016) use the NEMO ocean model, affecting model diversity),
27   and the range and degree of process realism (e.g. for aerosols, atmospheric chemistry and other Earth System
28   components). This section particularly explores the influence of model resolution and of complexity on
29   model performance (see also Section 8.5.1).
31   A key advance in CMIP6 compared to CMIP5 is the presence of high-resolution models that have
32   participated in HighResMIP. Resolution alone can significantly affect a model’s performance, with some
33   effects propagating to the global scale. Recent studies have shown that enhancing the horizontal resolution of
34   models is seen to significantly affect aspects of large-scale circulation as well as improve the simulation of
35   small-scale processes and extremes when compared to CMIP3 and CMIP5 models (Haarsma et al., 2016; see
36   also Section 11.4.3), with some models approaching 10 km resolution in the atmosphere (Kodama et al.,
37   2021) or ocean (Caldwell et al., 2019; Gutjahr et al., 2019; Roberts et al., 2019; Chang et al., 2020; Semmler
38   et al., 2020).
40   As discussed in Section 3.3, CMIP6 models reproduce observed large-scale mean surface temperature
41   patterns as well as their CMIP5 predecessors, but biases in surface temperature in the mean of HighResMIP
42   models are smaller than those in the mean of the corresponding standard resolution CMIP6 configurations of
43   the same models (Figure 3.3; Section The extent and causes of improvements due to increased
44   horizontal resolutions in the atmosphere and ocean domains depend on the model (Kuhlbrodt et al., 2018;
45   Roberts et al., 2018, 2019; Sidorenko et al., 2019), although they typically involve better process
46   representation (for example of ocean currents and atmospheric storms) which can lead to reduced biases in
47   top of atmosphere radiation and cloudiness. Precipitation has likewise improved in CMIP6 versus CMIP5
48   models, but biases remain. The high resolution (< 25 km) class of models participating in HighResMIP
49   compares regionally better against observations than the standard resolution CMIP6 models (of order 100
50   km, Figure 3.13; Section, partly because of an improved representation of orographic (mountain-
51   induced) precipitation which constitutes a major fraction of precipitation on land, but other processes also
52   play an important role (Vannière et al., 2019). However, there are also large parts of the tropical ocean where
53   precipitation in high-resolution models is not improved compared to standard resolution CMIP6 models
54   (Vannière et al., 2019).
56   Additionally, the representation of surface and deeper ocean mean temperature is improved in models with
     Do Not Cite, Quote or Distribute                        3-97                                          Total pages: 202
     Final Government Distribution                      Chapter 3                                    IPCC AR6 WGI

 1   higher horizontal resolution (Sections and with systematic improvements in coupled tropical
 2   Atlantic sea surface temperature and precipitation biases at higher resolutions (Roberts et al., 2019, single
 3   model; Vannière et al., 2019), the North Atlantic cold bias (Bock et al., 2020, multi-model; Roberts et al.,
 4   2018, 2019; Caldwell et al., 2019; all single models) as well as deep ocean biases (Small et al., 2014; Griffies
 5   et al., 2015; Caldwell et al., 2019; Gutjahr et al., 2019; Roberts et al., 2019; Chang et al., 2020, all single
 6   model studies). Atlantic ocean transports (heat and volume) are also generally improved compared to
 7   observations (Grist et al., 2018; Caldwell et al., 2019; Docquier et al., 2019; Roberts et al., 2019, 2020c;
 8   Chang et al., 2020), as well as some aspects of air-sea interactions (Wu et al., 2019a, single model; Bellucci
 9   et al., 2021, multi-model). However, warm-biased sea surface temperatures in the Southern Ocean got worse
10   in comparison to standard resolution CMIP6 models (Bock et al., 2020). AR5 noted problems with the
11   simulation of clouds in this region which were later attributed to a lack of supercooled liquid clouds (Bodas-
12   Salcedo et al., 2016). Mesoscale ocean processes are critical to maintaining the Southern Ocean stratification
13   and response to wind forcing (Marshall and Radko, 2003; Hallberg and Gnanadesikan, 2006), and their
14   explicit representation requires even higher ocean resolution (Hallberg, 2013). Similarly, atmospheric
15   convection remains unresolved even in the highest-resolution climate models participating in HighResMIP.
16   However, there is also evidence of improvements in the frequency, distribution and interannual variability of
17   tropical cyclones (Roberts et al., 2020a, 2020b), particularly in the NH (see further discussion in Section
18, moisture budget (Vannière et al., 2019), and interaction with the ocean (Scoccimarro et al., 2017,
19   single model). At higher resolution the track density of tropical cyclones is increased practically everywhere
20   where tropical cyclones occur. Modelling of some climate extremes is shown to be improved in explosively
21   developing extra-tropical cyclones (Vries et al., 2019; Jiaxiang et al., 2020), blocking (Fabiano et al., 2020;
22   Schiemann et al., 2020; Section and European extreme precipitation due to a better representation of
23   the North Atlantic storm track (Van Haren et al., 2015) and orographic boundary conditions (Schiemann et
24   al., 2018).
26   In CMIP6 a number of Earth system models have increased the realism by which key biogeochemical
27   aspects of the coupled Earth system are represented, affecting for example the carbon and nitrogen cycles,
28   aerosols, and atmospheric chemistry (e.g., Cao et al., 2018; Gettelman et al., 2019; Lin et al., 2019;
29   Mauritsen et al., 2019; Séférian et al., 2019; Sellar et al., 2019; Sidorenko et al., 2019; Swart et al., 2019;
30   Dunne et al., 2020; Seland et al., 2020; Wu et al., 2020; Ziehn et al., 2020). In addition to increased process
31   realism, the level of coupling between the physical climate and biogeochemical components of the Earth
32   system has also been enhanced in some models (Mulcahy et al., 2020) as well as across different
33   biogeochemical components (see Section 5.4 for a discussion and Table 5.4 for an overview). For example,
34   the nitrogen cycle is now simulated in several ESMs (Zaehle et al., 2015; Davies-Barnard et al., 2020). This
35   advance accounts for the fertilization effect nitrogen availability has on vegetation and carbon uptake,
36   reducing uncertainties in the simulations of the carbon uptake responses to physical climate change (Section
37   3.6.1) and to CO2 increases (Arora et al., 2020), thus improving confidence in the simulated airborne fraction
38   of CO2 emissions (Jones and Friedlingstein, 2020) and better constraining remaining carbon budgets (Section
39   5.5). Such advances also allow investigation of land-based climate mitigation options (e.g. through changes
40   in land management and associated terrestrial carbon uptake (Mahowald et al., 2017; Pongratz et al., 2018)
41   or interactions between different facets of the managed Earth system, such as interactions between mitigation
42   efforts targeting climate warming and air quality (West et al., 2013)). A number of developments also
43   explicitly target improved simulation of the past.
45   Further such ESM developments include: (i) Apart from the nitrogen cycle, extending terrestrial carbon
46   cycle models to simulate interactions between the carbon cycle and other nutrient cycles, such as
47   phosphorous, that are known to play an important role in limiting future plant uptake of CO2 (Zaehle et al.,
48   2015). (ii) Introducing explicit coupling between interactive atmospheric chemistry and aerosol schemes
49   (Gettelman et al., 2019; Sellar et al., 2019), which has been shown to affect estimates of historical aerosol
50   radiative forcing (Karset et al., 2018). Furthermore, interactive treatment of atmospheric chemistry in a full
51   ESM supports investigation of interactions between climate and air quality mitigation efforts, such as in
52   AerChemMIP (Collins et al., 2017) as well as interactions between stratospheric ozone recovery and global
53   warming (Morgenstern et al., 2018). (iii) Couplings between components of Earth system models have been
54   extended to increase their utility for studying future interactions across the full Earth system, such as
55   between ocean biogeochemistry and cloud-aerosol processes (Mulcahy et al., 2020), vegetation and impacts
     Do Not Cite, Quote or Distribute                      3-98                                      Total pages: 202
     Final Government Distribution                     Chapter 3                                    IPCC AR6 WGI

 1   on dust production (Kok et al., 2018), production of secondary organic aerosols (SOA, Zhao et al., 2017) and
 2   Equilibrium Climate Sensitivity (ECS) whereby enhanced CO2 fertilization of land vegetation causes
 3   changes in regional surface albedo (Andrews et al., 2019). Increased coupling between physical climate and
 4   biogeochemical processes in a single ESM, along with an increased number of interactively represented
 5   processes, such as permafrost thaw, vegetation, wildfires and continental ice sheets increases our ability to
 6   investigate the potential for abrupt and interactive changes in the Earth system (see Sections 4.7.3 and 5.4.9
 7   and Box 5.1). Table 5.4 provides an overview of recent advances with representing the carbon cycle in
 8   ESMs.
10   In summary, both high-resolution and high-complexity models have been evaluated as part of CMIP6. In
11   comparison with standard resolution CMIP6 models, higher resolution probed under the HighResMIP
12   activity (Haarsma et al., 2016) improves aspects of the simulation of climate (particularly concerning sea
13   surface temperature) but discrepancies remain and there are some regions, such as parts of the Southern
14   Ocean, where currently attainable resolution produces inferior performance (high confidence). Such model
15   behaviour can indicate deficiencies in model physics that are not simply associated with resolution. In
16   several cases, high-complexity ESMs that include additional interactions and thus have potential for
17   additional associated model errors nevertheless perform as well as their low-complexity counterparts,
18   illustrating that interactively simulating these Earth System components as part of the climate system is now
19   well established.

     Do Not Cite, Quote or Distribute                     3-99                                     Total pages: 202
     Final Government Distribution                       Chapter 3                                    IPCC AR6 WGI

 1   Frequently Asked Questions
 3   FAQ 3.1:     How do we Know Humans are Responsible for Climate Change?
 5   The dominant role of humans in driving recent climate change is clear. This conclusion is based on a
 6   synthesis of information from multiple lines of evidence, including direct observations of recent changes in
 7   Earth’s climate; analyses of tree rings, ice cores, and other long-term records documenting how the climate
 8   has changed in the past; and computer simulations based on the fundamental physics that govern the climate
 9   system.
11   Climate is influenced by a range of factors. There are two main natural drivers of variations in climate on
12   timescales of decades to centuries. The first is variations in the sun’s activity, which alter the amount of
13   incoming energy from the sun. The second is large volcanic eruptions, which increase the number of small
14   particles (aerosols) in the upper atmosphere that reflect sunlight and cool the surface—an effect that can last
15   for several years (see also FAQ 3.2). The main human drivers of climate change are increases in the
16   atmospheric concentrations of greenhouse gases and of aerosols from burning fossil fuels, land use and other
17   sources. The greenhouse gases trap infrared radiation near the surface, warming the climate. Aerosols, like
18   those produced naturally by volcanoes, on average cool the climate by increasing the reflection of sunlight.
19   Multiple lines of evidence demonstrate that human drivers are the main cause of recent climate change.
21   The current rates of increase of the concentration of the major greenhouse gases (carbon dioxide, methane
22   and nitrous oxide) are unprecedented over at least the last 800,000 years. Several lines of evidence clearly
23   show that these increases are the results of human activities. The basic physics underlying the warming
24   effect of greenhouse gases on the climate has been understood for more than a century, and our current
25   understanding has been used to develop the latest generation climate models (see FAQ 3.3). Like weather
26   forecasting models, climate models represent the state of the atmosphere on a grid and simulate its evolution
27   over time based on physical principles. They include a representation of the ocean, sea ice and the main
28   processes important in driving climate and climate change.
30   Results consistently show that such climate models can only reproduce the observed warming (black line in
31   FAQ 3.1, Figure 1) when including the effects of human activities (grey band in FAQ 3.1, Figure 1), in
32   particular the increasing concentrations of greenhouse gases. These climate models show a dominant
33   warming effect of greenhouse gas increases (red band, which shows the warming effects of greenhouse gases
34   by themselves), which has been partly offset by the cooling effect of increases in atmospheric aerosols (blue
35   band). By contrast, simulations that include only natural processes, including internal variability related to El
36   Niño and other similar variations, as well as variations in the activity of the sun and emissions from large
37   volcanoes (green band in FAQ 3.1, Figure 1), are not able to reproduce the observed warming. The fact that
38   simulations including only natural processes show much smaller temperature increases indicates that natural
39   processes alone cannot explain the strong rate of warming observed. The observed rates can only be
40   reproduced when human influence is added to the simulations.
42   Moreover, the dominant effect of human activities is apparent not only in the warming of global surface
43   temperature, but also in the pattern of warming in the lower atmosphere and cooling in the stratosphere,
44   warming of the ocean, melting of sea ice, and many other observed changes. An additional line of evidence
45   for the role of humans in driving climate change comes from comparing the rate of warming observed over
46   recent decades with that which occurred prior to human influence on climate. Evidence from tree rings and
47   other paleoclimate records shows that the rate of increase of global surface temperature observed over the
48   past fifty years exceeded that which occurred in any previous 50-year period over the past 2000 years (see
49   FAQ 2.1).
51   Taken together, this evidence shows that humans are the dominant cause of observed global warming over
52   recent decades.
     Do Not Cite, Quote or Distribute                      3-100                                      Total pages: 202
     Final Government Distribution                         Chapter 3                                      IPCC AR6 WGI

 2   FAQ 3.1, Figure 1: Observed warming (1850-2018) is only reproduced in simulations including human influence.
 3                      Global surface temperature changes in observations, compared to climate model simulations of the
 4                      response to all human and natural forcings (grey band), greenhouse gases only (red band), aerosols
 5                      and other human drivers only (blue band) and natural forcings only (green band). Solid coloured
 6                      lines show the multi-model mean, and coloured bands show the 5–95% range of individual
 7                      simulations.
 9   [END FAQ 3.1, FIGURE 1 HERE]

     Do Not Cite, Quote or Distribute                       3-101                                        Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 1   FAQ 3.2:     What is Natural Variability and How has it Influenced Recent Climate Changes?
 3   Natural variability refers to variations in climate that are caused by processes other than human influence.
 4   It includes variability that is internally generated within the climate system and variability that is driven by
 5   natural external factors. Natural variability is a major cause of year-to-year changes in global surface
 6   climate and can play a prominent role in trends over multiple years or even decades. But the influence of
 7   natural variability is typically small when considering trends over periods of multiple decades or longer.
 8   When estimated over the entire historical period (1850–2020), the contribution of natural variability to
 9   global surface warming of -0.23°C–0.23°C is small compared to the warming of about 1.1°C observed
10   during the same period, which has been almost entirely attributed to the human influence.
12   Paleoclimatic records (indirect measurements of climate that can extend back many thousands of years) and
13   climate models all show that global surface temperatures have changed significantly over a wide range of
14   time scales in the past. One of these reasons is natural variability, which refers to variations in climate that
15   are either internally generated within the climate system or externally driven by natural changes. Internal
16   natural variability corresponds to a redistribution of energy within the climate system (for example via
17   atmospheric circulation changes similar to those that drive the daily weather) and is most clearly observed as
18   regional, rather than global, fluctuations in surface temperature. External natural variability can result from
19   changes in the Earth’s orbit, small variations in energy received from the sun, or from major volcanic
20   eruptions. Although large orbital changes are related to global climate changes of the past, they operate on
21   very long time scales (i.e., thousands of years). As such, they have displayed very little change over the past
22   century and have had very little influence on temperature changes observed over that period. On the other
23   hand, volcanic eruptions can strongly cool the Earth, but this effect is short-lived and their influence on
24   surface temperatures typically fades within a decade of the eruption.
26   To understand how much of observed recent climate change has been caused by natural variability (a process
27   referred to as attribution), scientists use climate model simulations. When only natural factors are used to
28   force climate models, the resulting simulations show variations in climate on a wide range of time scales in
29   response to volcanic eruptions, variations in solar activity, and internal natural variability. However, the
30   influence of natural climate variability typically decreases as the time period gets longer, such that it only has
31   mild effects on multi-decadal and longer trends (FAQ 3.2, Figure 1).
33   Consequently, over periods of a couple of decades or less, natural climate variability can dominate the
34   human-induced surface warming trend – leading to periods with stronger or weaker warming, and sometimes
35   even cooling (FAQ 3.2, Figure 1, left and center). Over longer periods, however, the effect of natural
36   variability is relatively small (FAQ 3.2, Figure 1, right). For instance, over the entire historical period (1850–
37   2019), natural variability is estimated to have caused between -0.23°C and +0.23°C of the observed surface
38   warming of about 1.1°C. This means that the bulk of the warming has been almost entirely attributed to
39   human activities, particularly emissions of greenhouse gases (see FAQ 3.1).
41   Another way to picture natural variability and human influence is to think of a person walking a dog. The
42   path of the walker represents the human-induced warming, while their dog represents natural variability.
43   Looking at global surface temperature changes over short periods is akin to focusing on the dog. The dog
44   sometimes moves ahead of the owner and other times behind. This is similar to natural variability that can
45   weaken or amplify warming on the short term. In both cases it is difficult to predict where the dog will be or
46   how the climate will evolve in the near future. However, if we pull back and focus on the slow steady steps
47   of the owner, the path of the dog is much clearer and more predictable, as it follows the path of its owner.
48   Similarly, human influence on the climate is much clearer over longer time periods.
53   FAQ 3.2, Figure 1: Annual (left), decadal (middle) and multi-decadal (right) variations in average global surface
54                      temperature. The thick black line is an estimate of the human contribution to temperature
55                      changes, based on climate models, whereas the green lines show the combined effect of natural

     Do Not Cite, Quote or Distribute                      3-102                                       Total pages: 202
    Final Government Distribution                        Chapter 3                                         IPCC AR6 WGI

1                      variations and human-induced warming, different shadings of green represent different
2                      simulations, which can be viewed as showing a range of potential pasts. The influence of natural
3                      variability is shown by the green bars, and it decreases with longer time scales. The data is sourced
4                      from the CESM1 large ensemble.
6   [END FAQ 3.2, FIGURE 1 HERE]

    Do Not Cite, Quote or Distribute                       3-103                                           Total pages: 202
     Final Government Distribution                      Chapter 3                                   IPCC AR6 WGI

 1   FAQ 3.3:     Are Climate Models Improving?
 3   Yes, climate models have improved and continue to do so, becoming better at capturing complex and small-
 4   scale processes and at simulating present-day mean climate conditions. This improvement can be measured
 5   by comparing climate simulations against historical observations. Both the current and previous generations
 6   of models show that increases in greenhouse gases cause global warming. While past warming is well
 7   simulated by the new generation models as a group, some individual models simulate past warming that is
 8   either below or above what is observed. The information about how well models simulate past warming, as
 9   well as other insights from observations and theory, are used to refine this Report’s projections of global
10   warming.
12   Climate models are important tools for understanding past, present and future climate change. They are
13   sophisticated computer programs that are based on fundamental laws of physics of the atmosphere, ocean,
14   ice, and land. Climate models perform their calculations on a three-dimensional grid made of small bricks or
15   grid cells of about 100 km across. Processes that occur on scales smaller than the model grid cells (such as
16   the transformation of cloud moisture into rain) are treated in a simplified way. This simplification is done
17   differently in different models. Some models include more processes and complexity than others; some
18   represent processes in finer detail (smaller grid cells) than others. Hence the simulated climate and climate
19   change vary between models.
21   Climate modelling started in the 1950s and, over the years, models have become increasingly sophisticated
22   as computing power, observations and our understanding of the climate system have advanced. The models
23   used in the IPCC First Assessment Report published in 1990 correctly reproduced many aspects of climate
24   (FAQ 1.1). The actual evolution of the climate since then has confirmed these early projections, when
25   accounting for the differences between the simulated scenarios and actual emissions. Models continue to
26   improve and get better and better at simulating the large variety of important processes that affect climate.
27   For example, many models now simulate complex interactions between different aspects of the Earth system,
28   such as the uptake of carbon by vegetation on land and by the ocean, and the interaction between clouds and
29   air pollutants. While some models are becoming more comprehensive, others are striving to represent
30   processes at higher resolution, for example to better represent the vortices and swirls in currents responsible
31   for much of the transport of heat in the ocean.
33   Scientists evaluate the performance of climate models by comparing historical model simulations to
34   observations. This evaluation includes comparison of large-scale averages as well as more detailed regional
35   and seasonal variations. There are two important aspects to consider: (1) how models perform individually
36   and (2) how they perform as a group. The average of many models often compares better against
37   observations than any individual model, since errors in representing detailed processes tend to cancel each
38   other out in multi-model averages.
40   As an example, FAQ 3.3 Figure 1 compares simulations from the three most recent generations of models
41   (available around 2005, 2010 and present) with observations of three climate variables. It shows the
42   correlation between simulated and observed patterns, where a value of 1 represents perfect agreement. Many
43   individual models of the new generation perform significantly better, as indicated by values closer to 1. As a
44   group, each generation out-performs the previous generation: the multi-model average (shown by the longer
45   lines) is progressively closer to 1. The vertical extent of the colored bars indicates the range of model
46   performance across each group. The top of the bar moves up with each generation, indicating improved
47   performance of the best performing models from one generation to the next. In the case of precipitation, the
48   performance of the worst performing models is similar in the two most recent model generations, increasing
49   the spread across models.
51   Developments in the latest generation of climate models, including new and better representation of physical,
52   chemical and biological processes, as well as higher resolution, have improved the simulation of many
53   aspects of the Earth system. These simulations, along with the evaluation of the ability of the models to
54   simulate past warming as well as the updated assessment of the temperature response to a doubling of CO2 in
55   the atmosphere, are used to estimate the range of future global warming (FAQ 7.3).
     Do Not Cite, Quote or Distribute                     3-104                                     Total pages: 202
     Final Government Distribution                       Chapter 3                                     IPCC AR6 WGI

 5   FAQ 3.3, Figure 1: Pattern correlations between models and observations of three different variables: surface
 6                      air temperature, precipitation and sea level pressure. Results are shown for the three most
 7                      recent generations of models, from the Coupled Model Intercomparison Project (CMIP): CMIP3
 8                      (orange), CMIP5 (blue) and CMIP6 (purple). Individual model results are shown as short lines,
 9                      along with the corresponding ensemble average (long line). For the correlations the yearly
10                      averages of the models are compared with the reference observations for the period 1980-1999,
11                      with 1 representing perfect similarity between the models and observations. CMIP3 simulations
12                      performed in 2004-2008 were assessed in the IPCC Fourth Assessment, CMIP5 simulations
13                      performed in 2011-2013 were assessed in the IPCC Fifth Assessment, and CMIP6 simulations
14                      performed in 2018-2021 are assessed in this report.
16   [END FAQ 3.3, FIGURE 1 HERE]

     Do Not Cite, Quote or Distribute                      3-105                                       Total pages: 202