It is unequivocal that human influence has warmed the atmosphere, ocean and land since pre-industrial times. Combining the evidence from across the climate system increases the level of confidence in the attribution of observed climate change to human influence and reduces the uncertainties associated with assessments based on single variables. Large-scale indicators of climate change in the atmosphere, ocean, cryosphere and at the land surface show clear responses to human influence consistent with those expected based on model simulations and physical understanding. {3.8.1}

For most large-scale indicators of climate change, the simulated recent mean climate from the latest generation Coupled Model Intercomparison Project Phase 6 (CMIP6) climate models underpinning this assessment has improved compared to the Coupled Model Intercomparison Project Phase 5 (CMIP5) models assessed in AR5 (high confidence). Some differences from observations remain, for example in regional precipitation patterns. High-resolution models exhibit reduced biases in some but not all aspects of surface and ocean climate (medium confidence), and most Earth system models, which include biogeochemical feedbacks, perform as well as their lower-complexity counterparts (medium confidence). The multi-model mean captures most aspects of observed climate change well (high confidence). The multi-model mean captures the proxy-reconstructed global-mean surface air temperature (GSAT) change during past high- and low-CO₂ climates (high confidence) and the correct sign of temperature and precipitation change in most assessed regions in the mid-Holocene (medium confidence). The simulation of paleoclimates on continental scales has improved compared to AR5 (medium confidence), but models often underestimate large temperature and precipitation differences relative to the present day (high confidence). {3.8.2}

The likely range of human-induced warming in global-mean surface air temperature (GSAT) in 2010–2019 relative to1850–1900 is 0.8°C–1.3°C, encompassing the observed warming of 0.9°C–1.2°C, while the change attributable to natural forcings is only −0.1°C to +0.1°C. The best estimate of human-induced warming is 1.07°C. Warming can now be attributed since 1850–1900, instead of since 1951 as done in AR5, thanks to a better understanding of uncertainties and because observed warming is larger. The likely ranges for human-induced GSAT and global mean surface temperature (GMST) warming are equal (medium confidence). Attributing observed warming to specific anthropogenic forcings remains more uncertain. Over the same period, forcing from greenhouse gases¹ likely increased GSAT by 1.0°C–2.0°C, while other anthropogenic forcings including aerosols likely decreased GSAT by 0.0°C–0.8°C. It is very likely that human-induced greenhouse gas increases were the main driver² of tropospheric warming since comprehensive satellite observations started in 1979, and extremely likely that human-induced stratospheric ozone depletion was the main driver of cooling in the lower stratosphere between 1979 and the mid-1990s. {3.3.1}

The CMIP6 model ensemble reproduces the observed historical global surface temperature trend and variability with biases small enough to support detection and attribution of human-induced warming (very high confidence). The CMIP6 historical simulations assessed in this report have an ensemble mean global surface temperature change within 0.2°C of the observations over most of the historical period, and observed warming is within the 5–95% range of the CMIP6 ensemble. However, some CMIP6 models simulate a warming that is either above or below the assessed 5–95% range of observed warming. CMIP6 models broadly reproduce surface temperature variations over the past millennium, including the cooling that follows periods of intense volcanism (medium confidence). For upper air temperature, there is medium confidence that most CMIP5 and CMIP6 models overestimate observed warming in the upper tropical troposphere by at least 0.1°C per decade over the period 1979 to 2014. The latest updates to satellite-derived estimates of stratospheric temperature have resulted in decreased differences between simulated and observed changes of global mean temperature through the depth of the stratosphere (medium confidence). {3.3.1}

The slower rate of GMST increase observed over 1998–2012 compared to 1951–2012 was a temporary event followed by a strong GMST increase (very high confidence). Improved observational datasets since AR5 show a larger GMST trend over 1998–2012 than earlier estimates. All the observed estimates of the 1998–2012 GMST trend lie within the 10th–90th percentile range of CMIP6 simulated trends (high confidence). Internal variability, particularly Pacific Decadal Variability, and variations in solar and volcanic forcings partly offset the anthropogenic surface warming trend over the 1998–2012 period (high confidence). Global ocean heat content continued to increase throughout this period, indicating continuous warming of the entire climate system (very high confidence). Since 2012, GMST has warmed strongly, with the past five years (2016–2020) being the warmest five-year period in the instrumental record since at least 1850 (high confidence). {Cross-Chapter Box 3.1, 3.3.1; 3.5.1}

It is likely that human influence has contributed to³ moistening in the upper troposphere since 1979. Also, there is medium confidence that human influence contributed to a global increase in annual surface specific humidity, and medium confidence that it contributed to a decrease in surface relative humidity over mid-latitude Northern Hemisphere continents during summertime. {3.3.2}

It is likely that human influence has contributed to observed large-scale precipitation changes since the mid-20th century. New attribution studies strengthen previous findings of a detectable increase in Northern Hemisphere mid- to high-latitude land precipitation (high confidence). Human influence has contributed to strengthening the zonal mean precipitation contrast between the wet tropics and dry subtropics (medium confidence). Yet, anthropogenic aerosols contributed to decreasing global land summer monsoon precipitation from the 1950s to 1980s (medium confidence). There is also medium confidence that human influence has contributed to high-latitude increases and mid-latitude decreases in Southern Hemisphere summertime precipitation since 1979 associated with the trend of the Southern Annular Mode toward its positive phase. Despite improvements, models still have deficiencies in simulating precipitation patterns, particularly over the tropical ocean (high confidence). {3.3.2, 3.3.3, 3.5.2}

Human-induced greenhouse gas forcing is the main driver of the observed changes in hot and cold extremes on the global scale (virtually certain) and on most continents (very likely ). It is likely that human influence, in particular due to greenhouse gas forcing, is the main driver of the observed intensification of heavy precipitation in global land regions during recent decades. There is high confidence in the ability of models to capture the large-scale spatial distribution of precipitation extremes over land. The magnitude and frequency of extreme precipitation simulated by CMIP6 models are similar to those simulated by CMIP5 models (high confidence). {Cross-Chapter Box 3.2}

It is likely that human influence has contributed to the poleward expansion of the zonal mean Hadley cell in the Southern Hemisphere since the 1980s. There is medium confidence that the observed poleward expansion of the zonal mean Hadley cell in the Northern Hemisphere is within the range of internal variability. The causes of the observed strengthening of the Pacific Walker circulation since the 1980s are not well understood, and the observed strengthening trend is outside the range of trends simulated in the coupled models (medium confidence). While CMIP6 models capture the general characteristics of the tropospheric large-scale circulation (high confidence), systematic biases exist in the mean frequency of atmospheric blocking events, especially in the Euro-Atlantic sector, some of which reduce with increasing model resolution (medium confidence). {3.3.3}

It is very likely that anthropogenic forcing, mainly due to greenhouse gas increases, was the main driver of Arctic sea ice loss since the late 1970s. There is new evidence that increases in anthropogenic aerosols have offset part of the greenhouse gas-induced Arctic sea ice loss since the 1950s (medium confidence). In the Arctic, despite large differences in the mean sea ice state, loss of sea ice extent and thickness during recent decades is reproduced in all CMIP5 and CMIP6 models (high confidence). By contrast, global climate models do not generally capture the small observed increase in Antarctic sea ice extent during the satellite era, and there is low confidence in attributing the causes of this change. {3.4.1}

It is very likely that human influence contributed to the observed reductions in Northern Hemisphere spring snow cover since 1950. The seasonal cycle in Northern Hemisphere snow cover is better reproduced by CMIP6 than by CMIP5 models (high confidence). Human influence was very likely the main driver of the recent global, near-universal retreat of glaciers. It is very likely that human influence has contributed to the observed surface melting of the Greenland Ice Sheet over the past two decades, and there is medium confidence in an anthropogenic contribution to recent overall mass loss from the Greenland Ice Sheet. However, there is only limited evidence, with medium agreement, of human influence on Antarctic Ice Sheet mass balance through changes in ice discharge. {3.4.2, 3.4.3}

It is extremely likely that human influence was the main driver of the ocean heat content increase observed since the 1970s, which extends into the deeper ocean (very high confidence). Since AR5, there is improved consistency between recent observed estimates and model simulations of changes in upper (<700 m) ocean heat content, when accounting for both natural and anthropogenic forcings. Updated observations and model simulations show that warming extends throughout the entire water column (high confidence), with CMIP6 models simulating 58% of industrial-era heat uptake (1850–2014) in the upper layer (0–700 m), 21% in the intermediate layer (700–2000 m) and 22% in the deep layer (>2000 m). The structure and magnitude of multi-model mean ocean temperature biases have not changed substantially between CMIP5 and CMIP6 (medium confidence). {3.5.1}

It is extremely likely that human influence has contributed to observed near-surface and subsurface ocean salinity changes since the mid-20th century. The associated pattern of change corresponds to fresh regions becoming fresher and salty regions becoming saltier (high confidence). Changes to the coincident atmospheric water cycle and ocean-atmosphere fluxes (evaporation and precipitation) are the primary drivers of the observed basin-scale salinity changes (high confidence). The observed depth-integrated basin-scale salinity changes have been attributed to human influence, with CMIP5 and CMIP6 models able to reproduce these patterns only in simulations that include greenhouse gas increases (medium confidence). The basin-scale changes are consistent across models and intensify through the historical period (high confidence). The structure of the biases in the multi-model mean has not changed substantially between CMIP5 and CMIP6 (medium confidence). {3.5.2}

Combining the attributable contributions from glaciers, ice-sheet surface mass balance and thermal expansion, it is very likely that human influence was the main driver of the observed global mean sea level rise since at least 1971. Since AR5, studies have shown that simulations that exclude anthropogenic greenhouse gases are unable to capture the sea level rise due to thermal expansion (thermosteric) during the historical period and that model simulations that include all forcings (anthropogenic and natural) most closely match observed estimates. It is very likely that human influence was the main driver of the observed global mean thermosteric sea level increase since 1970. {3.5.3, 3.5.1, 3.4.3}

While observations show that the Atlantic Meridional Overturning Circulation (AMOC) has weakened from the mid-2000s to the mid-2010s (high confidence) and the Southern Ocean upper overturning cell has strengthened since the 1990s (low confidence), observational records are too short to determine the relative contributions of internal variability, natural forcing, and anthropogenic forcing to these changes (high confidence). No changes in Antarctic Circumpolar Current transport or meridional position have been observed. The mean zonal and overturning circulations of the Southern Ocean and the mean overturning circulation of the North Atlantic (the Atlantic Meridional Overturning Circulation, AMOC) are broadly reproduced by CMIP5 and CMIP6 models. However, biases are apparent in the modelled circulation strengths (high confidence) and their variability (medium confidence). {3.5.4}

The main driver of the observed increase in the amplitude of the seasonal cycle of atmospheric CO₂ is enhanced fertilization of plant growth by the increasing concentration of atmospheric CO₂ (medium confidence) . However, there is only low confidence that this CO₂ fertilization has also been the main driver of observed greening because land management is the dominating factor in some regions. Earth system models simulate globally averaged land carbon sinks within the range of observation-based estimates (high confidence), but global-scale agreement masks large regional disagreements. {3.6.1}

It is virtually certain that the uptake of anthropogenic CO₂ was the main driver of the observed acidification of the global surface open ocean. The observed increase in CO₂ concentration in the subtropical and equatorial North Atlantic since 2000 is likely associated in part with an increase in ocean temperature, a response that is consistent with the expected weakening of the ocean carbon sink with warming. Consistent with AR5 there is medium confidence that deoxygenation in the upper ocean is due in part to human influence. There is high confidence that Earth system models simulate a realistic time evolution of the global mean ocean carbon sink. {3.6.2}

It is very likely that human influence has contributed to the observed trend towards the positive phase of the Southern Annular Mode (SAM) since the 1970s and to the associated strengthening and southward shift of the Southern Hemispheric extratropical jet in austral summer. The influence of ozone forcing on the SAM trend has been small since the early 2000s compared to earlier decades, contributing to a weaker SAM trend observed over 2000–2019 (medium confidence). Climate models reproduce the summertime SAM trend well, with CMIP6 models outperforming CMIP5 models (medium confidence). By contrast, the cause of the Northern Annular Mode (NAM) trend towards its positive phase since the 1960s and associated northward shifts of the Northern Hemispheric extratropical jet and storm track in boreal winter is not well understood. Models reproduce the observed spatial features and variance of the SAM and NAM very well (high confidence). {3.3.3, 3.7.1, 3.7.2}

Human influence has not affected the principal tropical modes of interannual climate variability or their associated regional teleconnections beyond the range of internal variability (high confidence). Further assessment since AR5 confirms that climate and Earth system models are able to reproduce most aspects of the spatial structure and variance of the El Niño–Southern Oscillation and Indian Ocean Basin and Dipole modes (medium confidence). However, despite a slight improvement in CMIP6, some underlying processes are still poorly represented. In the Tropical Atlantic basin, which contains the Atlantic Zonal and Meridional modes, major biases in modelled mean state and variability remain. {3.7.3 to 3.7.5}

There is medium confidence that anthropogenic and volcanic aerosols contributed to observed changes in the Atlantic Multi-decadal Variability (AMV) index and associated regional teleconnections since the 1960s, but there is low confidence in the magnitude of this influence. There is high confidence that internal variability is the main driver of Pacific Decadal Variability (PDV) observed since pre-industrial times, despite some modelling evidence for potential human influence. Uncertainties remain in quantification of the human influence on AMV and PDV due to brevity of the observational records, limited model performance in reproducing related sea surface temperature (SST) anomalies despite improvements from CMIP5 to CMIP6 (medium confidence), and limited process understanding of their key drivers. {3.7.6, 3.7.7}

Regression-based methods, also known as fingerprinting methods, have been widely used for detection of climate change and attribution of the change to different external drivers. Initially, these methods were applied to detect changes in global surface temperature, and were then extended to other climate variables at different time and spatial scales (e.g., Hegerl et al., 1996; Hasselmann, 1997; Allen and Tett, 1999; Gillett et al., 2003b; Zhang et al., 2007; Min et al., 2008a, 2011). These approaches are based on multivariate linear regression and assume that the observed change consists of a linear combination of externally forced signals plus internal variability, which generally holds for large-scale variables (Hegerl and Zwiers, 2011). The regressors are the expected space–time response patterns to different climate forcings (fingerprints), and the residuals represent internal variability. Fingerprints are usually estimated from climate model simulations following spatial and temporal averaging. A regression coefficient which is significantly greater than zero implies that a detectable change is identified in the observations. When the confidence interval of the regression coefficient includes unity and is inconsistent with zero, the magnitude of the model simulated fingerprints is assessed to be consistent with the observations, implying that the observed changes can be attributed in part to a particular forcing. Variants of linear regression have been used to address uncertainty in the fingerprints due to internal variability (Allen and Stott, 2003) as well as structural model uncertainty (Huntingford et al., 2006).

In order to improve the signal-to-noise ratio, observations and model-simulated responses are usually normalized by an estimate of internal variability derived from climate model simulations. This procedure requires an estimate of the inverse covariance matrix of the internal variability, and some approaches have been proposed for more reliable estimation of this (Ribes et al., 2009). A signal can be spuriously detected due to too-small noise, and hence simulated internal variability needs to be evaluated with care. Model-simulated variability is typically checked through comparing modelled variance from unforced simulations with the observed residual variance using a standard residual consistency test (Allen and Tett, 1999), or an improved one (Ribes and Terray, 2013). Imbers et al. (2014) tested the sensitivity of detection and attribution results to different representations of internal variability associated with short-memory and long-memory processes. Their results supported the robustness of previous detection and attribution statements for the global mean temperature change but they also recommended the use of a wider variety of robustness tests.

Some recent studies focused on the improved estimation of the scaling factor (regression coefficient) and its confidence interval. Hannart et al. (2014) described an inference procedure for scaling factors which avoids making the assumption that model error and internal variability have the same covariance structure. An integrated approach to optimal fingerprinting was further suggested in which all uncertainty sources (i.e., observational error, model error, and internal variability) are treated in one statistical model without a preliminary dimension reduction step (Hannart, 2016). Katzfuss et al. (2017) introduced a similar integrated approach based on a Bayesian model averaging. On the other hand, DelSole et al. (2019) suggested a bootstrap method to better estimate the confidence intervals of scaling factors even in a weak-signal regime. It is notable that some studies do not optimize fingerprints, as uncertainty in the covariance introduces a further layer of complexity, but results in only a limited improvement in detection (Polson and Hegerl, 2017).

Another fingerprinting approach uses pattern similarity between observations and fingerprints, in which the leading empirical orthogonal function obtained from the time-evolving multi-model forced simulation is usually defined as a fingerprint (e.g., Santer et al., 2013; Marvel et al., 2019; Bonfils et al., 2020). Observations and model simulations are then projected onto the fingerprint to measure the degree of spatial pattern similarity with the expected physical response to a given forcing. This projection provides the signal time series, which is in turn tested against internal variability, as estimated from long control simulations. As a way to extend this pattern-based approach to a high-dimensional detection variable at daily time scales, Sippel et al. (2019, 2020) proposed using the relationship pattern with a global climate change metric as a fingerprint. To solve the high-dimensional regression problem which makes regression coefficients not well constrained, they incorporated a statistical learning technique based on a regularized linear regression, which optimizes a global warming signal by giving lower weight to regions with large internal variability.

Considering the difficulty in accounting for climate modelling uncertainties in the regression-based approaches, Ribes et al. (2017) introduced a new statistical inference framework based on an additivity assumption and likelihood maximization, which estimates climate model uncertainty based on an ensemble of opportunity and tests whether observations are inconsistent with internal variability and consistent with the expected response from climate models. The method was further developed by Ribes et al. (2021), who applied it to narrow the uncertainty range in the estimated human-induced warming. Hannart and Naveau (2018), on the other hand, extended the application of standard causal theory (Pearl, 2009) to the context of detection and attribution by converting a time series into an event, and calculating the probability of causation, an approach which maximizes the causal evidence associated with the forcing. On the other hand, Schurer et al. (2018) employed a Bayesian framework to explicitly consider climate modelling uncertainty in the optimal regression method. Application of these approaches to attribution of large-scale temperature changes supports a dominant anthropogenic contribution to the observed global warming.

Climate change signals can vary with time and discriminant analysis has been used to obtain more accurate estimates of time-varying signals, and has been applied to different variables such as seasonal temperatures (Jia and DelSole, 2012) and the South Asian monsoon (Srivastava and DelSole, 2014). The same approach was applied to separate aerosol forcing responses from other forcings (X. Yan et al., 2016) and results using climate model output indicated that detectability of the aerosol response is maximized by using a combination of temperature and precipitation data. Paeth et al. (2017) introduced a detection and attribution method applicable for multiple variables based on a discriminant analysis and a Bayesian classification method. Finally, a systematic approach has been proposed to translating quantitative analysis into a description of confidence in the detection and attribution of a climate response to anthropogenic drivers (Stone and Hansen, 2016).

Overall, these new fingerprinting and other probabilistic methods for detection and attribution as well as efforts to better incorporate the associated uncertainties have addressed a number of shortcomings in previously applied detection and attribution techniques. They further strengthen the confidence in attribution of observed large-scale changes to a combination of external forcings as assessed in the following sections.

Seasonal snow cover is a defining climate feature of the northern continents. It is therefore of considerable interest that climate models correctly simulate this feature. It is discussed in more detail in Section 9.5.3, and observational aspects of snow cover are assessed in Section 2.3.2.2.

The AR5 noted the strong linear correlation between Northern Hemisphere snow cover extent (SCE) and annual-mean surface air temperature in CMIP5 models. It was assessed as likely that there had been an anthropogenic contribution to observed reductions in Northern Hemisphere snow cover since 1970 (Bindoff et al., 2013). The AR5 assessed that CMIP5 models reproduced key features of observed snow cover well, including the seasonal cycle of snow cover over northern regions of Eurasia and North America, but had more difficulties in more southern regions with intermittent snow cover. The AR5 also found that CMIP5 models underestimated the observed reduction in spring snow cover over this period (Figure 3.22; see also Brutel-Vuilmet et al., 2013; Thackeray et al., 2016; Santolaria-Otín and Zolina, 2020). This behaviour has been linked to how the snow-albedo feedback is represented in models (Thackeray et al., 2018a). The CMIP5 multi-model ensemble has been shown to represent the snow-albedo feedback more realistically than CMIP3, although models from some individual modelling centres have not improved or have even got worse (Thackeray et al., 2018a). There is still a systematic overestimation of the albedo of boreal forest covered by snow (Thackeray et al., 2015; Y. Li et al., 2016). Consequently, the snow albedo feedback might have been overestimated by CMIP5 models (Section 9.5.3; Xiao et al., 2017).

Figure 3.22 | Time series of Northern Hemisphere March–April mean snow cover extent (SCE) from observations, CMIP5 and CMIP6 simulations. The observations (grey lines) are updated Brown-NOAA (Brown and Robinson, 2011), Mudryk et al. (2020), and GLDAS2. CMIP5 (top) and CMIP6 (bottom) simulations of the response to natural plus anthropogenic forcing are shown in brown, natural forcing only in green, and the pre-industrial control simulation range is presented in blue. Five-year mean anomalies are shown for the 1923–2017 period with the x-axis representing the centre years of each five-year mean. CMIP5 all forcing simulations are extended by using RCP4.5 scenario simulations after 2005 while CMIP6 all forcing simulations are extended by using SSP2-4.5 scenario simulations after 2014. Shading indicates 5th–95th percentile ranges for CMIP5 and CMIP6 all and natural forcings simulations, and solid lines are ensemble means, based on all available ensemble members with equal weight given to each model (Section 3.2). The blue vertical bar indicates the mean 5th–95th percentile range of pre-industrial control simulation anomalies, based on non-overlapping segments. The numbers in brackets indicate the number of models used. Anomalies are relative to the average over 1971–2000. For models, SCE is restricted to grid cells with land fraction ≥50%. Greenland is excluded from the total area summation. Figure is modified from Paik and Min (2020), their Figure 1. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

CMIP6 models improve on CMIP5 models in producing slightly increased SCE versus CMIP5, correcting the low bias in CMIP5 (Mudryk et al., 2020). The linear relationship noted above between GSAT and SCE also exists in CMIP6 (Mudryk et al., 2020). Like CMIP5, the CMIP6 models capture the negative trend in spring snow cover that has occurred in recent decades (Figure 3.22). However, the median CMIP6 model now produces slightly stronger post-1981 declines in the March to April mean SCE than the CMIP5 median (Mudryk et al., 2020). Until about 1980, the models produce a generally stable March to April SCE, but after that a substantial decline, reaching a loss of about 2 × 10⁶km² in 2012–2017 relative to the 1971–2000 average. Compared to earlier studies which found that models underestimate observed trends for the 1979–2005 period (Brutel-Vuilmet et al., 2013), both CMIP5 and CMIP6 models show improved agreement with the observations over the period to 2017 (Figure 3.22). One remaining concern is a failure of most CMIP6 models to correctly represent the relationship between snow cover extent and snow mass, reflecting too slow seasonal increases and decreases of SCE in the models (Mudryk et al., 2020).

Several CMIP5 and CMIP6 based studies have consistently attributed the observed Northern Hemisphere spring SCE changes (Section 2.3.2.2) to anthropogenic influences (Rupp et al., 2013; Najafi et al., 2016; Paik and Min, 2020), with the observed changes being found to be inconsistent with natural variability alone. Similarly, spring snow thickness (Snow Water Equivalent) changes on the scale of the Northern Hemisphere have been attributed to greenhouse gas forcing (Jeong et al., 2017). Using individual forcing simulations from multiple CMIP6 models, Paik and Min (2020) detected greenhouse gas influence in the observed decrease of early spring SCE between 1925 and 2019, which was found to be separable from the responses to other forcings.

In summary, it is very likely that anthropogenic influence contributed to the observed reductions in Northern Hemisphere springtime snow cover since 1950. CMIP6 models better represent the seasonality and geographical distribution of snow cover than CMIP5 simulations (high confidence). Both CMIP5 and CMIP6 models simulate strong declines in spring SCE during recent years, in general agreement with observations, causing the multi-model mean decreasing trend in spring SCE to now better agree with observations than in earlier evaluations. Evidence has yet to emerge that interactions between vegetation and snow, found problematic in CMIP5, have improved in CMIP6 models (Section 9.5.3). Such deficiencies in the representation of snow in climate models mean there is medium confidence in the simulation of snow cover over the northern continents in CMIP6 model simulations. The models consistently link snow extent to surface air temperature (Figure 9.24). With warming of near-surface air linked to anthropogenic influence, and particularly to greenhouse gas increases (Section 3.3.1.1), this provides additional evidence that reductions in snow cover are also caused by human activity.

While (Chapter 9 (Sections 9.4 and 9.5) discusses process understanding for glaciers and ice sheets, as well as evaluation of global and regional-scale glacier and ice-sheet models, our focus here is on the attribution of large-scale changes in glaciers and ice sheets. Land ice in the form of glaciers has been included in CMIP climate and Earth system models as components of the land surface models for many years. However, their representation is simplified and is omitted altogether in the less complex modelling systems. In CMIP3 (Meehl et al., 2007) and CMIP5 (Taylor et al., 2012) land ice area fraction, a component of land surface models, was defined as a time-independent quantity, and in most model configurations was preset at the simulation initialization as a permanent land feature. In CMIP6 considerable progress has been made in improving and evaluating the representation of modelled land ice. For glaciers, an example is the expansion of the Joint UK Land Environment Simulator (JULES) land surface model to enable elevated tiles, and hence more accurately simulate the altitudinal atmospheric effects on glaciers (Shannon et al., 2019). Moreover, standalone glacier models have now been systematically compared in GlacierMIP (Hock et al., 2019a; Marzeion et al., 2020). The Antarctic and Greenland Ice Sheets were absent in global climate models that pre-date CMIP6 (Eyring et al., 2016a), however some preliminary analyses that used results from CMIP5 to drive standalone ice-sheet models were included in AR5 (Church et al., 2013a). For the first time in CMIP, the latest CMIP6 phase includes a coordinated effort to simulate temporally evolving ice sheets within the Ice Sheet Model Intercomparison Project (ISMIP6; Box 9.3; Nowicki et al., 2016). Our understanding of aspects of the global water storage contained in glaciers and ice sheets, and their contribution to sea-level rise, has improved since AR5 and SROCC (Hock et al., 2019b; Meredith et al., 2019) both in models and observations (see assessment of observations and model evaluation for the Greenland Ice Sheet in Sections 2.3.2.4.1 and 9.4.1; Antarctica in Sections 2.3.2.4.2 and 9.4.2; and glaciers in Sections 2.3.2.3 and 9.5.1).

Ocean temperature and ocean heat content are key physical variables considered for climate model evaluation and are primary indicators of a changing ocean climate. This section assesses the performance of climate models in representing the mean state ocean temperature and heat content (Section 3.5.1.1), with a particular focus on the tropical oceans given the importance of air-sea coupling in these areas (Section 3.5.1.2). This is followed by an assessment of detection and attribution studies of changes in ocean temperature and heat content (Section 3.5.1.3). Changes in global surface temperature are assessed in Section 3.3.1.1.

3.5.1.1 Sea Surface and Zonal Mean Ocean Temperature Evaluation

In CMIP3 and CMIP5 models, large SST biases were found in the mid- and high latitudes (Flato et al., 2013). In CMIP6, the Northern Hemisphere mid-latitude surface temperature biases appear to be marginally improved in the multi-model mean when contrasted to CMIP5 despite large biases remaining in a few models (Figures 3.23a and 3.24). There is a decreased spread of the zonal mean SST error between 50°N and 30°S, relative to CMIP5 (Figure 3.24a). On the other hand, the Southern Ocean’s warm surface temperature bias remains (Figure 3.23a; Beadling et al., 2020), and is on average larger in CMIP6 than in CMIP5 models (Figures 3.23a and 3.24). This warm bias is often associated with persistent overlying atmospheric cloud biases (Hyder et al., 2018). Several other large biases also appear to remain largely unresolved in CMIP6, particularly warm biases in excess of 1°C along the equatorial eastern continental boundaries of the tropical Atlantic and Pacific Oceans (Figure 3.23a).

Figure 3.23 | Multi-model mean bias of (a) sea surface temperature and (b) near-surface salinity, defined as the difference between the CMIP6 multi-model mean and the climatology from the World Ocean Atlas 2018. The CMIP6 multi-model mean is constructed with one realization of 46 CMIP6 historical experiments for the period 1995–2014 and the climatology from the World Ocean Atlas 2018 is an average over all available years (1955–2017). Uncertainty is represented using the advanced approach: No overlay indicates regions with robust signal, where ≥66% of models show change greater than the variability threshold and ≥80% of all models agree on sign of change; diagonal lines indicate regions with no change or no robust signal, where <66% of models show a change greater than the variability threshold; crossed lines indicate regions with conflicting signal, where ≥66% of models show change greater than the variability threshold and <80% of all models agree on sign of change. For more information on the advanced approach, please refer to Cross-Chapter Box Atlas.1. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Figure 3.24 | Biases in zonal mean and equatorial sea surface temperature (SST) in CMIP5 and CMIP6 models. CMIP6 (red), CMIP5 (blue) and HighResMIP (green) multi-model mean (a) zonally averaged SST bias; (b) equatorial SST bias; and (c) equatorial SST compared to observed mean SST (black line) for 1979–1999. The inter-model 5th and 95th percentiles are depicted by the respective shaded range. Model climatologies are derived from the 1979–1999 mean of the historical simulations, using one simulation per model. The Hadley Centre Sea Ice and Sea Surface Temperature version 1 (HadISST) (Rayner et al., 2003) observational climatology for 1979–1999 is used as the reference for the error calculation in (a) and (b); and for observations in (c). Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Overall, the simulated and observed trends in SST patterns are generally consistent for the historical period (Olonscheck et al., 2020). The CMIP6 models generally represent the observed pattern of trends better than the CMIP5 models, and observed trends fall within the range of simulated trends over a larger area for CMIP6 models than for CMIP5 models (Olonscheck et al., 2020).

The CMIP5 multi-model mean zonally averaged subsurface ocean temperature showed warm biases between 200 m and 2000 m (mid-depth) over most latitudes, with exceptions in the Southern Ocean (>60°S, 100–2000 m) and upper (0–400 m) Arctic Ocean. Cold biases were simulated near the surface (0–200 m) at most latitudes (Flato et al., 2013). CMIP6 biases are broadly consistent with those reported in CMIP5 for the near-surface (<200 m) and mid-depth (200–2000 m) ocean (Voldoire et al., 2019b; Beadling et al., 2020; Y. Zhu et al., 2020). The warm bias begins between 100 and 400 m depth in all three basins, however, it is most prominent in the Atlantic Ocean, with a maximum magnitude in the equatorial latitudes, as in CMIP5 (Figure 3.25). In the Pacific, the large warm biases are mostly seen in the subtropical regions (30°N–60°N and 30°S–60°S). The cool near surface tropical bias is most prominent in the Pacific Ocean and also present in the Atlantic, with a smaller magnitude (Figure 3.25). Relative to CMIP5, the most prominent difference is an increase to the mid-depth (300–2000 m) warm bias in CMIP6 and a change in sign of the bias from cold to warm for the Southern Ocean mid-depth (>60°S) from CMIP5 to CMIP6 (Figure 3.25). Compared to CMIP3 and CMIP5, there is improved agreement between most CMIP6 models and observations in their representation of the zonal mean temperature of the upper 100 m of the Southern Ocean (Beadling et al., 2020).

Figure 3.25 | CMIP6 potential temperature and salinity biases for the global ocean, Atlantic Ocean, Pacific Ocean and Indian Ocean. Shown in colour are the time-mean differences between the CMIP6 historical multi-model climatological mean and observations, zonally averaged for each basin (excluding marginal and regional seas). The observed climatological values are obtained from the World Ocean (Atlas 2018 (WOA18, 1981–2010; Prepared by the Ocean Climate Laboratory, National Oceanographic Data Center, Silver Spring, MD, USA), and are shown as labelled black contours for each of the basins. The simulated annual mean climatologies for 1981 to 2010 are calculated from available CMIP6 historical simulations, and the WOA18 climatology utilized synthesized observed data from 1981 to 2010. Output from a total of 30 available CMIP6 models is used for the temperature panels (left column) and 28 models for the salinity panels (right column). Potential temperature units are °C and salinity units are the Practical Salinity Scale 1978 [PSS-78]. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Focusing on the deep ocean (>2000 m), the CMIP6 ensemble mean shows a prominent and consistent warm bias (Figure 3.25), in all basins except the equatorial and northern Pacific, which contrasts to a cold bias seen in CMIP5 (Flato et al., 2013). We note that while an updated observational temperature dataset is used in this assessment (WOA09 was used in AR5, while WOA18, 1981–2010 is used in AR6), the deep-ocean warm bias remains and is approaching double the magnitude (about 0.5°C) of the equivalent CMIP5 multi-model mean bias, a feature which is particularly prominent in the Atlantic and southern Indian Oceans. Increased horizontal resolution as well as the choice of the vertical coordinate are reported to partly improve these biases in some models (Adcroft et al., 2019; Rackow et al., 2019; Hewitt et al., 2020).

Since AR5, there has been growing evidence that the representation of mean surface and deeper ocean temperatures in coupled climate models can be improved by increasing the horizontal resolution both in the ocean and the atmosphere (e.g., Small et al., 2014; Hewitt et al., 2016; Iovino et al., 2016; Roberts et al., 2019). At an ocean resolution of around 1°, which is typical of CMIP6 models, some processes are parameterized rather than explicitly resolved, leading to a compromise in their dynamical representation. An increase in the model resolution allows for processes to be explicitly resolved, and can for example, enhance the simulation of eddies, thus improving simulated vertical eddy transport, and reducing temperature drifts in the deeper ocean (Griffies et al., 2015; von Storch et al., 2016). For some models, the mean absolute error in ocean temperature below 500 m is smaller in the high resolution version compared to the low resolution version, particularly in eddy-active regions such as the North Atlantic (Rackow et al., 2019). Increasing the horizontal resolution of individual climate models often leads to an overall decrease in the surface temperature biases over regions where they persisted through earlier CMIP generations, such as the central and western equatorial Pacific, as well as the North and tropical Atlantic (Figure 3.3e; Roberts et al., 2019; Hewitt et al., 2020). Despite this, as a group the four HighResMIP models included in Figures 3.3e and 3.24 do not on average show smaller SST biases than the CMIP6 multi-model mean, demonstrating the importance of factors other than resolution in contributing to SST biases.

In summary, there is little improvement in the multi-model mean sea surface and zonal mean ocean temperatures from CMIP5 to CMIP6 (medium confidence). Nevertheless, the CMIP6 models show a somewhat more realistic pattern of SST trends (low confidence).

3.5.1.3 Ocean Heat Content Change Attribution

The ocean plays an important role as the Earth’s primary energy store. The AR5 and SROCC assessed that the ocean accounted for more than 90% of the Earth’s energy change since the 1970s (Rhein et al., 2013; Bindoff et al., 2019). These assessments are consistent with recent studies assessed in Section 7.2 and Cross-Chapter Box 9.1, which find that 91% of the observed change in Earth’s total energy from 1971 to 2018 was stored in the ocean (von Schuckmann et al., 2020). The AR5 concluded that anthropogenic forcing has very likely made a substantial contribution to ocean warming above 700 m, whereas below 700 m, limited measurements restricted the assessment of ocean heat content changes in AR5 and prevented a robust comparison between observations and models (Bindoff et al., 2013).

With the recent increase in ocean sampling by Argo to 2000 m (Roemmich et al., 2015; Riser et al., 2016; von Schuckmann et al., 2016) and the resulting improvements in estimates of ocean heat content (Abraham et al., 2013; Balmaseda et al., 2013; Durack et al., 2014b; Cheng et al., 2017; von Schuckmann et al., 2020), a more quantitative assessment of the global ocean heat content changes that extends into the intermediate ocean (700–2000 m) over the more recent period (from 2005 to the present; Durack et al., 2018) can be performed. Observed ocean heat content changes are discussed in Section 2.3.3.1, where it is reported that it is virtually certain that the global upper ocean (0–700 m) and very likely that the global intermediate ocean (700–2000 m) warmed substantially from 1971 to the present. Further, ocean layer warming contributions are reported as 61% (0–700 m), 31% (700–2000 m) and 8% (>2000 m) for the 1971 to 2018 period (Table 2.7). CMIP5 model simulations replicate this partitioning fairly well for the industrial-era (1865 to 2017) throughout the upper (0–700 m, 65%), intermediate (700–2000 m, 20%) and deep (>2000 m, 15%) layers (Gleckler et al., 2016; Durack et al., 2018). The corresponding warming percentages for the multi-model mean of a subset of CMIP6 simulations over the 1850–2014 period are 58% for the upper, 21% for the intermediate, and 22% for the deep-ocean layers (Figure 3.26). These results are consistent with SROCC which assessed that it is virtually certain that both the upper and intermediate ocean warmed from 2004 to 2016, with an increased rate of warming since 1993 (Bindoff et al., 2019). The spatial distribution of these changes for different ocean depths is assessed in Section 9.2.2.1.

Figure 3.26 | Global ocean heat content in CMIP6 simulations and observations. Time series of observed (black) and simulated (red) global ocean heat content anomalies with respect to 1995–2014 for the full ocean depth (left-hand panel); upper layer: 0–700 m (top right-hand panel); intermediate layer: 700–2000 m (middle right-hand panel); and the abyssal ocean: >2000 m (bottom right-hand panel). The best estimate observations (black solid line) for the period of 1971–2018, along with very likely ranges (black shading) are from Section 2.3.3.1. For the models (1860–2014), ensemble members from 15 CMIP6 models are used to calculate the multi-model mean values (red solid line) after averaging across simulations for each independent model. The very likely ranges in the simulations are shown in red shading. Simulation drift has been removed from all CMIP6 historical runs using a contemporaneous portion of the linear fit to each corresponding pre-industrial control run (Gleckler et al., 2012). Units are zettajoules (ZJ; 10²¹ joule). Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

The multi-model means of both CMIP5 and CMIP6 historical simulations forced with time varying natural and anthropogenic forcing shows robust increases in ocean heat content in the upper (0–700 m) and intermediate (700–2000 m) ocean (high confidence) (Figure 3.26; Cheng et al., 2016, 2019; Gleckler et al., 2016; Bilbao et al., 2019; Tokarska et al., 2019). Temporary (<10 years) surface and subsurface cooling during and after large volcanic eruptions is also captured in the upper-ocean, and global mean ocean heat content (Balmaseda et al., 2013). The ocean heat content increase is also reflected in the corresponding ocean thermal expansion which is a leading contributor to global mean sea level rise (Sections 3.5.3.2 and 9.2.4, and Box 9.1).

For the period 1971–2014, the rate of ocean heat uptake for the global ocean in the CMIP6 models is about 6.43 [2.08–8.66] ZJ yr^–1, with the upper, intermediate and deeper layers respectively accounting for 68%, 16% and 16% of the full depth global heat uptake (Figure 3.26). Overall, the simulated ocean heat content changes are consistent with the updated and improved observational analyses, within the very likely uncertainty range defined for each (see also (Section 2.3.3.1, Table 2.7; Domingues et al., 2008; Purkey and Johnson, 2010; Levitus et al., 2012; Good et al., 2013; Cheng et al., 2017; Ishii et al., 2017; Zanna et al., 2019) as well as with the ocean components of total Earth heating assessed in Section 7.2.2.2, Table 7.1. Nevertheless, large uncertainties remain, particularly in the deeper layers due to the poor temporal and spatial sampling coverage, particularly in the Atlantic, Southern and Indian Oceans (Garry et al., 2019). The very likely ranges of the simulated trends for the full ocean depth and below 2000 m fall within the very likely range of observed uptake during the last two decades. In the intermediate layer, the multi-model ensemble mean mostly stays above the observed 5th–95th percentile range before the year 2000, and below that range after 2000. For the upper ocean, some individual model realizations show a reduced ocean heat content increase during the 1970s and 1980s, which is then compensated by a greater warming than the observations from the early 1990s. These discrepancies have been linked with a temporary increase in the Southern Ocean deep water formation rate, as well as with the models’ strong aerosol cooling effects and high equilibrium climate sensitivity (see also Section 7.5.6 and Box 7.2; Andrews et al., 2019, 2020; Golaz et al., 2019; Dunne et al., 2020; Winton et al., 2020).

Nevertheless, simulations show that the rate of ocean heat uptake has doubled in the past few decades, when contrasted to the rate over the complete 20th century (Figure 3.26), with over a third of the accumulated heat stored below 700 m (Cheng et al., 2016, 2019; Gleckler et al., 2016; Durack et al., 2018). The Southern Ocean shows the strongest ocean heat uptake that penetrates to deeper layers (Section 9.2.3.2), whereas ocean heat content increases in the Pacific and Indian Oceans largely occur in the upper layers (Bilbao et al., 2019).

Since AR5, the attribution of ocean heat content increases to anthropogenic forcing has been further supported by more detection and attribution studies. These studies have shown that contributions from natural forcing alone cannot explain the observed changes in ocean heat content in either the upper or intermediate ocean layers, and a response to anthropogenic forcing is clearly detectable in ocean heat content (Gleckler et al., 2016; Bilbao et al., 2019; Tokarska et al., 2019). Moreover, a response to greenhouse gas forcing is detectable independently of the response to other anthropogenic forcings (Bilbao et al., 2019; Tokarska et al., 2019), which has offset part of the greenhouse gas induced warming. Further evidence is provided by the agreement between observed and simulated changes in global thermal expansion associated with the ocean heat content increase when both natural and anthropogenic forcings are included in the simulations (Section 3.5.3.2), though internal variability plays a larger role in driving basin-scale thermosteric sea level trends (Bilbao et al., 2015). Over the Southern Ocean, warming is detectable over the late 20th century and is largely attributable to greenhouse gases (Swart et al., 2018; Hobbs et al., 2021), while other anthropogenic forcings such as ozone depletion have been shown to mitigate the warming in some of the CMIP5 simulations (Swart et al., 2018; Hobbs et al., 2021). The use of the mean temperature above a fixed isotherm rather than fixed depth further strengthens a robust detection of the anthropogenic response in the upper ocean (Weller et al., 2016), and better accounting for internal variability in the upper ocean (Rathore et al., 2020), helps explain reported hemispheric asymmetry in ocean heat content change (Durack et al., 2014b).

In summary, there is strong evidence for an improved understanding of the observed global ocean heat content increase. It is extremely likely that human influence was the main driver of the ocean heat content increase observed since the 1970s, which extends into the deeper ocean (very high confidence). Updated observations, like model simulations, show that warming extends throughout the entire water column (high confidence).

While ocean assessments have primarily focused on temperature changes, improved observational salinity products since the early 2000s have supported more assessment of long-term ocean salinity change and variability from AR4 (Bindoff et al., 2007) to AR5 across both models and observations (Flato et al., 2013; Rhein et al., 2013). The AR5 assessed that it was very likely that anthropogenic forcings have made a discernible contribution to surface and subsurface ocean salinity changes since the 1960s. The SROCC augmented these insights, noting that observed high latitude freshening and warming have very likely made the surface ocean less dense with a stratification increase of between 2.18% and 2.42% from 1970 to 2017 (Bindoff et al., 2019). A recent observational analysis has expanded on these assessments, suggesting a very marked summertime density contrast enhancement across the mixed layer base of 6.2–11.6% per decade, driven by changes in temperature and salinity, which is more than six times larger than previous estimates (Sallée et al., 2021). An idealized ocean modelling study suggests that the enhanced stratification can account for a third of the salinity enhancement signal since 1990 (Zika et al., 2018). Thus, there has been an expansion of observed global- and basin-scale salinity change assessment literature since AR5, with many new studies reproducing the key patterns of long-term salinity change reported in AR5 (Rhein et al., 2013), and linking these through modelling studies to coincident changes in evaporation–precipitation patterns at the ocean surface (Sections 2.3.1.3, 3.3.2, 8.2.2.1 and 9.2.2).

Unlike SSTs, simulated sea surface salinity (SSS) does not provide a direct feedback to the atmosphere. However, some recent work has identified indirect radiative feedbacks through sea-salt aerosol interactions (Ayash et al., 2008; Amiri-Farahani et al., 2019; Z. Wang et al., 2019) that can act to strengthen tropical cyclones, and increase precipitation (Balaguru et al., 2012, 2016; Grodsky et al., 2012; Reul et al., 2014; Jiang et al., 2019). The absence of a direct feedback is one of the primary reasons why salinity simulation is difficult to constrain in ocean modelling systems, and why deviations from the observed near-surface salinity mean state between models and observations are often apparent (Durack et al., 2012; Shi et al., 2017).

3.5.2.2 Salinity Change Attribution

AR5 concluded that it was very likely that anthropogenic forcings had made a discernible contribution to surface and subsurface ocean salinity changes since the 1960s (Bindoff et al., 2013; Rhein et al., 2013). It highlighted that the spatial patterns of salinity trends, and the mean fields of salinity and evaporation minus precipitation are all similar, with an enhancement to Atlantic Ocean salinity and freshening in the Pacific and Southern Oceans. Since AR5 all subsequent work on assessing observed and modelled salinity changes has confirmed these results.

Considerable changes to observed broad- or basin-scale ocean near-surface salinity fields have been reported (see Section 2.3.3.2), and these have been linked to changes in the evaporation minus precipitation patterns at the ocean surface through model simulations, typically expressing a pattern of change where climatological mean fresh regions become fresher and corresponding salty regions becoming saltier (Durack et al., 2012, 2013; Zika et al., 2015; Lago et al., 2016; Skliris et al., 2016, 2018; Cheng et al., 2020), also broadly present in the CMIP6 multi-model mean (Figure 3.27). At basin-scales, the depth-integrated effect of mean salinity changes as captured in halosteric sea level for the top 0 to 2000 m has also been assessed based on observational products, and these results mirror near-surface patterns in the CMIP5 and CMIP6 models, with most areas that are becoming fresher at the surface exhibiting increases in halosteric sea level, and areas becoming saltier exhibiting decreases (Durack et al., 2014a; Figure 3.28). Further investigations using observations and models together have tied the long-term patterns of surface and subsurface salinity changes to coincident changes to the evaporation minus precipitation field over the ocean (Durack et al., 2012, 2013; Durack, 2015; Levang and Schmitt, 2015; Zika et al., 2015, 2018; Grist et al., 2016; Lago et al., 2016; Cheng et al., 2020), however the rate of these changes through time continues to be an active area of active research (Skliris et al., 2014; Zika et al., 2015, 2018; Cheng et al., 2020; Sallée et al., 2021).

Figure 3.27 | Maps of multi-decadal salinity trends for the near-surface ocean. Units are Practical Salinity Scale 1978 [PSS-78] per decade. (Top) The best estimate (Section 2.3.3.2) observed trend (1950–2019, Durack and Wijffels, 2010). (Bottom) Simulated trend from the CMIP6 historical experiment multi-model mean (1950–2014). Black contours show the climatological mean salinity in increments of 0.5 PSS-78 (thick lines 1 PSS-78). Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Figure 3.28 | Long-term trends in halosteric and thermosteric sea level in CMIP6 models and observations. Units are mm yr^–1. In The right-hand column, three observed maps of 0 to 2000 m halosteric sea level trends are shown: top (D&W) from Durack and Wijffels (2010), 1950–2019, updated; upper-middle (EN4) from Good et al. (2013), 1950–2019, updated; and lower-middle (Ishii) from Ishii et al. (2017), 1955–2019, updated. Bottom-right: the CMIP6 historical multi-model mean (1950–2014). Red and orange colours show a halosteric contraction (enhanced salinity) and blue and green a halosteric expansion (reduced salinity). In The left-hand column , basin-integrated halosteric (top) and thermosteric (bottom) trends for the Atlantic and Pacific, the two largest ocean basins, are shown, where Pacific anomalies are presented on the x-axis and Atlantic on the y-axis. Observational estimates are presented in black, CMIP6 historical (all forcings) simulations are shown in orange squares, with the multi-model mean shown as a dark orange diamond with a black bounding box. CMIP6 hist-nat (historical natural forcings only) simulations are shown in green squares with the multi-model mean as a dark green diamond with a black bounding box. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Climate change detection and attribution studies have considered salinity, with the first of these assessed in AR5 (Bindoff et al., 2013). Since that time, the positive detection conclusions (Stott et al., 2008; Pierce et al., 2012; Terray et al., 2012) have been supported by a number of more recent and independent assessments which have reproduced the multi-decadal basin-scale patterns of change in observations and models (Figures 3.27 and 3.28; Durack et al., 2014a; Durack, 2015; Levang and Schmitt, 2015; Skliris et al., 2016). Observed depth-integrated basin responses, contrasting the Pacific and Atlantic basins (freshening Pacific and enhanced salinity Atlantic) were also shown to be replicated in most historical (natural and anthropogenically forced) simulations, with this basin contrast absent in CMIP5 and CMIP6 natural-only simulations that exclude anthropogenic forcing (Durack et al., 2014a; Figure 3.28).

While observational sparsity considerably limits quantification of regional changes, a recent study by Friedman et al. (2017) assessed salinity changes in the Atlantic Ocean from 1896 to 2013 and confirmed the pattern of mid-to-low latitude enhanced salinity and high latitude North Atlantic freshening over this period exists even after accounting for the effects of the NAO and AMO.

Considering the bulk of evidence, it is extremely likely that human influence has contributed to observed near-surface and subsurface salinity changes across the globe since the mid-20th century. All available multi-decadal assessments have confirmed that the associated pattern of change corresponds to fresh regions becoming fresher and salty regions becoming saltier (high confidence). CMIP5 and CMIP6 models are only able to reproduce these patterns in simulations that include greenhouse gas increases (medium confidence). Changes to the coincident atmospheric water cycle and ocean-atmosphere fluxes (evaporation and precipitation) are the primary drivers of the basin-scale observed salinity changes (high confidence). This result is supported by all available observational assessments, along with a growing number of climate modelling studies targeted at assessing ocean and water cycle changes. The basin-scale changes are consistent across models and intensify on centennial scales from the historical period through to the projections of future climate (high confidence).

In keeping with the scope of this chapter, this section addresses global and basin-scale sea level changes, whereas regional and local sea level changes are assessed in Section 9.6. In AR5, the observed sea level budget was closed by considering all contributing factors including ocean warming, mass contributions from terrestrial storage, glaciers, and the Antarctic and Greenland ice sheets (Church et al., 2013b). The SROCC found that the observed global mean sea level (GMSL) rise is consistent within uncertainties with the sum of the estimated observed contributions for 1993–2015 and 2006–2015.

3.5.3.2 Sea Level Change Attribution

The SROCC concluded with high confidence that the dominant cause of GMSL rise since 1970 is anthropogenic forcing. Prior to that, AR5 had concluded that it is very likely that there has been a substantial contribution from anthropogenic forcings to GMSL rise since the 1970s. Since AR5, several studies have identified a human contribution to observed sea level change resulting from a warming climate as manifest in thermosteric sea level change and the contribution from melting glaciers and ice sheets.

For the global mean thermosteric sea level change, Slangen et al. (2014) showed the importance of anthropogenic forcings (combined greenhouse gas and aerosol forcings) for explaining the magnitude of the observed changes between 1957 and 2005, considering the full depth of the ocean and natural forcings in order to capture the variability (see also Figure 3.29). Over the 1950–2005 period, Marcos and Amores (2014) found that human influence explains 87% of the 0–700 m global thermosteric sea level rise. Both thermosteric and regional dynamic patterns of sea level change in individual forcing experiments from CMIP5 were considered by Slangen et al. (2015) who showed that responses to anthropogenic forcings are significantly different from both internal variability and inter-model differences and that although greenhouse gas and anthropogenic aerosol forcings produce opposite GMSL responses, there are differences in the response on regional scales. Based on these studies, we conclude that it is very likely that anthropogenic forcing was the main driver of the observed global mean thermosteric sea level change since 1970.

Figure 3.29 | Simulated and observed global mean sea level change due to thermal expansion for CMIP6 models and observations relative to the baseline period 1850–1900. Historical simulations are shown in brown, natural only in green, greenhouse gas only in grey, and aerosol only in blue (multi-model means shown as thick lines, and shaded ranges between the 5th and 95th percentile). The best estimate observations (black solid line) for the period of 1971–2018, along with very likely ranges (black shading) are from Section 2.3.3.1 and are shifted to match the multi-model mean of the historical simulations for the 1995–2014 period. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

In an attribution study of the sea-level contributions of glaciers, Marzeion et al. (2014) found that between 1991 and 2010, the anthropogenic fraction of global glacier mass loss was 69 ± 24% (see also Section 3.4.3.1). Slangen et al. (2016) considered all quantifiable components of the GMSL budget and showed that anthropogenically forced changes account for 69 ± 31% of the observed sea level rise over the period 1970 to 2005, whereas natural forcings combined with internal variability have a much smaller effect – only contributing 9 ± 18% of the change over the same period. These studies indicate that about 70% of the combined change in glaciers, ice-sheet surface mass balance and thermal expansion since 1970 can be attributed to anthropogenic forcing, and that this percentage has increased over the course of the 20th century. Detection studies on GMSL change in the 20th century (Becker et al., 2014; Dangendorf et al., (2015) found that observed total GMSL change in the 20th century was inconsistent with internal variability. Dangendorf et al. (2015) determined that for 1900 to 2011 at least 45% of GMSL change is human-induced. A study that developed a semi-empirical model to link sea-level change to observed GMST change concluded that at least 41% of the 20th century sea-level rise would not have happened in the absence of the century’s increasing GMST and that there was a 95% probability that by 1970 GMSL was higher than that which would have occurred in the absence of increasing GMST (Kopp et al., 2016). Richter et al. (2020) compared modelled sea level change with the satellite altimeter observations from 1993 to 2015; a period short enough that internal variability can dominate the spatial pattern of change. They found that when GMSL is not removed, model simulated zonally averaged sea level trends are consistent with altimeter observations globally as well as in each ocean basin and much larger than might be expected from internal variability. Using spatial correlation, Fasullo and Nerem (2018) showed that the satellite altimeter trend pattern is already detectable.

We note that current detection and attribution studies do not yet include all processes that are important for sea-level change (see Section 9.6). However, based on the body of literature available, we conclude that the main driver of the observed GMSL rise since at least 1971 is very likely anthropogenic forcing. The assessed period starts in 1971 for consistency with observations assessed in Cross-Chapter Box 9.1.

Circulation of the ocean, whether it be wind or density driven, plays a prominent role in the heat and freshwater transport of the Earth system (Buckley and Marshall, 2016). Thus, its accurate representation is crucial for the realistic representation of water mass properties, and replication of observed changes driven by atmosphere-land-ocean coupling. Here, we assess the ability of CMIP models to reproduce the observed large-scale ocean circulation, along with assessment of the detection and attribution of any anthropogenically-driven changes. We also note that the process-based understanding of these circulation changes and circulation changes occurring at smaller scales is assessed in Section 9.2.3.

3.5.4.1 Atlantic Meridional Overturning Circulation (AMOC)

The Atlantic Meridional Overturning Circulation (AMOC) represents a large-scale flow of warm salty water northward at the surface and a return flow of colder water southward at depth. As such, its mean state plays an important role in transporting heat in the climate system, while its variability can act to redistribute heat (see Sections 2.3.3.4.1 and 9.2.3.1 for more details). Paleo-climatic and model evidence suggest that changes in AMOC strength have played a prominent role in past transitions between warm and cool climatic phases (e.g., Dansgaard et al., 1993; Ritz et al., 2013).

The AR5 concluded that while climate models suggested that an AMOC slowdown would occur in response to anthropogenic forcing, the short direct observational AMOC record precluded it from being used to support this model finding. Chapter 2 reports with high confidence, a weakening of the AMOC was observed in the mid-2000s to the mid-2010s, while again also noting that the observational record was too short to determine whether this is a significant trend or a manifestation of decadal and multi-decadal variability (Section 2.3.3.4.1). Indirect evidence of AMOC weakening since at least the 1950s is also presented, but confidence in this longer-term decrease was low Section 2.3.3.4.1).

Despite the additional six years or so of observations since AR5, the evaluation of the AMOC in models continues to be severely hampered by the geographically sparse and temporally short observational record. The longest continuous observational estimates of the AMOC are based on measurements taken at 26°N by the RAPID-MOCHA array (Smeed et al., 2018). Basic evaluation of the AMOC at 26°N shows that the CMIP5 and CMIP6 multi-model mean overturning strength is comparable with RAPID (Reintges et al., 2017; Weijer et al., 2020), but the model range is large (12–29 sverdrups (Sv)) for CMIP5 (Zhang and Wang, 2013); and 10–31 Sv for CMIP6 (Weijer et al., 2020) (Figure 3.30a). It is noted that deviations of AMOC strength in CMIP5 models have been related to global-scale sea surface temperature biases (C. Wang et al., 2014). Both coupled and ocean-only models also underestimate the depth of the AMOC cell (Danabasoglu et al., 2014; Weijer et al., 2020; Figure 3.30a). Paleo-climatic evidence has also raised questions regarding the accuracy of the representation of the strength and depth of the modelled AMOC during past periods (Otto-Bliesner et al., 2007; Muglia and Schmittner, 2015). Overall, however, both the CMIP5 and CMIP6 model ensembles simulate the general features of the AMOC mean state reasonably well, but there is a large spread in the latitude and depth of the maximum overturning, and the maximum AMOC strength (Figure 3.30a).

Figure 3.30 | Observed and CMIP6 simulated AMOC mean state, variability and long-term trends. (a) AMOC meridional stream function profiles at 26.5°N from the historical CMIP5 (1860–2004) and CMIP6 (1860–2014) simulations compared with the mean maximum overturning depth (horizontal grey line) and magnitude (vertical grey line) from the RAPID observations (2004–2018). The distributions of model ranges of AMOC maximum magnitude and depth are respectively displayed near the x- and y-axis. (b) Distributions of overlapping eight-year AMOC trends from individual CMIP6 historical simulations (pink box plots) are plotted along with the combined distributions of all available CMIP5 (blue boxplot) and CMIP6 (red boxplot) models. For reference, the observed eight-year trend calculated between 2004 and 2012 is also shown as a horizontal grey line (following Roberts et al., 2014). (c) Distributions of interannual AMOC variability from individual CMIP6 model historical simulations, along with the combined distributions of all available CMIP5 and CMIP6 models. Interannual variability in models and observations is estimated as annual mean (April–March) differences, and the horizontal grey line is the observed value for 2009/2010 minus 2008/2009 (following Roberts et al., 2014). (d–f) Distributions of linear AMOC trends calculated over various time periods (see panel titles) in CMIP6 simulations forced with: greenhouse gas forcing only (GHG), natural forcing only (NAT), anthropogenic aerosol forcing only (AER) and all forcing combined (Historical; HIST). (a–f) Boxes indicate the 25th to 75th percentile range, whiskers indicate 1st and 99th percentiles in (a-c) and 5th and 95th percentile in (d-f), and dots indicate outliers, while the horizontal black line is the multi-model mean trend. In (d–f) the multi-model mean trend is also written above each distribution. The multi-model distributions in (a–c) were produced with one historical ensemble member per model for which the AMOC variable was available (listed), while those in (d–f) were produced with the detection and attribution simulation datasets utilized by Menary et al. (2020). Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

The short length of the observed time-series (RAPID has measured the AMOC since 2004), sparse observations, observational uncertainties (Sinha et al., 2018), as well as significant observed variability on interannual and longer time scales, makes comparison with modelled AMOC variability challenging. RAPID observations show that the overturning at 26°N was 2.9 Sv weaker in the multi-year average of 2008–2012 relative to 2004–2008 and 2.5 Sv weaker in 2012–2017 relative to 2004–2008 (Smeed et al., 2014, 2018) (see also (Section 2.3.3.4.1). As expected, this weakening was accompanied by a significant reduction in northward heat transport (Bryden et al., 2020). CMIP5 and CMIP6 models produce a forced weakening of the AMOC over the 2012–2017 period relative to 2004–2008, but at 26°N the multi-model mean response is substantially weaker than the observed AMOC decline over the same period. The discrepancy between the modelled multi-model mean (i.e., the forced response) and the RAPID observed AMOC changes has led studies to suggest that the observed weakening over 2004–2017 is largely due to internal variability (Yan et al., 2018). However, comparison of observed RAPID AMOC variability with modelled variability also reveals that most CMIP5 models appear to underestimate the interannual and decadal time scale AMOC variability (Roberts et al., 2014; Yan et al., 2018), and, although the overall variance is larger in CMIP6 than in CMIP5, similar results are found analysing the CMIP6 models (Figure 3.30b,c). It is currently unknown why most models underestimate this AMOC variability, or whether they are underestimating the internal or externally forced components. This underestimation of AMOC variability may also have potential implications for detection and attribution, the relationship between AMOC and AMV (see Section 3.7.7), and near-term predictions. There is also emerging evidence, based on analysis of freshwater transports, that the AMOC in CMIP5-era models is too stable, largely due to systematic biases in ocean salinity (W. Liu et al., 2017; Mecking et al., 2017). Such a systematic bias may potentially be linked with the underestimation of both simulated AMOC internal variability through eddy-mean flow interactions that are poorly represented in standard CMIP-class model resolution (Leroux et al., 2018), and externally forced change.

As reported in Section 2.3.3.4.1, estimates of AMOC since at least 1950, which are generated from observed surface temperatures or sea surface height, suggest the AMOC weakened through the 20th century (low confidence) (Ezer et al., 2013; Caesar et al., 2018). Over the same period, the CMIP5 multi-model mean showed no significant net forced response in AMOC (Cheng et al., 2013). However, a significant forced change is simulated in the CMIP6 multi-model mean, where a clear increase of the AMOC is seen over the 1940–1985 period (Figure 3.30e; Menary et al., 2020). Although there is general agreement that the influence of greenhouse gases acts to a weaken the modelled AMOC (Delworth and Dixon, 2006; Caesar et al., 2018), changes in solar, volcanic and anthropogenic aerosol emissions can lead to temporary changes in AMOC on decadal- to multi-decadal time scales (Delworth and Dixon, 2006; Menary et al., 2013; Menary and Scaife, 2014; Swingedouw et al., 2017; Undorf et al., 2018b). As such, the simulated net forced response in AMOC is a balance between the different forcing factors (Section 9.2.3.1; Delworth and Dixon, 2006; Menary et al., 2020). The differing AMOC response of CMIP5 and CMIP6 models during the historical period has been associated with stronger aerosol effective radiative forcing in the CMIP6 models (Menary et al., 2020), such that the aerosol-induced AMOC increase during the 1940–1985 period overcomes the greenhouse gas induced decline (Figure 3.30e). However, models simulate a range of anthropogenic aerosol effective radiative forcing and a range of historical AMOC trends in CMIP6 (Menary et al., 2020) and there remains considerable uncertainty over the realism of the CMIP6 AMOC response during the 20th century (Figure 3.30d–f) due to disagreement among the differing lines of evidence. For example, ocean reanalysis (Jackson et al., 2019) and forced ocean model simulations (Robson et al., 2012; Danabasoglu et al., 2016), which show AMOC changes that are broadly consistent with the CMIP6 response, appear to disagree with observational estimates of AMOC over the historical period (Ezer et al., 2013; Caesar et al., 2018). It is noted, however, that the relatively short length of the forced ocean simulations and ocean reanalysis precludes a comparable assessment of 20th century trends. Furthermore, despite the similar AMOC evolution seen in forced ocean model simulations and the CMIP6 models, it is unclear whether the same underlying mechanisms are responsible for the changes.

In summary, models do not support robust assessment of the role of anthropogenic forcing in the observed AMOC weakening between the mid-2000s and the mid-2010s, which is assessed to have occurred with high confidence in Section 2.3.3.4.1, as the changes are outside of the range of modelled AMOC trends (regardless of whether they are forced or internally generated) in most models. Thus, we have low confidence that anthropogenic forcing has influenced the observed changes in AMOC strength in the post-2004 period. In addition, there remains considerable uncertainty over the realism of the CMIP6 AMOC response during the 20th century due to disagreement among the differing lines of observational and modelled evidence (i.e., historical AMOC estimates, ocean reanalysis, forced ocean simulations and historical CMIP6 simulations). Thus, we have low confidence that anthropogenic forcing has had a significant influence on changes in AMOC strength during the 1860–2014 period.

The AR5 did not make attribution statements on changes in global carbon sinks. The IPCC Special Report on Climate Change and Land (SRCCL) assessed with high confidence that global vegetation photosynthetic activity has increased over the last 2–3 decades (Jia et al., 2019). That increase was attributed to direct land use and management changes, as well as to CO₂ fertilization, nitrogen deposition, increased diffuse radiation and climate change (high confidence). The AR5 assessed with high confidence that CMIP5 Earth System Models (ESMs) simulate the global mean land and ocean carbon sinks within the range of observation-based estimates (Flato et al., 2013). The IPCC SRCCL, however, noted the remaining shortcomings of carbon cycle schemes in ESMs (Jia et al., 2019), which for example do not properly incorporate thermal responses of respiration and photosynthesis, and frequently omit representations of permafrost thaw (Comyn-Platt et al., 2018), the nitrogen cycle (R.Q. Thomas et al., 2015) and its influence on vegetation dynamics (Jeffers et al., 2015), the phosphorus cycle (Fleischer et al., 2019), and accurate implications of carbon store changes for a range of land use and land management options (Erb et al., 2018; Harper et al., 2018) (see Sections 5.2.1.4.1 and 5.4, Figure 5.24 and Table 5.4 for details).

This section considers three main large-scale indicators of climate change relevant to the terrestrial carbon cycle: atmospheric CO₂ concentration, atmosphere-land CO₂ fluxes, and leaf area index. These indicators were chosen because they have been the target of attribution studies. Other indicators, like land use and management, and wildfires, relate to human influence but are discussed in Chapter 5. Chapter 7 discusses energetic consequences of changes in the terrestrial carbon cycle in Section 7.4.2.5.2. CMIP5 and CMIP6 ESMs are most often run with prescribed observed historical changes in atmospheric CO₂ concentration and diagnose CO₂ emissions consistent with these. Such calculations require that the models simulate realistic changes in the terrestrial carbon cycle over the historical period, as changes to land carbon stores will influence the size of CO₂ emissions consistent with prescribed CO₂ pathways, and associated remaining carbon budgets (Section 5.5). Such testing of existing models is needed while also recognising there are process representations still requiring inclusion.

Since AR5, atmospheric inversion studies have further tested or constrained models, while new datasets have been used to constrain specific parts of the terrestrial carbon cycle such as plant respiration (Huntingford et al., 2017). Figure 3.31 compares historical emissions-driven CMIP6 simulations of global mean atmospheric CO₂ concentration and net ocean and land carbon fluxes to the assessed CO₂ concentration and fluxes from the Global Carbon Project (Friedlingstein et al., 2019). For 2014, the CMIP6 models simulate a range of CO₂ concentrations centred around the observed value of 397 ppmv, with a range of 381 to 412 ppmv. GSAT anomalies simulated over the historical period are very similar in models that simulate or prescribe changes in atmospheric CO₂ concentrations (Figures 3.31b and 3.4a). Most models simulate realistic temporal evolution of the global net ocean and land carbon fluxes, although model spread is larger over land (Figure 3.31c,d; see also Sections 3.6.2 and 5.4.5.2, and Figure 5.24). Although literature published soon after AR5 highlighted the importance of representing nitrogen limitation on plant growth (Peng and Dan, 2015; R.Q. Thomas et al., 2015), more recent studies note that models without nitrogen limitation can still be consistent with the latest estimates of historical carbon cycle changes (Arora et al., 2020; Meyerholt et al., 2020). Uncertainties in the photosynthetic response to atmospheric CO₂ concentrations at global scales, shifts in carbon allocation and turnover, land-use change (Hoffman et al., 2014; Wieder et al., 2019), and water limitation are also important influences on land carbon fluxes.

Figure 3.31 | Evaluation of historical emissions-driven CMIP6 simulations for 1850–2014. Observations (black) are compared to simulations of global mean (a) atmospheric CO₂ concentration (ppmv), with observations from the National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA ESRL; Dlugokencky and Tans, 2020); (b) surface air temperature anomaly (°C) with respect to the 1850–1900 mean, with observations from HadCRUT4 (Morice et al., 2012); (c) land carbon uptake (PgC yr^–1), (d) ocean carbon uptake (PgC yr^–1), both with observations from the Global Carbon Project (GCP; Friedlingstein et al., 2019) and grey shading indicating the observational uncertainty. Land and ocean carbon uptakes are plotted using a 10-year running mean for better visibility. The ocean uptake is offset to 0 in 1850 to correct for pre-industrial riverine-induced carbon fluxes. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

All models and observational estimates agree that interannual variability in net CO₂ uptake is much larger over land than over the ocean. Studies demonstrate that regional variations in both the trends and the yearly strength of the terrestrial carbon sink are considerable. Land carbon uptake is dominated by the extratropical northern latitudes (see also Section 5.4.5.3 and Figure 5.25; Ciais et al., 2019) because the tropics may have become a net source of carbon (Baccini et al., 2017). At local to regional scales, the dominant driver of yearly sink strength variations is water availability, but at continental to global scales, temperature anomalies are the dominant driver (Section 5.2.1.4.2; Jung et al., 2017). The major role of levels of water stored in the ground in influencing land-atmosphere CO₂ exchange has also been confirmed through simultaneous analysis of satellite gravimetry and atmospheric CO₂ levels (Humphrey et al., 2018). When considered globally, simulated land and ocean carbon sinks fall within the range of observation-based estimates with high confidence. But there is also high confidence that that apparent success arises for the wrong reasons, as models underestimate the Northern Hemisphere carbon sink, as discussed in Section 5.4.5.3.

The seasonal cycle in atmospheric CO₂, which is driven by the drawdown of carbon by photosynthesis on land during the summer and release by respiration during the winter, has increased in amplitude since the start of systematic monitoring (Figure 3.32; see also (Section 2.3.4.1). This trend, which is larger at higher latitudes of the Northern Hemisphere, was first reported by Keeling et al. (1996) and has continued. Changes in vegetation productivity have also been observed, as well as longer growing seasons (Park et al., 2016). However, a slow down of the increasing trend has been noted, linked to a slow down of both vegetation greening and growing-season length increases (Buermann et al., 2018; Z. Li et al., 2018; K. Wang et al., 2020). Figure 3.32 shows that CMIP6 terrestrial carbon cycle models partially capture the increasing amplitude of the seasonal cycle of the land carbon sink, also seen in observational reconstructions. However, the identification of the human influence that contributes most strongly to these changes in the seasonal cycle is debated.

Figure 3.32 | Relative change in the amplitude of the seasonal cycle of global land carbon uptake in the historical CMIP6 simulations from 1961–2014. Net biosphere production estimates from 19 CMIP6 models (red), the data-led reconstruction JMA-TRANSCOM (Maki et al., 2010; dotted) and atmospheric CO₂ seasonal cycle amplitude changes from observations (global as dashed line, Mauna Loa Observatory (MLO) (Dlugokencky et al., 2020) in bold black). Seasonal cycle amplitude is calculated using the curve fit algorithm package from the National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA ESRL). Relative changes are referenced to the 1961–1970 mean and for short time series adjusted to have the same mean as the model ensemble in the last 10 years. Interannual variation was removed with a nine-year Gaussian smoothing. Shaded areas show the one sigma model spread (grey) for the CMIP6 ensemble and the one sigma standard deviation of the smoothing (red) for the CO₂ MLO observations. Inset: average seasonal cycle of ensemble mean net biosphere production and its one sigma model spread for 1961–1970 (orange dashed line, light orange shading) and 2005–2014 (solid green line, green shading). Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Proposed causes of the trend in the amplitude of the seasonal cycle of CO₂, and its amplification at higher latitudes, include increases in the summer productivity and/or increases in the magnitude of winter respiration of northern ecosystems (Barichivich et al., 2013; Graven et al., 2013; Forkel et al., 2016; Wenzel et al., 2016), increases in productivity throughout the Northern Hemisphere by CO₂ fertilization, and increases in the productivity of agricultural crops in northern mid-latitudes (Gray et al., 2014; Zeng et al., 2014). Recent studies have attempted to quantify the different contributions by comparing atmospheric CO₂ observations with ensembles of land surface model simulations. Piao et al. (2017) found that CO₂ fertilization of photosynthesis is the main driver of the increase in the amplitude of the seasonal cycle of atmospheric CO₂ but noted that climate change drives the latitudinal differences in that increase. North of 40°N, Bastos et al. (2019) also found CO₂ fertilization to be the most likely driver, with warming at northern high latitudes contributing a decrease in amplitude, in contrast to earlier conclusions (Graven et al., 2013; Forkel et al., 2016), and agricultural and land use changes making only a small contribution. For temperate regions of the Northern Hemisphere, K. Wang et al. (2020) found that the importance of CO₂ fertilization is decreased by drought stress, but also found only a small contribution from agricultural and land use changes. However, many global models do not include nitrogen fertilization, changes to crop cultivars or irrigation effects, with the latter associated with deficiencies in simulated terrestrial water cycling (H. Yang et al., 2018). All these factors influence the capability of models to simulate accurately the seasonal cycle in atmosphere-land CO₂ exchanges. Model comparisons to the atmospheric CO₂ concentration record for Barrow, Alaska, suggest that models underestimate current levels of carbon fixation (Winkler et al., 2019) and have deficiencies in their phenological representation of greenness levels, particularly for autumn (Z. Li et al., 2018). Based on these studies and noting the uncertainty in the processes ultimately driving changes in atmospheric CO₂ seasonal cycles (Section 5.2.1.4), we assess with medium confidence that fertilization by anthropogenic increases in atmospheric CO₂ concentrations is the main driver of the increase in the amplitude of the seasonal cycle of atmospheric CO₂.

Detection and attribution methods have been applied to leaf area index, which represents ‘greenness’ and general photosynthetic productivity (see Section 2.3.4.3). Nitrogen deposition and land cover change trends remain small compared to variability, so attributing changes in leaf area index to those processes is difficult. Using three satellite products and ten land models, Zhu et al. (2016) found increases in leaf area index (greening) over 25–50% of global vegetated areas, and they attributed 70% of this greening to CO₂ fertilization, although they found that land use change can dominate regionally. This is consistent with the attribution study of observed greening of Mao et al. (2016), and with Mao et al. (2013) who found that CO₂ fertilization was the dominant cause of enhanced vegetation growth, with latitudinal changes in leaf area index explained by the larger land surface warming in the Northern Hemisphere. These conclusions are also consistent with those of Zhu et al. (2017), who found a dominant role for CO₂ fertilization in driving leaf area index changes in an attribution study in which land models were first weighted by performance. However, Chen et al. (2019) has challenged these results by showing that greening in India and China was driven by land-use change.

Leaf area index increases attributed to CO₂ fertilization are due to a direct raised physiological response. However, for drylands, CO₂-induced stomatal closure may act to conserve soil moisture and thereby indirectly drive higher photosynthesis through higher water use efficiency (Lu et al., 2016). In models with nitrogen deposition, there is evidence that this simulated effect also influences leaf area index trends, however, because of a lack of literature based on large-scale land simulations including both nutrient limitation and crop intensification, it is not yet possible to make an attribution statement about their individual roles in leaf area index changes.

In summary, Earth system models simulate globally averaged land carbon sinks within the range of observation-based estimates (high confidence), but global-scale agreement masks large regional disagreements. Based on new studies that attribute changes in atmospheric CO₂ seasonal cycle to CO₂ fertilization, albeit counteracted by other factors, combined with the medium confidence that models represent the processes driving changes in the seasonal cycle, we assess that there is medium confidence that CO₂ fertilization is the main driver of the increase in the amplitude of the seasonal cycle of atmospheric CO₂. Based on available literature, CO₂ fertilization has been the main driver of the observed greening trend, but there is only low confidence in this assessment because of ongoing debate about the relative roles of CO₂ fertilization, high latitude warming, and land management, and the low number of models that represent the whole suite of processes involved.

Since CMIP5, there has been a general increase in ocean horizontal and vertical grid resolution in ocean model components (Arora et al., 2020; Séférian et al., 2020). The latter of these developments is particularly significant for projections of ocean stressors as it directly affects the representation of stratification. Updates in the representation of ocean biogeochemical processes between CMIP5 and CMIP6 have typically involved an increase in model complexity. Specific developments have been the more widespread inclusion of micronutrients, such as iron, variable stoichiometric ratios, more detailed representation of lower trophic levels including bacteria and the cycling and sinking of organic matter. CMIP6 biogeochemical model performance is generally an improvement on that of the parent CMIP5 generation of models (Séférian et al., 2020). The global representation of present-day air-sea carbon fluxes and surface chlorophyll concentrations show moderate improvements between CMIP5 and CMIP6. Similar improvements are seen in the representation of subsurface oxygen concentrations in most ocean basins, while the representation of surface macronutrient concentrations in CMIP6 is shown to have improved with respect to silicic acid but declined slightly with respect to nitrate. Model representation of the micronutrient iron has not improved substantially since CMIP5, but many more models are capable of representing iron. In addition, a comparison of the carbon concentration and carbon climate feedbacks shows no significant change between CMIP5 and CMIP6 (Arora et al., 2020).

Since AR5, research has also focused on the detection and attribution of regional patterns in ocean biogeochemical change relating to interior deoxygenation, air-sea CO₂ flux, and ocean carbon uptake and associated acidification. Characterization of flux variability requires understanding of the suite of physical and biological processes including transport, heat fluxes, interior ventilation, biological production and gas exchange which can have very different controls on seasonal versus interannual time scales in both the North Pacific (Ayers and Lozier, 2012) and North Atlantic (Breeden and McKinley, 2016). In the Southern Ocean, models have difficulty reproducing the observed seasonal cycle and interannual variability, making attribution particularly challenging (Lovenduski et al., 2016; Mongwe et al., 2016, 2018).

The AR5 concluded that oxygen concentrations have decreased in the open ocean since 1960 and such decreases can be attributed in part to human influence with medium confidence. The decrease in ocean oxygen content in the upper 1000 m, between 1970 and 2010, is further confirmed in SROCC (medium confidence), with the oxygen minimum zone expanding in volume (see also Section 5.3.3.2). Observed oxygen declines over the last several decades (Stendardo and Gruber, 2012; Stramma et al., 2012; Schmidtko et al., 2017) match model estimates in the surface ocean (Oschlies et al., 2017) but are much larger than model derived estimates in the interior (Bopp et al., 2013; Cocco et al., 2013). Some of this difference has been interpreted as due to a lack of representation of coastal eutrophication in these models (Breitburg et al., 2018), but much of it remains unexplained. This disparity is particularly apparent in the eastern Pacific oxygen minimum zone, where some CMIP5 models showed increasing trends whereas observations show a strong decrease (Cabré et al., 2015). However, proxy reconstructions suggest that over the last century the ocean may have in fact undergone increases in oxygen in the most oxygen poor regions (Deutsch et al., 2014). As discussed in Section 5.3.1, ocean oxygen went through wide oscillations on multi-centennial time scales through the last deglaciation, with abrupt warming resulting in loss of oxygen in subsurface waters of the North Pacific (Praetorius et al., 2015). The global upper ocean oxygen inventory is negatively correlated with ocean heat content with a regression coefficient comparable to that found in ocean models (Ito et al., 2017). Variability and trends in the observed upper ocean oxygen concentration are mainly driven by the apparent oxygen utilization component with small contributions from oxygen solubility, suggesting that changing ocean circulation, mixing, and/or biochemical processes, rather than thermally induced solubility effects may be the main drivers of observed deoxygenation. The spatial distribution of the ocean deoxygenation in the interior of the ocean as well as over coastal areas is further assessed in Section 5.3.

As one of the most commonly observed surface parameters, the partial pressure of CO₂ has been the topic of considerable detection and attribution work. In North Atlantic subtropical and equatorial biomes, warming has been shown to be a significant and persistent contributor to the observed increase in the partial pressure of CO₂ since the mid‐2000s with long‐term warming leading to a reduction in ocean carbon uptake (Fay and McKinley, 2013), and with both the partial pressure of CO₂ and associated carbon uptake demonstrating strong predictability as a function of interannual to decadal climate state (H. Li et al., 2016; Li and Ilyina, 2018). In the Southern Ocean however, detection and attribution of surface trends in the partial pressure of CO₂ has proven more elusive and dependent on methodology, with some studies suggesting that Southern Ocean carbon uptake slowed from about 1990 to 2006 and subsequently strengthened from 2007 to 2010 (Lovenduski et al., 2008; Fay et al., 2014; Ritter et al., 2017). Other studies have suggested that poor representation of the seasonal cycle in the Southern Ocean may confound the models’ ability to represent changes in the partial pressure of CO₂ in the Southern Ocean (Nevison et al., 2016; Mongwe et al., 2018).

Section 5.2.1.3 assesses that both observational reconstructions based on the partial pressure of CO₂ and ocean biogeochemical models show a quasi-linear increase in the ocean sink of anthropogenic CO₂ from 1.0 ± 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 (high confidence). During the 1990s, the global net flux of CO₂ into the ocean is estimated to have weakened to 0.8 ± 0.5 PgC yr^–1 while in 2000 and thereafter, it is estimated to have strengthened considerably to rates of 2.0 ± 0.5 PgC yr^–1, associated with changes in SST, the surface concentration of dissolved inorganic carbon and alkalinity, and decadal variations in atmospheric forcing (Landschützer et al., 2016, see also Section 5.2).

Ocean acidification is one of the most detectible metrics of environmental change and was well covered in AR5, in which it was assessed that the uptake of anthropogenic CO₂ had very likely resulted in acidification of surface waters (Bindoff et al., 2013). Since then, observations and simulations of multi-decadal trends in surface carbon chemistry have increased in robustness. The evidence on ocean pH decline had further strengthened in SROCC with good agreement found between CMIP5 models and observations and an assessment that the ocean was continuing to acidify in response to ongoing carbon uptake (Bindoff et al., 2019). An observed decrease in global surface open ocean pH is assessed in Section 2.3.3.5 to be virtually certain to have occurred with a rate of 0.003–0.026 per decade for the past 40 years. The ocean acidification has occurred not only in the surface layer but also in the interior of the ocean (Sections 2.3.3.5 and 5.3.3). Rates have been observed to be between −0.015 and −0.020 per decade in mode and intermediate waters of the North Atlantic through the combined effect of increased anthropogenic and remineralized carbon (Ríos et al., 2015) and acidification has been observed down to 3000 m in the deep water formation regions (Perez et al., 2018). There has also been considerable improvement in detection and attribution of anthropogenic CO₂ versus eutrophication-based acidification in coastal waters (Wallace et al., 2014).

The increased evidence in recent studies supports an assessment that it is virtually certain that the uptake of anthropogenic CO₂ was the main driver of the observed acidification of the global surface open ocean. The observed increase in acidification over the North Atlantic subtropical and equatorial regions since 2000 is likely associated in part with an increase in ocean temperature, a response which corresponds to the expected weakening of the ocean carbon sink with warming. Due to strong internal variability, systematic changes in carbon uptake in response to climate warming have not been observed in most other ocean basins at present. We further assess, consistent with AR5 and SROCC, that deoxygenation in the upper ocean is due in part to anthropogenic forcing, with medium confidence. There is high confidence that Earth system models simulate a realistic time evolution of the global mean ocean carbon sink.

The Northern Annular Mode (NAM; also known as the Arctic Oscillation) is an oscillation of atmospheric mass between the Arctic and northern mid-latitudes, analogous to the Southern Annular Mode (SAM; Section 3.7.2). It is the leading mode of variability of sea-level pressure in the northern extratropics but also has a clear fingerprint through the troposphere up to the lower stratosphere, with maximum expression in boreal winter (Kidston et al., 2015). The North Atlantic Oscillation (NAO) can be interpreted as the regional expression of the NAM and captures most of the related variance in the troposphere over a broad North Atlantic/Europe domain. Indices measuring the state of the NAO correlate highly with those of the NAM, and teleconnection patterns for both modes are rather similar (Feldstein and Franzke, 2006). A detailed description of the NAM and the NAO as well as their associated teleconnection over land is given in Annex IV.2.1.

AR5 found that while models simulated correctly most of the spatial properties of the NAM, substantial inter-model differences remained in the details of the associated teleconnection patterns over land (Flato et al., 2013). The AR5 reported that most models did not reproduce the observed positive trend of the NAO/NAM indices during the second half of the 20th century. It was unclear to what extent this failure reflected model shortcomings and/or if the observed trend could be simply related to pronounced internal climate variability. The AR5 accordingly did not make an attribution assessment for the NAO/NAM.

New studies since AR5 continue to find that CMIP5 models reproduce the spatial structure and magnitude of the NAM reasonably well (Lee and Black, 2013; Zuo et al., 2013; Davini and Cagnazzo, 2014; Ying et al., 2014; Ning and Bradley, 2016; Deser et al., 2017b; Gong et al., 2017) although the North Pacific SLP anomalies remain generally too strong (Zuo et al., 2013; Gong et al., 2017) and the subtropical North Atlantic lobe of SLP anomalies conversely too weak (Ning and Bradley, 2016) in many models. Such overall biases noted in both CMIP3 and CMIP5 (Davini and Cagnazzo, 2014) persist in CMIP6 historical simulations, even though the multi-model multi-member ensemble mean spatial correlation between modelled and observed NAM is slightly higher (Figure 3.33a,d,g). Regarding the NAO, the majority of CMIP5 models very successfully simulate its spatial structure (Lee et al., 2019) and its associations with extratropical jet, storm track and blocking variations over a broad North-Atlantic/Europe domain (Davini and Cagnazzo, 2014) and over land through teleconnections (Volpi et al., 2020). The good performance of the models is confirmed in CMIP6 with a marginal improvement of the averaged observation-model spatial correlation (Figure 3.33b,e,h) and better skill based on other evaluation metrics (Fasullo et al., 2020). The slight underestimation of the SLP anomalies related to the NAO centres of actions over the Azores and Greenland–Iceland–Norwegian Seas remain unchanged compared to CMIP5.

Figure 3.33 | Model evaluation of NAM, NAO and SAM in boreal winter. Regression of Mean Sea Level Pressure (MSLP) anomalies (in hPa) onto the normalized principal component (PC) of the leading mode of variability obtained from empirical orthogonal decomposition of the boreal winter (December–February) MSLP poleward of 20°N for the observed Northern Annular Mode (NAM, a), over 20°N–80°N, 90°W–40°E for the North Atlantic Oscillation as shown by the black sector (NAO, b), and poleward of 20°S for the Southern Annular Mode (SAM, c) for the JRA-55 reanalysis. Cross marks indicate regions where the anomalies are not significant at the 10% level based on a t-test. The period used to calculate the NAO/NAM is 1958–2014 but 1979–2014 for the SAM. (d–f) Same but for the multi-model ensemble (MME) mean from CMIP6 historical simulations. Models are weighted in compositing to account for differences in their respective ensemble size. Diagonal lines show regions where less than 80% of the runs agree in sign. (g–i) Taylor diagrams summarizing the representation of the modes in models and observations following Lee et al. (2019) for CMIP5 (light blue) and CMIP6 (red) historical simulations. The reference pattern is taken from JRA-55 (a–c). The ratio of standard deviation to that of the reference observations (radial distance), spatial correlation (radial angle) and resulting root-mean-squared errors (solid isolines) are given for individual ensemble members (crosses) and for other observational products (ERA5 and NOAA 20CR version 3, black dots). Coloured dots stand for weighted multi-model mean statistics for CMIP5 (blue) and CMIP6 (light red) as well as for AMIP simulations from CMIP6 (orange). (j–l) Histograms of the trends built from all individual ensemble members and all the models (brown bars). Vertical lines in black show all the observational estimates. The orange, light red, and light blue lines indicate the weighted multi-model mean of CMIP6 AMIP, CMIP6 and CMIP5 historical simulations, respectively. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

CMIP5 models with a model top within the stratosphere seriously underestimate the amplitude of the variability of the wintertime NAM expression in the stratosphere, in contrast to CMIP5 models which extend well above the stratopause (Lee and Black, 2015). However, even in the latter models, the stratospheric NAM events, and their downward influence on the troposphere, are insufficiently persistent (Charlton-Perez et al., 2013; Lee and Black, 2015). Increased vertical resolution does not show any significant added value in reproducing the structure and magnitude of the tropospheric NAM (Lee and Black, 2013) nor in the NAO predictability as assessed in a seasonal prediction context with a multi-model approach (Butler et al., 2016). On the other hand, there is mounting evidence that a correct representation of the Quasi Biennal Oscillation, extratropical stratospheric dynamics (the polar vortex and sudden stratospheric warmings), and related troposphere-stratosphere coupling, as well as their interplay with ENSO, are important for NAO/NAM timing (Scaife et al., 2016; Karpechko et al., 2017; Domeisen, 2019; Domeisen et al., 2019), in spite of underestimated troposphere–stratosphere coupling found in models compared to observations (O’Reilly et al., 2019b).

The observed trend of the NAM and NAO indices is positive in winter when calculated from the 1960s (Section 2.4.1.1) but it includes large multi-decadal variability, which means that the nature of the trend should be interpreted with caution (Gillett et al., 2013). The multi-model multi-member ensemble mean of the trend estimated from historical simulations over that period is very close to zero for both CMIP5 and CMIP6 (Figures 3.33j,k and 3.34a). Even if one cannot rule out that 1958–2014 was an exceptional period of variability, the observational estimates of the wintertime NAO trend lie outside the 5th–95th percentile range of the distribution of trends in the CMIP6 historical simulations, and the observed NAM trends over the same period lie above the 90th percentile. There is a tendency for the CMIP5 models to systematically underestimate the level of multi-decadal versus interannual variability of the winter NAO and jet stream compared to observations (X. Wang et al., 2017; Bracegirdle et al., 2018; Simpson et al., 2018). Results from CMIP6 (Figure 3.33j,k) and over the 1958–2019 period (Figure 3.34a) confirm this conclusion and seriously question the ability of the models to simulate long-term fluctuations of the NAO/NAM, independently of its forced or internal origins.

Figure 3.34 | Attribution of observed seasonal trends in the annular modes to forcings. Simulated and observed trends in NAM indices over 1958–2019 (a) and in SAM indices over 1979–2019 (b) and over 2000–2019 (c) for boreal winter (December–February average; DJF) and summer (June–August average; JJA). The indices are based on the difference of the normalized zonally averaged monthly mean sea level pressure between 35°N and 65°N for the NAM and between 40°S and 65°S for the SAM as defined in Jianping and Wang (2003) and Gong and Wang (1999), respectively; the unit is decade^–1. Ensemble mean, interquartile ranges and 5th and 95th percentiles are represented by empty boxes and whiskers for pre-industrial control simulations and historical simulations. The number of ensemble members and models used for computing the distribution is given in the upper-left legend. Grey lines show observed trends from the ERA5 and JRA-55 reanalyses. Multi-model multi-member ensemble means of the forced component of the trends as well as their 5–95% confidence intervals assessed from t-statistics, are represented by filled boxes, based on CMIP6 individual forcing simulations from DAMIP ensembles; greenhouse gases in brown, aerosols in light blue, stratospheric ozone in purple and natural forcing in green. Models with at least three ensemble members are used for the filled boxes, with black dots representing the ensemble means of individual models. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Dedicated SST-forced stand-alone atmospheric model experiments (AMIP) suggest that ocean forcing appears to play a role in decadal variability of the NAO and associated fluctuations in the strength of the jet (Woollings et al., 2015). In particular, Atlantic and Indian Ocean SST anomalies (Fletcher and Cassou, 2015; Baker et al., 2019; Douville et al., 2019; Dhame et al., 2020) may have contributed to the long-term positive trend of the winter NAO/NAM over the 20th century, but there is only low confidence in such a causal relationship because of the limitation of the imposed SST approach in AMIP and the uncertainties in observed SST trends among datasets used as forcing of the atmospheric model. The representation of the NAM and NAO spatial structure is slightly improved in AMIP ensembles (Figure 3.33g,h), which also produce slightly larger trends than the historical simulations for the NAO, but not for the NAM.

When calculated over the most recent two decades, the wintertime NAM/NAO trend is weakly negative since the mid-1990s (Hanna et al., 2015). Recent studies based on observations (Gastineau and Frankignoul, 2015) and dedicated modelling experiments (Davini et al., 2015; Peings and Magnusdottir, 2016) suggest that the recent dominance of negative NAM/NAO could be partly related to the latest shift of the Atlantic Multi-decadal Variability (AMV) to a warm phase (Sections 2.4.4 and 3.7.7). Some recent modelling studies also find that the Arctic sea ice decline might be partly responsible for more recurrent negative NAM/NAO (Peings and Magnusdottir, 2013; B.M. Kim et al., 2014; Nakamura et al., 2015), while other studies do not robustly identify such responses in models (see also Cross-Chapter Box 10.1).

In contrast to winter, the observed trend of the NAO index over 1958–2014 is overall negative in summer and is associated with more recurrent blocking conditions over Greenland, in particular since the mid-1990s, thus contributing to the acceleration of melting of the Arctic sea ice (Section 3.4.1.1) and Greenland Ice Sheet (Section 3.4.3.2; Fettweis et al., 2013; Hanna et al., 2015; Ding et al., 2017). The origin of the negative trend of the summer NAO has not been clearly identified, and is hypothesized to be the result of combined influences (Lim et al., 2019), though trends in summertime NAO should also be interpreted with caution because of the presence of strong multi-decadal variability. The recent observed negative NAO prevalence and related blocking over Greenland is not present in any of the CMIP5 models (Hanna et al., 2018).

Regarding the influence of external forcings since pre-industrial times, AR5 noted that CMIP5 models tend to show an increase in the NAM in response to greenhouse gas increases (Bindoff et al., 2013). Based on the CMIP5 historical ensemble, Gillett and Fyfe (2013) however showed that such a trend is not significant in all seasons. A multi-model assessment of eight CMIP5 models found a NAM increase in response to greenhouse gases, but no robust influence of aerosol changes (Gillett et al., 2013). As for ozone depletion, there is no robust detectable influence on long-term trends of the NAO/NAM (Karpechko et al., 2018) in contrast to the SAM (Section 3.7.2), but there are indications that extreme Arctic ozone depletion events and their surface expression are linked to an anomalously strong NAM episodes (Calvo et al., 2015; Ivy et al., 2017). However, the direction of causality here is not clear.

Conclusions on external forcing influences on the NAM are supported by CMIP6 results based on single forcing ensembles (Figure 3.34a). Positive trends are found in historical simulations over 1958–2019 in boreal winter and are mainly driven by greenhouse gas increases. No significant trends are simulated in response to anthropogenic aerosols, stratospheric ozone or natural forcing. Albeit weak and not statistically significant, the sign of the multi-model mean forced response due to natural forcing is consistent with the observed reduction of solar activity since the 1980s (Section 2.2.1) whose influence would have favoured the negative phase of wintertime NAM/NAO based on the fingerprint of the nearly periodical 11-year solar cycle extracted from models (Scaife et al., 2013; Andrews et al., 2015; Thiéblemont et al., 2015) or observations (Gray et al., 2016; Lüdecke et al., 2020). But such an NAO response to solar forcing remains highly uncertain and controversial, being contradicted by longer proxy records over the last millennium (Sjolte et al., 2018) and modelling evidence (Gillett and Fyfe, 2013; Chiodo et al., 2019). For all seasons and for all individual forcings, uncertainties remain in the estimation of the forced response in the NAM trend as evidenced by considerable model spread (Figure 3.34a) and because the simulated forced component has small amplitude compared to internal variability.

Despite new efforts since AR5 to reconstruct the NAO beyond the instrumental record, it is still very challenging to assess the role of external forcings in the apparent multi-decadal to centennial variability present throughout the last millennium. Large uncertainties remain in the reconstructed NAO index that are sensitive to the types of proxies and statistical methods (Trouet et al., 2012; Ortega et al., 2015; Anchukaitis et al., 2019; Cook et al., 2019; Hernández et al., 2020; Michel et al., 2020) and reconstructed NAO variations are often not reproduced using pseudo-proxy approaches in models (Lehner et al., 2012; Landrum et al., 2013). At low frequency, it remains challenging to evaluate if the observed or reconstructed signal corresponds to an actual change in the NAO intraseasonal to interannual intrinsic properties or rather to a change in the mean background atmospheric circulation changes projecting on a specific phase of the mode. Consequently, conflicting results emerge in the attribution of reconstructed long-term variations in the NAO to solar forcing, whose influence thus remains controversial (Gómez-Navarro and Zorita, 2013; Moffa-Sánchez et al., 2014; Ortega et al., 2015; Ait Brahim et al., 2018; Sjolte et al., 2018; Xu et al., 2018). Influences from major volcanic eruptions appear to be more robust (Ortega et al., 2015; Swingedouw et al., 2017) even if some modelling experiments question the amplitude of the response, which mostly projects on the positive phase of the NAM/NAO (Bittner et al., 2016). The forced response is dependent on the strength, seasonal timing and location of the eruption but may also depend on the mean climate background state (Zanchettin et al., 2013) and/or the phases of the main modes of decadal variability such as the AMV (Section 3.7.7; Ménégoz et al., 2018).

Finally, there is some evidence of an apparent signal-to-noise problem referred to as ‘paradox’ in seasonal and decadal hindcasts of the NAO over the period 1979–2018 (Scaife and Smith, 2018), which suggests that the NAO response to external forcing, SST or sea ice anomalies could be too weak in models. The weakness of the signal has been related to troposphere-stratosphere coupling which is too intermittent (O’Reilly et al., 2019b) and to chronic model biases in the persistence of NAO/NAM daily regimes, which is critically underestimated in coupled models (Strommen and Palmer, 2019; Zhang and Kirtman, 2019), and which does not exhibit significant improvement when model resolution is increased (Fabiano et al., 2020). Note, however, that the apparent signal-to-noise problem may be dependent on the period analysed over the 20th century, which questions its interpretation as a general characteristic of coupled models (Weisheimer et al., 2020).

In summary, CMIP5 and CMIP6 models are skilful in simulating the spatial features and the variance of the NAM/NAO and associated teleconnections (high confidence). There is limited evidence for a significant role for anthropogenic forcings in driving the observed multi-decadal variations of the NAM/NAO from the mid 20th century. Confidence in attribution is low: (i) because there is a large spread in the modelled forced responses which is overwhelmed anyway by internal variability; (ii) because of the apparent signal-to-noise problem; and (iii) because of the chronic inability of models to produce a range of trends which encompasses the observed estimates over the last 60 years.

The Southern Annular Mode (SAM) consists of a meridional redistribution of atmospheric mass around Antarctica (Figure 3.33c,f), associated with a meridional shift of the jet and surface westerlies over the Southern Ocean. SAM indices are variously defined as the difference in zonal-mean sea level pressure or geopotential height between middle and high latitudes or via a principal-component analysis (Annex IV.2.2). Observational aspects of the SAM are assessed in Section 2.4.1.2.

AR5 assessed that CMIP5 models have medium performance in reproducing the SAM with biases in pattern (Flato et al., 2013). It also concluded that the trend of the SAM toward its positive phase in austral summer since the mid-20th century is likely to be due in part to stratospheric ozone depletion, and there was medium confidence that greenhouse gases have also played a role (Bindoff et al., 2013). Based on proxy reconstructions, AR5 found with medium confidence that the positive SAM trend since 1950 was anomalous compared to the last 400 years (Masson-Delmotte et al., 2013).

Additional research has shown that CMIP5 models reproduce the spatial structure of the SAM well, but tend to overestimate its variability in austral summer at interannual time scales, although this variability is within the observational uncertainty (Figure 3.33c,f,i; Zheng et al., 2013; Schenzinger and Osprey, 2015). This is related to the models’ tendency to simulate slightly more persistent SAM anomalies in summer compared to reanalyses (Schenzinger and Osprey, 2015; Bracegirdle et al., 2020). This may be due in part to too weak a negative feedback from tropospheric planetary waves (Simpson et al., 2013). CMIP6 models show improved performance in reproducing the spatial structure and interannual variance of the SAM in summer based on Lee et al. (2019) diagnostics (Figure 3.33i), with a better match of its trend with reanalyses over 1979–2014 (Figure 3.33l), more realistic persistence and improved positioning of the westerly jet, which in CMIP5 models on average is located too far equatorward (Bracegirdle et al., 2020; Grose et al., 2020). In CMIP5, it is also found that models which extend above the stratopause tend to simulate stronger summertime trends in the late 20th century than their counterparts with tops within the stratosphere (Rea et al., 2018; Son et al., 2018), though other differences between these sets of models, such as additional physical processes operating in the stratosphere or interactive ozone chemistry, may have also affected these results (Gillett et al., 2003a; Sigmond et al., 2008; Rea et al., 2018). At the surface, Ogawa et al. (2015) demonstrate with an atmospheric model the importance of sharp mid-latitude SST gradients for stratospheric ozone depletion to affect the SAM in summer. These studies imply that the well resolved stratosphere combined with finer ocean horizontal resolution has contributed to the stronger simulated trends in CMIP6 than in CMIP5.

CMIP6 historical simulations capture the observed positive trend of the summertime SAM when calculated from the 1970s to the 2010s (Figure 3.34b). J.L. Thomas et al. (2015) found that the chance of the observed 1980–2004 trend occurring only due to internal variability is less than 10% in many of the CMIP5 models, and results from CMIP6 models suggest that the chance of the 1979–2019 trend being due to internal variability could be even lower (Figure 3.34b). Although paleo-reconstructions of the SAM index are uncertain and vary in terms of long-term trends (Section 2.4.1.2), new reconstructions show that the 60-year summertime SAM trend since the mid-20th century is outside the 5th–95th percentile range of the trends in the pre-industrial variability, which matches the trend range of CMIP5 pre-industrial control simulations well (Dätwyler et al., 2018).

In general agreement with AR5, new research continues to indicate that both stratospheric ozone depletion and increasing greenhouse gases have contributed to the trend of the SAM during austral summer toward its positive phase in recent decades (Solomon and Polvani, 2016), with the ozone depletion influence dominating (Gerber and Son, 2014; Son et al., 2018). In CMIP6 historical simulations there are significant positive SAM trends over the 1979–2019 period in austral summer, although the contribution from ozone forcing evaluated with the four available models is not significant (Figure 3.34b). Three of these models share the same standard prescribed ozone forcing and produce significantly positive SAM trends over an extended period (1957–2019). The fourth model, MRI-ESM2-0, has the option of interactive ozone chemistry. Its ozone-only experiment is forced by prescribed ozone derived from its own historical simulations and produces a negative SAM trend associated with weak ozone depletion (Morgenstern et al., 2020). Morgenstern et al. (2014) and Morgenstern (2021) find an indirect influence of greenhouse gases on the SAM via induced ozone changes in coupled chemistry-climate simulations, which differ from the prescribed ozone simulations shown in Figure 3.34b. Since about 1997, the effective abundance of ozone-depleting halogen has been decreasing in the stratosphere (WMO, 2018), leading to a stabilization or even a reversal of stratospheric ozone depletion (Sections 2.2.5.2 and 6.3.2.2). The ozone stabilization and slight recovery since about 2000 may have caused a pause in the summertime SAM trend (Figure 3.34c; Saggioro and Shepherd, 2019; Banerjee et al., 2020), although some influence from internal variability cannot be ruled out. While some studies find an anthropogenic aerosol influence on the summertime SAM (Gillett et al., 2013; Rotstayn, 2013), recent studies with larger multi-model ensembles find that this effect is not robust (Steptoe et al., 2016; Choi et al., 2019), consistent with CMIP6 single forcing ensembles (Figure 3.34). In the CMIP5 simulations, volcanic stratospheric aerosol has a significant weakening effect on the SAM in autumn and winter (Cross-Chapter Box 4.1; Gillett and Fyfe, 2013), but there is no evidence that this effect leads to a significant multi-decadal trend since the late 20th century. Beyond external forcing, Fogt et al. (2017) show a significant association of tropical SST variability with the summertime SAM trend since the mid-20th century in agreement with Lim et al. (2016), who, however, demonstrate that such a teleconnection between the summertime SAM and El Niño–Southern Oscillation (Annex IV.2.3), found in observations, is missing in many CMIP5 models.

On longer time scales, last millennium experiments from CMIP5 models fail to capture multicentennial variability evident in the reconstructions for the pre-industrial era (Abram et al., 2014; Dätwyler et al., 2018), which is also the case in those from available CMIP6 models (Figure 3.35). However, there is large uncertainty among reconstructions (Section 2.4.1.2). It is therefore unclear whether this disagreement reflects this observational uncertainty, whether forcings such as variations in the imposed insolation may be too weak, whether models are insufficiently sensitive to such variations, or whether internal variability including that associated with tropical Pacific variability is under-represented (Abram et al., 2014). The explanation could be a combination of all these factors. However, despite the aforementioned limitations of the reconstructions, Section 2.4.1.2 assesses that the recent positive trend in the SAM is likely unprecedented in at least the past millennium (medium confidence). CMIP5 and CMIP6 last-millennium simulations only capture the present anomalous state during the final decades of the simulations which are dominated by human influence; this state is also outside the range of simulated variability characteristic of pre-industrial times.

Figure 3.35 | Southern Annular Mode (SAM) indices in the last millennium. (a) Annual-mean SAM reconstructions by Abram et al. (2014) and Dätwyler et al. (2018). (b) The annual-mean SAM index defined by Gong and Wang (1999) in CMIP5 and CMIP6 last millennium simulations extended by historical simulations. All indices are normalized with respect to 1961–1990 means and standard deviations. Thin lines and thick lines show seven-year and 70-year moving averages, respectively. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

In summary, it is very likely that anthropogenic forcings have contributed to the observed trend of the summer SAM toward its positive phase since the 1970s. This assessment is supported by further model studies that confirm the human influence on the summertime SAM with improved models since AR5. While ozone depletion contributed to the trend from the 1970s to the 1990s (medium confidence), its influence has been small since 2000, leading to a weaker summertime SAM trend over 2000–2019 (medium confidence). Climate models reproduce the spatial structure of the summertime SAM observed since the late 1970s well (high confidence). CMIP6 models reproduce the spatiotemporal features and recent multi-decadal trend of the summertime SAM better than CMIP5 models (medium confidence). However, there is a large spread in the intensity of the SAM response to ozone and greenhouse gas changes in both CMIP5 and CMIP6 models (high confidence), which limits the confidence in the assessment of the ozone contribution to the observed trends. CMIP5 and CMIP6 models do not capture multicentennial variability of the SAM found in proxy reconstructions (low confidence). This confidence level reflects that it is unclear whether this is due to a model or an observational shortcoming.

The El Niño–Southern Oscillation (ENSO), which is generated via seasonally modulated interactions between the tropical Pacific ocean and atmosphere, influences severe weather, rainfall, river flow and agricultural production over large parts of the world (McPhaden et al., 2006). In fact, the remote climate influence of ENSO is so large that knowledge of its current phase and forecasts of its future phase largely underpin many seasonal rainfall and temperature forecasts worldwide (Annex IV.2.3).

AR5 noted that there have been clear improvements in the simulation of ENSO through previous generations of CMIP models (Flato et al., 2013), such that many CMIP5 models displayed behaviour that was qualitatively similar to that of the observed ENSO (Guilyardi et al., 2012). However, systematic errors were identified in the models’ representation of the tropical Pacific mean state and aspects of their interannual variability that affect quantitative comparisons. The AR5 assessment of ENSO concluded that the considerable observed inter-decadal modulations in ENSO amplitude and spatial pattern were largely consistent with unforced model simulations. Thus, there was low confidence in the role of a human-induced influence in these (Bindoff et al., 2013).

Observed ENSO amplitude, which is measured by the standard deviation of SST anomalies in a central equatorial Pacific region often referred to as the Nino 3.4 region, along with the lifecycle of events, are both reasonably well reproduced by most CMIP5 and CMIP6 models (Figure 3.36; Bellenger et al., 2014; Planton et al., 2021). The average CMIP5 model ENSO amplitude is slightly lower than that observed, while the average CMIP6 model ENSO amplitude is slightly higher than observed (Figure 3.36). The ENSO amplitude of the individual models, however, is highly variable across CMIP5 and CMIP6 models with many displaying either more or less variability than observed (Stevenson, 2012; Grose et al., 2020; Planton et al., 2021).

Figure 3.36 | Life cycle of (left) El Niño and (right) La Niña events in observations (black) and historical simulations from CMIP5 (blue; extended with RCP4.5) and CMIP6 (red). An event is detected when the December ENSO index value in year zero exceeds 0.75 times its standard deviation for 1951–2010. (a, b) Composites of the ENSO index (°C). The horizontal axis represents month relative to the reference December (the grey vertical bar), with numbers in parentheses indicating relative years. Shading and lines represent 5th–95th percentiles and multi-model ensemble means, respectively. (c, d) Mean durations (months) of El Niño and La Niña events defined as number of months in individual events for which the ENSO index exceeds 0.5 times its December standard deviation. Each dot represents an ensemble member from the model indicated on the vertical axis. The boxes and whiskers represent multi-model ensemble means, interquartile ranges and 5th and 95th percentiles of CMIP5 and CMIP6. The CMIP5 and CMIP6 multi-model ensemble means and observational values are indicated at the top right of each panel. The multi-model ensemble means and percentile values are evaluated after weighting individual members with the inverse of the ensemble size of the same model, so that individual models are equally weighted irrespective of their ensemble sizes. The ENSO index is defined as the SST anomaly averaged over the Niño 3.4 region (5°S–5°N, 170°W–120°W). All results are based on five-month running mean SST anomalies with triangular-weights after linear detrending. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

ENSO events are often synchronized to the seasonal cycle in the observations, as the associated SST anomalies tend to peak in boreal winter (November to January) and be at their weakest in the boreal spring (March to April) (Harrison and Larkin, 1998; Larkin and Harrison, 2002). The majority of CMIP5 and CMIP6 models broadly reproduce the seasonality of ENSO SST variability in the central equatorial Pacific (Taschetto et al., 2014; Abellán et al., 2017; Grose et al., 2020; Planton et al., 2021) (Figure 3.37). However, CMIP5 models, while displaying an improvement on CMIP3 models, appear to under-represent the magnitude of the seasonal variance modulation (Bellenger et al., 2014). This under-representation of seasonal variance modulation continues in CMIP6 models, which display no statistically significant difference in this behaviour when compared to CMIP5 models (Planton et al., 2021) (Figure 3.37).

Figure 3.37 | ENSO seasonality in observations (black) and historical simulations from CMIP5 (blue; extended with RCP4.5) and CMIP6 (red) for1951–2010. (a) Climatological standard deviation of the monthly ENSO index (SST anomaly averaged over the Niño 3.4 region; °C). Shading and lines represent 5th–95th percentiles and multi-model ensemble means, respectively. (b) Seasonality metric, which is defined for each model and each ensemble member as the ratio of the ENSO index climatological standard deviation in November–January (NDJ) to that in March–May (MAM). Each dot represents an ensemble member from the model indicated on the vertical axis. The boxes and whiskers represent the multi-model ensemble means, interquartile ranges and 5th and 95th percentiles of CMIP5 and CMIP6 individually. The CMIP5 and CMIP6 multi-model ensemble means and observational values are indicated at the top right of the panel. The multi-model ensemble means and percentile values are evaluated after weighting individual members with the inverse of the ensemble size of the same model, so that individual models are equally weighted irrespective of their ensemble sizes. All results are based on five-month running mean SST anomalies with triangular-weights after linear detrending. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Observations show strong multi-decadal modulation of ENSO variance throughout the 20th century, with the most recent period displaying larger variability while the mid-century displayed relatively low ENSO variability (Figure 2.36; Li et al., 2013; McGregor et al., 2013; Hope et al., 2017). As assessed in Section 2.4.2, ENSO amplitude since 1950 is higher than over the pre-industrial period from 1850 as far back as 1400 (medium confidence), but there is low confidence that it is higher than the variability over periods prior to 1400. This reported variance increase suggests that external forcing plays a role in the ENSO variance changes (Hope et al., 2017). However, large ensembles of single model or multiple model simulations do not find strong trends in ENSO variability over the historical period, suggesting that external forcing has not yet modulated ENSO variability with a magnitude that exceeds the range of internal variability (Hope et al., 2017; Maher et al., 2018b; Stevenson et al., 2019). This is consistent with the Chapter 2 assessment that there is no clear evidence for a recent sustained shift in ENSO beyond the range of variability on decadal to millennial time scales (Section 2.4.2). CMIP5 and CMIP6 models show a decrease in ENSO variance in the mid-Holocene (Brown et al., 2020), though not to the extent seen in paleo-proxy records (Emile-Geay et al., 2016). This suggests that both modelled and observed ENSO respond to changes in external forcing, but not necessarily in the same manner.

Most CMIP5 and CMIP6 models are found to represent the general structure of observed SST anomalies during ENSO events well (Kim and Yu, 2012; Taschetto et al., 2014; Brown et al., 2020; Grose et al., 2020). However, the majority of CMIP5 models display SST anomalies that: i) extend too far to the west (Taschetto et al., 2014; Capotondi et al., 2015); and ii) have meridional widths that are too narrow (Zhang and Jin, 2012) compared to the observations. CMIP6 models display a statistically significant improvement in the longitudinal representation of ENSO SST anomalies relative to CMIP5 models (Planton et al., 2021), however, systematic biases in the zonal extent and meridional width remain in CMIP6 models (Fasullo et al., 2020; Planton et al., 2021). The ENSO phase asymmetry, where observed strong El Niño events are larger and have a shorter duration than strong La Niña events (Ohba and Ueda, 2009; Frauen and Dommenget, 2010), is also under-represented in both CMIP5 and CMIP6 models (Zhang and Sun, 2014; Planton et al., 2021). In this instance, both CMIP5 and CMIP6 models typically display El Niño events that have a longer duration than those observed, La Niña events that have a similar duration to those observed, and there is very little asymmetry in the duration of El Niño and La Niña phases (Figure 3.36). Roberts et al. (2018) find an improvement in amplitude asymmetry in a HighResMIP model, but the under-representation remains.

The continuum of El Niño events are typically stratified into two types (often termed ‘flavours’), Central Pacific and East Pacific, where the name denotes the location of the events’ largest SST anomalies (Annex IV.2.3; Capotondi et al., 2015). As discussed in Section 2.4.2, the different types of events tend to produce distinct teleconnections and climatic impacts (e.g., Taschetto et al., 2020). The characteristics of El Niño events of these two flavours in CMIP5 were generally comparable to the observations (Taschetto et al., 2014). CMIP6 models, however, display a statistically significant improvement in the representation of this ENSO event-to-event SST anomaly diversity when compared with CMIP5 models (Planton et al., 2021). In addition to this ENSO event diversity, the short observational record also displays an increase in the number of the Central Pacific-type events in recent decades (Ashok et al., 2007; McPhaden et al., 2011), which has also been identified as unusual in the context of the last 500–800 years based on recent paleo-climatic reconstructions (Section 2.4.2; Y. Liu et al., 2017; Freund et al., 2019). However, the short observational record combined with observational (L’Heureux et al., 2013) and paleo-climatic reconstruction uncertainties preclude firm conclusions being made about the long-term changes in the occurrence of different El Niño event types. Initial analysis with a selected number of CMIP3 models suggested that there may be a forced component to this recent prominence of Central Pacific-type events (Yeh et al., 2009), but analysis since then suggests that this behaviour is (i) consistent with that expected from internal variability (Newman et al., 2011); and (ii) not apparent across the full CMIP5 ensemble of historical simulations (Taschetto et al., 2014). Analysis of single-model large ensembles suggests that changes to ENSO event type in response to historical radiative forcing are not significant (e.g., Stevenson et al., 2019). These same results, however, also suggest that multiple forcings can have significant influences on ENSO type and that the net response will depend on the accurate representation of the balance of these forcings (Stevenson et al., 2019).

The climatic effects of ENSO outside the tropical Pacific largely arise through atmospheric teleconnections that are induced by ENSO-driven changes in deep tropical atmospheric convection and heating (Yeh et al., 2018). The teleconnections to higher latitudes are forced by waves that propagate into the extratropics (Hoskins and Karoly, 1981) and respectively excite the Pacific-North American pattern (Horel and Wallace, 1981) and Pacific-South American pattern (Karoly, 1989; Irving and Simmonds, 2016) in the Northern and Southern Hemispheres. Given the influence of these teleconnections on climate and extremes around the globe, it is important to understand how well they are reproduced in CMIP models. What has also become clear is that spatial correlations of ENSO’s teleconnections calculated over relatively short periods (<100 years) may not be the most effective way to assess these relationships (Langenbrunner and Neelin, 2013; Perry et al., 2020). This is because the spatial patterns are significantly affected by internal atmospheric variability on relatively short time scales (<100 years; Batehup et al., 2015; Perry et al., 2020). However, looking at simplified metrics like the agreement in the sign of the teleconnections (Langenbrunner and Neelin, 2013), regional average teleconnection strength over land (Perry et al., 2020), or a combination of both (Power and Delage, 2018) provides a more robust depiction of the teleconnection representation. Examining sign agreement for the teleconnection patterns, ensembles of CMIP5 AMIP simulations display broad spatial regions with high sign agreement with the observations, suggesting that the model ensemble is producing useful information regarding the teleconnected precipitation signal (Langenbrunner and Neelin, 2013). Looking at regional averages of CMIP5 historical simulations, Power and Delage (2018) show that the average coupled model teleconnection pattern reproduces the sign of the observed teleconnections in the majority of the 25 regions analysed. The sign agreement between the observed teleconnection and the multi-model mean teleconnection remains strong in CMIP6 (18 out of 20 displayed regions; Figure 3.38), and the observed DJF (December–January–February) teleconnection strength falls within the modelled range in all of the displayed regions for temperature and precipitation. Note, however, that while there is broad agreement in ENSO teleconnections between CMIP6 models and observations during DJF (e.g., Fasullo et al., 2020), there are regions and seasons where the modelled teleconnection strength is outside the observed range (Chen et al., 2020).

Figure 3.38 | Model evaluation of ENSO teleconnection for near surface air temperature and precipitation in boreal winter (December–January–February). Teleconnections are identified by linear regression with the Niño 3.4 SST index based on Extended Reconstructed Sea Surface Temperature (ERSST) version 5 over the period 1958–2014. Maps show observed patterns for temperature from the Berkeley Earth dataset over land and from ERSST version 5 over ocean (°C, top) and for precipitation from GPCC over land (shading, mm day^–1) and GPCP worldwide (contours, period: 1979–2014). Distributions of regression coefficients (grey histograms) are provided for a subset of AR6 reference regions defined in Atlas.1.3 for temperature (top) and precipitation (bottom). All fields are linearly detrended prior to computation. Multi-model multi-member ensemble means are indicated by thick vertical black lines. Blue vertical lines show three observational estimates of temperature, based on Berkeley Earth, GISTEMP and CRUTS datasets, and two observational estimates of precipitation, based on GPCC and CRUTS datasets. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Most CMIP5 and CMIP6 models exhibit ENSO behaviour during the historical period that, to first order, is qualitatively similar to that of the observed ENSO. Many studies are now delving deeper into the models to understand if they are accurately producing the dynamics driving ENSO and its initiation (Jin et al., 2006; Bellenger et al., 2014; Vijayeta and Dommenget, 2018; Bayr et al., 2019; Planton et al., 2021). For both CMIP3 and CMIP5, diagnostics of ENSO event growth appear to show that the models, while producing ENSO variability that is qualitatively similar to that observed, do not represent the balance of the underlying dynamics well. The atmospheric Bjerknes feedback is too weak in the majority of models, while the surface heat flux feedback is also too weak in the majority of models. The former restricts event growth, while the latter restricts event damping, which when combined allow most models to produce variability in a range that is consistent with the observations (Bellenger et al., 2014; S.T. Kim et al., 2014; Vijayeta and Dommenget, 2018; Bayr et al., 2019). Analysis of ENSO representation in a subset of CMIP6 models by Planton et al. (2021) suggests that these issues remain.

To conclude, ENSO representation in CMIP5 models displayed a significant improvement from the representation of ENSO variability in CMIP3 models, which displayed much more intermodel spread in standard deviation, and stronger biennial periodicity (Guilyardi et al., 2012; Flato et al., 2013). In general, there has been no large step change in the representation of ENSO between CMIP5 and CMIP6, however, CMIP6 models appear to better represent some key ENSO characteristics (e.g., Brown et al., 2020; Planton et al., 2021). The instrumental record and paleo-proxy evidence through the Holocene all suggest that ENSO can display considerable modulations in amplitude, pattern and period (see also (Section 2.4.2). For the period since 1850, there is no clear evidence for a sustained shift in ENSO index beyond the range of internal variability. However, paleo-proxy evidence indicates with medium confidence that ENSO variability since 1950 is greater than at any time between 1400 and 1850 (Section 2.4.2). Coupled models display large changes of ENSO behaviour in the absence of external forcing changes, and little-to-no variance sensitivity to historical anthropogenic forcing. Thus, there is low confidence that anthropogenic forcing has led to the changes of ENSO variability inferred from paleo-proxy evidence.

Chapter 2 reports low confidence that the apparent change from East Pacific- to Central Pacific-type El Niño events that occurred in the last 20–30 years was representative of a long term change. While some climate models do suggest external forcing may affect the El Niño event type, most climate models suggest that what has been observed is well within the range of natural variability. Thus, there is low confidence that anthropogenic forcing has had an influence on the observed changes in El Niño event type.

The Indian Ocean Basin (IOB) and Dipole (IOD) modes are the two leading modes of interannual SST variability over the tropical Indian Ocean, featuring basin-wide warming/cooling and an east–west dipole of SST anomalies, respectively (Annex IV.2.4). The IOD mode is anchored to boreal summer to autumn by the air–sea feedback, and often develops in concert with ENSO. Driven by matured ENSO events, the IOB mode peaks in boreal spring and often persists into the subsequent summer. Similar patterns of Indian Ocean SST variability also dominate its decadal and longer time scale variability (Han et al., 2014b).

AR5 concluded that models show high and medium performance in reproducing the IOB and IOD modes, respectively (medium confidence), with difficulty in reproducing the persistence of the IOB and the pattern and magnitude of the IOD (Flato et al., 2013). There was low confidence that changes in the IOD were detectable or attributable to human influence (Bindoff et al., 2013).

Since AR5, CMIP5 model representation of these modes has been analysed in detail, finding that most of the models qualitatively reproduce the spatial and seasonal features of the IOB and IOD modes (Chu et al., 2014; Liu et al., 2014; W. Tao et al., 2016). Improvements in simulating the IOB mode since CMIP3 have been identified in reduced multi-model mean biases and inter-model spread (W. Tao et al., 2016). CMIP5 models overall capture the transition from the IOD to IOB modes during an ENSO event (W. Tao et al., 2016). The IOB mode is forced in part through a cross-equatorial wind–evaporation–SST feedback triggered by ENSO-forced anomalous ocean Rossby waves that propagate to the shallow climatological thermocline dome in the tropical south-western Indian Ocean (Du et al., 2009). Consistently, models with a deeper climatological thermocline dome produce a weaker and less persistent IOB mode (G. Li et al., 2015a; Zheng et al., 2016). The deep thermocline bias remains in the ensemble mean of CMIP5 models due to a common surface easterly wind bias over the equatorial Indian Ocean (Lee et al., 2013) associated with weaker South Asian summer monsoon circulation (G. Li et al., 2015b). However, the influence of this systematic bias may be compensated by other biases, resulting in a realistic IOB magnitude (W. Tao et al., 2016). Halder et al. (2021) found that CMIP6 models reproduce the IOB mode reasonably well, but did not evaluate the progress since CMIP5.

By contrast, the IOD magnitude is overestimated by CMIP5 models on average, though with noticeable improvements from CMIP3 models (Liu et al., 2014). The overestimation of the IOD magnitude remains in most of 34 CMIP6 models examined in McKenna et al. (2020) with worsening on average in July and August. A too steep climatological thermocline slope along the equator due to the surface easterly wind bias in boreal summer and autumn contributes to this IOD magnitude bias through an excessively strong Bjerknes feedback in CMIP5 (Liu et al., 2014; G. Li et al., 2015b; Hirons and Turner, 2018). The surface easterly bias and associated east–west SST gradient bias are not improved in CMIP6 (Long et al., 2020; Section 3.5.1.2.3), suggesting that the thermocline bias also remains. McKenna et al. (2020) additionally find degradation in the positive-negative asymmetry of the IOD but an improvement in IOD frequency in a subset of CMIP6 models compared to CMIP5. In terms of teleconnections, the equatorial surface easterly wind bias also affects the IOD-associated moisture transport anomalies toward tropical eastern Africa (Hirons and Turner, 2018) where the IOD is associated with strong precipitation anomalies in boreal autumn (Annex IV.2.4). CMIP5 and CMIP6 models capture the IOD teleconnection to Southern and Central Australian precipitation although it is weaker on average than observed, with no clear improvements from CMIP5 to CMIP6 (Grose et al., 2020). Strong IOD events could also influence the Northern Hemisphere extratropical circulation in winter and in particular the NAM (Section 3.7.1), based on interference between forced Rossby waves emerging from the Indian Ocean and climatological stationary waves (Fletcher and Cassou, 2015). The record positive phase of the NAO/NAM in winter 2019–2020 assessed over the instrumental era has been accordingly linked to the record IOD event of autumn 2019 (Hardiman et al., 2020), which has been associated with the devastating record fire season in Australia (Wang and Cai, 2020).

The observed Indian Ocean basin-average SST increase on multi-decadal and centennial time scales is well represented by CMIP5 historical simulations, and has been attributed to the effects of greenhouse gases offset in part by the effects of anthropogenic aerosols mainly through aerosol-cloud interactions (Dong and Zhou, 2014; Dong et al., 2014b). The observed SST trend is larger in the western than eastern tropical Indian Ocean, which leads to an apparent upward trend of the IOD index, but this trend is statistically insignificant (Section 2.4.3). CMIP5 models capture this warming pattern, which may be associated with Walker circulation weakening over the Indian Ocean due to greenhouse gas forcing (Dong and Zhou, 2014). However, strong internal decadal IOD-like variability and observational uncertainty preclude attribution (Cai et al., 2013; Han et al., 2014b; Gopika et al., 2020). Such a positive IOD-like change in equatorial zonal SST gradient suggests an increase in the frequency of extreme positive events (Cai et al., 2014) and skewness (Cowan et al., 2015) of the IOD mode. While there is some evidence of an increase in frequency of positive IOD events during the second half of the 20th century, the current level of IOD variability is not unprecedented in a proxy reconstruction for the last millennium (Section 2.4.3; Abram et al., 2020). Besides, the IOD magnitude in the late 20th century is not significantly different between CMIP5 simulations forced by historical and natural-only forcings, though this conclusion is based on only five selected ensemble members that realistically reproduce statistical features of the IOD (Blau and Ha, 2020). While selected CMIP5 models show weakening (Thielke and Mölg, 2019) and seasonality changes (Blau and Ha, 2020) in IOD-induced rainfall anomalies in tropical eastern Africa, no comparison with observational records has been made. Likewise, while a strengthening tendency of the ENSO-IOB mode correlation and resultant intensification of the IOB mode are found in historical or future simulations in selected CMIP5 models (Hu et al., 2014; Tao et al., 2015), such a change has not been detected in observational records.

After linear detrending, Pacific Decadal Variability (PDV; Annex IV.2.6; Section 3.7.6) has been suggested as a driver of decadal to multi-decadal variations in the IOB mode (Dong et al., 2016). However, correlation between the PDV and a decadal IOB index, defined from linearly detrended SST, changed from positive to negative during the 1980s (Han et al., 2014a). The increase in anthropogenic forcing and recovery from the eruptions of El Chichón in 1982 and Pinatubo in 1991 may have overwhelmed the PDV influence, and explain this change (Dong and McPhaden, 2017; L. Zhang et al., 2018a). However, the low statistical degrees of freedom hamper clear detection of human influence in this correlation change.

To summarize, there is medium confidence that changes in the interannual IOD variability in the late 20th century inferred from observations and proxy records are within the range of internal variability. There is no evidence of anthropogenic influence on the interannual IOB. On decadal- to multi-decadal time scales, there is low confidence that human influence has caused a reversal of the correlation between PDV and decadal variations in the IOB mode. The low confidence in this assessment is due to the short observational record, limited number of models used for the attribution, lack of model evaluation of the decadal IOB mode, and uncertainty in the contribution from volcanic aerosols. Nevertheless, CMIP5 models have medium overall performance in reproducing both the interannual IOB and IOD modes, with an apparently good performance in reproducing the IOB magnitude arising from compensation of biases in the formation process, and overly high IOD magnitude due to the mean state bias (high confidence). There is no clear improvement in the simulation of the IOD from CMIP5 to CMIP6 models, though there is only medium confidence in this assessment, since only a subset of CMIP6 models have been examined. There is no evidence for performance changes in simulating the IOB from CMIP5 to CMIP6 models.

The Atlantic Zonal Mode (AZM), often referred to as the Atlantic Equatorial Mode or Atlantic Niño, and the Atlantic Meridional Mode (AMM) are the two leading basin-wide patterns of interannual to decadal variability in the tropical Atlantic. Akin to ENSO in the Pacific, the term Atlantic Niño is broadly used to refer to years when the SSTs in the tropical eastern Atlantic basin along the cold tongue are significantly warmer than the climatological average. The AMM is characterized by anomalous cross-equatorial gradients in SST. Both modes are associated with altered strength of the Inter-tropical Convergence Zone (ITCZ) and/or latitudinal shifts in the ITCZ, which locally affect African and American monsoon systems and remotely affect tropical Pacific and Indian Ocean variability through inter-basins teleconnections. A detailed description of both AZM and AMM, as well as their associated teleconnection over land, is given in Annex IV.2.5

AR5 mentioned the considerable difficulty in simulating both Atlantic Niño and AMM despite some improvements in CMIP5 for some models (Flato et al., 2013). Severe biases in mean state and variance for both SST and atmospheric dynamics including rainfall (e.g., a double ITCZ) as well as teleconnections were reported. The AR5 highlighted the complexity of the tropical Atlantic biases, which were explained by multiple factors both in the ocean and atmosphere.

Since AR5, further analysis of the major persistent biases in models has been reported (Xu et al., 2014; Jouanno et al., 2017; Y. Yang et al., 2017; Dippe et al., 2018; Lübbecke et al., 2018; Voldoire et al., 2019a). Errors in equatorial and basin wide trade winds, cloud cover and ocean vertical mixing and dynamics both locally and in remote subtropical upwelling regions, key thermodynamic ocean–atmosphere feedbacks, and tropical land–atmosphere interaction have been shown to be detrimental to the representation of both the Atlantic Niño and AMM leading to poor teleconnectivity over land (Rodríguez-Fonseca et al., 2015; Wainwright et al., 2019) and between tropical basins (Ott et al., 2015).

Despite some improvements (Richter et al., 2014; Nnamchi et al., 2015), biases in the mean state are so large that the mean east–west temperature gradient at the equator along the thermocline remains opposite to observed in two thirds of the CMIP5 models (Section 3.5.1.2.2), which clearly affects the simulation of the Atlantic Niño and associated dynamics (Muñoz et al., 2012; Ding et al., 2015; Deppenmeier et al., 2016). The interhemispheric SST gradient is also systematically underestimated in models, with a too cold mean state in the northern part of the tropical Atlantic ocean and too warm conditions in the South Atlantic basin. The seasonality is poorly reproduced and the wind–SST coupling is weaker than observed so that altogether, and despite AMM-like variability in 20th century climate simulations, AMM is not the dominant Atlantic mode in all CMIP5 models (Liu et al., 2013; Amaya et al., 2017). These biases in mean state translate into biases in modelling the mean ITCZ (Flato et al., 2013). Similar biases were found in experiments using CMIP5 models but with different climate background states, such as Last Glacial Maximum, mid-Holocene and future scenario simulations (Brierley and Wainer, 2018). Analyses of CMIP6 show encouraging results in the representation of Atlantic Niño and AMM modes of variability in terms of amplitude and seasonality. Some models now display reduced biases in the spatial structure of the modes and related explained variance but persistent errors still remain on average in the timing of the modes and in the coupled nature of the modes, that is, the strength of the link between ocean (SST, mixed layer depth) and atmospheric (wind) anomalies (Richter and Tokinaga, 2020), as well as in the Atlantic Ocean equatorial east–west temperature gradient (Section 3.5.1.2.2, Figure 3.24).

There are some recent indications that increasing model resolution both vertically and horizontally, in the ocean and atmospheric component (Richter, 2015; Small et al., 2015; Harlaß et al., 2018), could partly alleviate some tropical Atlantic biases in mean state (Section 3.5.1.2.2), seasonality, interannual- to decadal-variability and associated teleconnectivity over land, such as with the West African monsoon (Steinig et al., 2018). Results from CMIP6 tend to confirm that increasing resolution is not the unique way to address the biases in the tropical Atlantic (Richter and Tokinaga, 2020). For instance, the inclusion of a stochastic physics scheme has a nearly equivalent effect in the improvement of the mean number and the strength distribution of tropical Atlantic cyclones (Vidale et al., 2021).

(Section 2.4.4 assess that there is low confidence in any sustained changes to the AZM and AMM variability in instrumental observations. Moreover, any attribution of possible human influence on the Atlantic modes and associated teleconnections is limited by the poor fidelity of CMIP5 and CMIP6 models in reproducing the mean tropical Atlantic climate, its seasonality and variability, despite hints of some improvement in CMIP6, as well as other sources of uncertainties related to limited process understanding in the observations (Foltz et al., 2019), the response of the tropical Atlantic climate to anthropogenic aerosol forcing (Booth et al., 2012; Zhang et al., 2013a) and the presence of strong multi-decadal fluctuations related to AMV (Section 3.7.7) and cross-tropical basin interactions (Martín-Rey et al., 2018; Cai et al., 2019). The fact that most models poorly represent the climatology and variability of the tropical Atlantic combined with the short observational record makes it difficult to place the recent observed changes in the context of internal multi-annual variability versus anthropogenic forcing.

In summary, based on CMIP5 and CMIP6 results, there is no robust evidence that the observed changes in either the Atlantic Niño or AMM modes and associated teleconnections over the second half of the 20th century are beyond the range of internal variability or have been influenced by natural or anthropogenic forcing. Considering the physical processes responsible for model biases in these modes, increasing resolution in both ocean and atmosphere components may be an opportunity for progress in the simulation of the tropical Atlantic changes as evidenced by some individual model studies (Roberts et al., 2018), but this needs confirmation from a multi-model perspective.

Pacific Decadal Variability (PDV) is the generic term for the modes of variability in the Pacific Ocean that vary on decadal to inter-decadal time scales. PDV and its related teleconnections encompass the Pacific Decadal Oscillation (PDO; Mantua et al., 1997; Zhang et al., 1997; Mantua and Hare, 2002), and an anomalous SST pattern in the North Pacific, as well as a broader structure associated with Pacific-wide SSTs termed the Inter-decadal Pacific Oscillation (IPO; Power et al., 1999; Folland et al., 2002; Henley et al., 2015). Since the PDO and IPO indices are highly correlated, this section assesses them together as the PDV (Annex IV.2.6).

AR5 mentioned an overall limited level of evidence for both CMIP3 and CMIP5 evaluation of the Pacific modes at inter-decadal time scales, leading to low confidence statements about the models’ performance in reproducing PDV (Flato et al., 2013) and similarly low confidence in the attribution of observed PDV changes to human influence (Bindoff et al., 2013).

The implication of PDV in the observed slowdown of the GMST warming rate in the early 2000s (Cross-Chapter Box 3.1) has triggered considerable research on decadal climate variability and predictability since AR5 (Meehl et al., 2013, 2016b; England et al., 2014; Dai et al., 2015; Steinman et al., 2015; Kosaka and Xie, 2016; Cassou et al., 2018). Many studies find that the broad spatial characteristics of PDV are reasonably well represented in unforced climate models (Newman et al., 2016; Henley, 2017) and in historical simulations in CMIP5 and CMIP6 (Figure 3.39), although there is sensitivity to the methodology used to remove the externally-forced component of the SST (Bonfils and Santer, 2011; Xu and Hu, 2018). Compared with CMIP3 models, CMIP5 models exhibit overall slightly better performance in reproducing PDV and associated teleconnections (Polade et al., 2013; Joshi and Kucharski, 2017), and also smaller inter-model spread (Lyu et al., 2016). CMIP6 models on average show slightly improved reproduction of the PDV spatial structure than CMIP5 models (Figure 3.39a–c; Fasullo et al., 2020). SST anomalies in the subtropical South Pacific lobe are, however, too weak relative to the equatorial and North Pacific lobes in CMIP5 pre-industrial control and historical simulations (Henley et al., 2017), a bias that remains in CMIP6 (Figure 3.39b).

Biases in the PDV temporal properties and amplitude are present in CMIP5 (Cheung et al., 2017; Henley, 2017). While model evaluation is severely hampered by short observational records and incomplete observational coverage before satellite measurements started, the duration of PDV phases appears to be shorter in coupled models than in observations, and correspondingly the ratio of decadal to interannual variance is underestimated (Figure 3.39e,f; Henley et al., 2017). This apparent bias may be associated with overly biennial behaviour of Pacific trade wind variability and related ENSO activity, leaving too weak variability on decadal time scales (Kociuba and Power, 2015). ENSO influence on the extratropical North Pacific Ocean at decadal time scales is also very diverse among both CMIP3 and CMIP5 models, being controlled by multiple factors (Nidheesh et al., 2017). In terms of amplitude, the variance of the PDV index after decadal filtering is significantly weaker in the concatenated CMIP5 ensemble than the three observational estimates used in Figure 3.39e (p<0.1 with an F-test). Consequently, the observed PDV fluctuations over the historical period often lie in the tails of the model distributions (Figure 3.39e,f). Even if one cannot rule out that the observed PDV over the instrumental era represents an exceptional period of variability, it is plausible that the tendency of the CMIP5 models to systematically underestimate the low frequency variance is due to an incomplete representation of decadal-scale mechanisms in these models. This situation is slightly improved in CMIP6 historical simulations but remains a concern (Fasullo et al., 2020). The results of McGregor et al. (2018) suggest that the under-representation of the variability stems from Atlantic mean SST biases (Section 3.5.1.2.2) through inter-basin coupling.

While PDV is primarily understood as an internal mode of variability (Si and Hu, 2017), there are some indications that anthropogenically induced SST changes project onto PDV and have contributed to its past evolution (Bonfils and Santer, 2011; Dong et al., 2014a; Boo et al., 2015; Xu and Hu, 2018). However, the level of evidence is limited because of the difficulty in correctly separating internal versus externally forced components of the observed SST variations, and because it is unclear whether the dynamics of the PDV are operative in this forced SST change pattern. Over the last two to three decades which encompass the period of slower GMST increase (Cross-Chapter Box 3.1), Smith et al. (2016) found that anthropogenic aerosols have driven part of the PDV change toward its negative phase. A similar result is shown in Takahashi and Watanabe (2016) who found intensification of the Pacific Walker circulation in response to aerosol forcing (Section 3.3.3.1.2). Indeed, CMIP6 models simulate a negative PDV trend since the 1980s (Figure 3.39f), which is much weaker than internal variability. However, a response to anthropogenic aerosols is not robustly identified in a large ensemble of a model (Oudar et al., 2018), across CMIP5 models (Hua et al., 2018), or in idealized model simulations (Kuntz and Schrag, 2016). Alternatively, inter-basin teleconnections associated with the warming of the North Atlantic Ocean related to the mid-1990s phase shift of the AMV (McGregor et al., 2014; Chikamoto et al., 2016; Kucharski et al., 2016; X. Li et al., 2016a; Ruprich-Robert et al., 2017), and also warming in the Indian Ocean (Luo et al., 2012; Mochizuki et al., 2016), could have favoured a PDV transition to its negative phase in the 2000s. Considering the possible influence of external forcing on Indian Ocean decadal variability (Section 3.7.4) and AMV (Section 3.7.7), any such human influence on PDV would be indirect through changes in these ocean basins, and then imported to the Pacific via inter-basin coupling. However, this human influence on AMV, and how consistently such inter-basin processes affect PDV phase shifts, are uncertain. Other modelling studies find that anthropogenic aerosols can influence the PDV (Verma et al., 2019; Amiri-Farahani et al., 2020; Dow et al., 2020). It is however unclear whether and how much those forcings contributed to the observed variations of PDV. In CMIP6 models, the temporal correlation of the multi-model ensemble mean PDV index with its observational counterpart is insignificant and negligible (Figure 3.39f), suggesting that any externally-driven component in historical PDV variations was weak. Lastly, the multi-model ensemble mean computed from CMIP6 historical simulations shows slightly stronger variation than the CMIP5 counterpart, suggesting a greater simulated influence from external forcings in CMIP6. Still, the fraction of the forced signal to the total PDV is very low (Figure 3.39f), in contrast to AMV (Section 3.7.7). Consistently, Liguori et al. (2020) estimate that the variance fraction of the externally-driven to total PDV is up to only 15% in a multi-model large ensemble of historical simulations. These findings support an assessment that PDV is mostly driven by internal variability since the pre-industrial era. The sensitivity of ensemble-mean PDV trends to the ensemble size (Oudar et al., 2018), and the dominance of the ensemble spread over the ensemble mean in the 60-year trend of the equatorial Pacific zonal SST gradient in large ensemble simulations (Watanabe et al., 2021), also support this statement.

Figure 3.39 | Model evaluation of the Pacific Decadal Variability (PDV). (a, b) Sea surface temperature (SST) anomalies (°C) regressed onto the Tripole Index (TPI; Henley et al., 2015) for 1900–2014 in (a) ERSST version 5 and (b) CMIP6 multi-model ensemble (MME) mean composite obtained by weighting ensemble members by the inverse of the model ensemble size. A 10-year low-pass filter was applied beforehand. Cross marks in (a) represent regions where the anomalies are not significant at the 10% level based on a t-test. Diagonal lines in (b) indicate regions where less than 80% of the runs agree in sign. (c) A Taylor diagram summarizing the representation of the PDV pattern in CMIP5 (each ensemble member is shown as a cross in light blue, and the weighted multi-model mean as a dot in dark blue), CMIP6 (each ensemble member is shown as a cross in red, and the weighted multi-model mean as a dot in orange) and observations over 40°S–60°N and 110°E–70°W. The reference pattern is taken from ERSST version 5 and black dots indicate other observational products: Hadley Centre Sea Ice and Sea Surface Temperature data set version 1 (HadISST version 1) and Centennial in situ Observation-Based Estimates of Sea Surface Temperature version 2 (COBE-SST2). (d) Autocorrelation of unfiltered annual TPI at lag one year and 10-year low-pass filtered TPI at lag 10 years for observations over 1900–2014 (horizontal lines), 115-year chunks of pre-industrial control simulations (open boxes) and individual historical simulations over 1900–2014 (filled boxes) from CMIP5 (blue) and CMIP6 (red). (e) As in (d), but showing standard deviation of the unfiltered and filtered TPI (°C). Boxes and whiskers show weighted multi-model means, interquartile ranges and 5th and 95th percentiles. (f) Time series of the 10-year low-pass filtered TPI (°C) in ERSST version 5, HadISST version 1 and COBE-SST2 observational estimates (black) and CMIP5 and CMIP6 historical simulations. The thick red and light blue lines are the weighted multi-model mean for the historical simulations in CMIP5 and CMIP6, respectively, and the envelopes represent the 5th–95th percentile ranges across ensemble members. The 5–95% confidence interval for the CMIP6 multi-model mean is given in thin dashed lines. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

In CMIP5 last millennium simulations, there is no consistency in temporal variations of PDV across the ensemble (Fleming and Anchukaitis, 2016). This supports the notion that PDV is internal in nature. However, this issue remains difficult to assess because paleoclimate reconstructions of PDV have too poor a level of agreement for a rigorous model evaluation in past millennia (Henley, 2017).

To conclude, there is high confidence that internal variability has been the main driver of the PDV since pre-industrial times, despite some modelling evidence for potential external influence. This assessment is supported by studies based on large ensemble simulations that found the dominance of internally-driven PDV, and the CMIP6-based assessment (Figure 3.39). As such, PDV is an important driver of decadal internal climate variability which limits detection of human influence on various aspects of decadal climate change on global to regional scales (high confidence). Model evaluation of PDV is hampered by short observational records, spatial incompleteness of observations before the satellite observation era, and poor agreement among paleoclimate reconstructions. Despite the limitations of these model-observation comparisons, CMIP5 models, on average, simulate broadly realistic spatial structures of the PDV, but with a clear bias in the South Pacific (medium confidence). CMIP5 models also very likely underestimate PDV magnitude. CMIP6 models tend to show better overall performance in spatial structure and magnitude of PDV, but there is low confidence in this assessment due to the lack of literature.

Atlantic Multi-decadal Variability (AMV) refers to a climate mode representing basin-wide multi-decadal fluctuations in surface temperatures in the North Atlantic (Figure 3.40a,f), with teleconnections particularly pronounced over the adjacent continents and the Arctic. The AMV phenomenon is usually assessed through SST anomalies averaged over the entire North Atlantic basin, hereafter the AMV index, but it is associated with many physical processes including three-dimensional ocean circulation, such as AMOC fluctuations (Section 3.5.4.1), gyre adjustments, and salt and heat transport in the entire North Atlantic and subarctic Atlantic basins. The AMV, together with the PDV, has been shown to have modulated GSAT on multi-decadal time scales since pre-industrial times (Cross-Chapter Box 3.1; T. Wu et al., 2019a; Li et al., 2020). A detailed description of the AMV as well as its associated teleconnection over land is given in Annex IV.2.7.

Figure 3.40 | Model evaluation of Atlantic Multi-decadal Variability (AMV). (a, b) Sea surface temperature (SST) anomalies (°C) regressed onto the AMV index defined as the 10-year low-pass filtered North Atlantic (0°–60°N, 80°W–0°E) area-weighted SST* anomalies over 1900–2014 in (a) ERSST version 5 and (b) the CMIP6 multi-model ensemble (MME) mean composite obtained by weighting ensemble members by the inverse of each model’s ensemble size. The asterisk denotes that the global mean SST anomaly has been removed at each time step of the computation. Cross marks in (a) represent regions where the anomalies are not significant at the 10% level based on a t-test. Diagonal lines in (b) show regions where less than 80% of the runs agree in sign. (c) A Taylor diagram summarizing the representation of the AMV pattern in CMIP5 (each ensemble member is shown as a cross in light blue, and the weighted multi-model mean is shown as a dot in dark blue), CMIP6 (each ensemble member is shown as a cross in red, and the weighted multi-model mean is shown as a dot in orange) and observations over [0°–60°N, 80°W–0°E]. The reference pattern is taken from ERSST version 5 and black dots indicate other observational products (HadISST version 1 and COBE-SST2). (d) Autocorrelation of unfiltered annual AMV index at lag one year and 10-year low-pass filtered AMV index at lag 10 years for observations over 1900–2014 (horizontal lines), 115-year chunks of pre-industrial control simulations (open boxes) and individual historical simulations over 1900–2014 (filled boxes) from CMIP5 (blue) and CMIP6 (red). (e) As in (d), but showing standard deviation of the unfiltered and filtered AMV indices (°C). Boxes and whiskers show the weighted multi-model means, interquartile ranges and 5th and 95th percentiles. (f) Time series of the AMV index (°C) in ERSST version 5, HadISST version 1 and COBE-SST2 observational estimates (black) and CMIP5 and CMIP6 historical simulations. The thick red and light blue lines are the weighted multi-model mean for the historical simulations in CMIP5 and CMIP6, respectively, and the envelopes represent the 5th–95th percentile ranges obtained from all ensemble members. The 5–95% confidence interval for the CMIP6 multi-model mean is shown by the thin dashed line. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

AR5 assessed, based on climate models, that the AMV was primarily internally-driven alongside some contribution from external forcings (mainly anthropogenic aerosols) over the late 20th century (Bindoff et al., 2013; Flato et al., 2013). But AR5 also concluded that models show medium performance in reproducing the observed AMV, with difficulties in simulating the time scale, the spatial structure and the coherency between all the physical processes involved (Flato et al., 2013).

Climate models analysed since AR5 continue to simulate AMV-like variability as part of their internal variability. This statement is mostly based on CMIP5 pre-industrial control and historical simulations (Wouters et al., 2012; Schmith et al., 2014; Menary et al., 2015; Ruprich-Robert and Cassou, 2015; Brown et al., 2016b; Chen et al., 2016; Kim et al., 2018a) and is also true for the CMIP6 models (Menary et al., 2018; Voldoire et al., 2019b). Models also continue to support links to a wide array of remote climate influences through atmospheric teleconnections (Martin et al., 2014; Ruprich-Robert et al., 2017, 2018; Monerie et al., 2019; Qasmi et al., 2020; Ruggieri et al., 2021). Even if debate remains (Clement et al., 2015; Cane et al., 2017; Mann et al., 2020), there is now stronger evidence for a crucial role of oceanic dynamics in internal AMV that is primarily linked to the AMOC and its interplay with the NAO (Zhang et al., 2013a; Müller et al., 2015; O’Reilly et al., 2016b, 2019a; Delworth et al., 2017; Zhang, 2017; Sun et al., 2019; Kim et al., 2020). However, considerable diversity in the spatio-temporal properties of the simulated AMV is found in both pre-industrial control and historical CMIP5 experiments (Zhang and Wang, 2013; Wills et al., 2019). Such model diversity is presumably associated with the wide range of coupled processes associated with AMV (Baker et al., 2017; Woollings et al., 2018a) including large-scale atmospheric teleconnections and regional feedbacks relating to tropical clouds, Arctic sea ice in the subarctic basins and Saharan dust, whose relative importance and interactions across time scales are specific to each model (Martin et al., 2014; Brown et al., 2016b).

Additional studies since AR5 corroborate that CMIP5-era models tend to underestimate many aspects of observed AMV and its SST fingerprint. On average, the duration of modelled AMV episodes is too short, the magnitude of AMV is too weak and its basin-wide SST spatial structure is limited by the poor representation of the link between the tropical North Atlantic and the subpolar North Atlantic/Nordic seas (Martin et al., 2014; Qasmi et al., 2017). Such mismatches between observed and simulated AMV (Figure 3.40c–e) have been associated with intrinsic model biases in both mean state (Menary et al., 2015; Drews and Greatbatch, 2016) and variability in the ocean and overlying atmosphere. For instance, compared to available observational data CMIP5 models underestimate the ratio of decadal to interannual variability of the main drivers of AMV, namely the AMOC, NAO and related North Atlantic jet variations (Section 3.7.1; Bracegirdle et al., 2018; Kim et al., 2018b; Simpson et al., 2018; Yan et al., 2018), which has strong implications for the simulated temporal statistics of AMV, AMV-induced teleconnections (Ault et al., 2012; Menary et al., 2015) and AMV predictability.

The increase of AMV variance in CMIP6 models (stronger magnitude and longer duration) seems to be explained by the enhanced variability in the subpolar North Atlantic SST (Figure 3.40b,c), which is particularly pronounced in some models, associated with greater variability in the AMOC (Section 3.5.4.1; Voldoire et al., 2019a; Boucher et al., 2020) and greater GMST multi-decadal variability (Section 3.3.1 and Figure 3.40c–f; Voldoire et al., 2019b; Parsons et al., 2020). The decadal variance in SST in the subpolar North Atlantic seems now to be slightly overestimated in CMIP6 compared to observational estimates, while the AMV-related tropical SST anomalies remain weaker in line with CMIP5 (Figure 3.40b). The mechanisms producing the tropical-extratropical relationship at decadal time scales remain poorly understood despite stronger evidence since AR5 for the importance of the subpolar gyre SST anomalies in generating tropical changes through atmospheric teleconnection (Caron et al., 2015; Ruprich-Robert et al., 2017; Kim et al., 2020). Significant discrepancies remain in the simulated AMV spatial pattern when historical simulations are compared to multivariate observations (Yan et al., 2018; Robson et al., 2020).

There is additional evidence since AR5 that external forcing has been playing an important role in shaping the timing and intensity of the observed AMV since pre-industrial times (Bellomo et al., 2018; Andrews et al., 2020). The time synchronisation between observed and multi-model mean AMV SST indices is significant in both CMIP5 and CMIP6 historical simulations, while the explained variance of the forced response in CMIP6 appears stronger (Figure 3.40d–f). The competition between greenhouse gas warming and anthropogenic sulphate aerosol cooling has been proposed to be particularly important over the latter half of the 20th century (Booth et al., 2012; Steinman et al., 2015; Murphy et al., 2017; Undorf et al., 2018a; Haustein et al., 2019). The latest observed AMV shift from the cold to the warm phase in the mid-1990s at the surface ocean is well captured in the CMIP6 forced component and may be associated with the lagged response to increased AMOC due to strong anthropogenic aerosol forcing over 1955–1985 (Menary et al., 2020) in combination with the rapid response through surface flux processes to declining aerosol forcing and increasing greenhouse gas influence since then. However, natural forcings may have also played a significant role. For instance, volcanic forcing has been shown to contribute in part to the cold phases of the AMV-related SST anomalies observed in the 20th century (Terray, 2012; Bellucci et al., 2017; Swingedouw et al., 2017; Birkel et al., 2018). Over the last millennium, natural forcings including major volcanic eruptions and fluctuations in solar activity are thought to have driven a larger fraction of the multi-decadal variations in the AMV than in the industrial era, with some interplay with internal processes (Otterå et al., 2010; Knudsen et al., 2014; Moffa-Sánchez et al., 2014; J. Wang et al., 2017; Malik et al., 2018; Mann et al., 2021), but other studies question the role of natural forcings over this period (Zanchettin et al., 2014; Lapointe et al., 2020).

Model evaluation of the AMV phenomenon remains difficult because of short observational records (especially of detailed process-based observations), the lack of stationarity in the variance, spatial patterns and frequency of the AMV assessed from modelled SST (Qasmi et al., 2017), difficulties in estimating the forced signals in both historical simulations and observations (Tandon and Kushner, 2015), and because of probable interplay between internally and externally-driven processes (Watanabe and Tatebe, 2019). Furthermore, models simulate a large range of historical anthropogenic aerosol forcing (Smith et al., 2020) and questions often referred to as signal-to-noise paradox have been raised concerning the models’ ability to correctly simulate the magnitude of the response of AMV-related atmospheric circulation phenomena, such as the NAO (Section 3.7.1), to both internally and externally generated changes (Scaife and Smith, 2018). Related methodological and epistemological uncertainties also call into question the relevance of the traditional basin-average SST index to assessing the AMV phenomenon (Zanchettin et al., 2014; Frajka-Williams et al., 2017; Haustein et al., 2019; Wills et al., 2019).

To summarize, results from CMIP5 and CMIP6 models together with new statistical techniques to evaluate the forced component of modelled and observed AMV, provide robust evidence that external forcings have modulated AMV over the historical period. In particular, anthropogenic and volcanic aerosols are thought to have played a role in the timing and intensity of the negative (cold) phase of AMV recorded from the mid-1960s to mid-1990s and subsequent warming (medium confidence). However, there is low confidence in the estimated magnitude of the human influence. The limited level of confidence is primarily explained by difficulties in accurately evaluating model performance in simulating AMV. The evaluation is severely hampered by short instrumental records but also, equally importantly, by the lack of detailed and coherent long-term process-based observations (for example of the AMOC, aerosol optical depth, surface fluxes and cloud changes), which limit our process understanding. In addition, studies often rely solely on simplistic SST indices that may be hard to interpret (Zhang et al., 2016) and may mask critical physical inconsistencies in simulations of the AMV compared to observations (Zhang, 2017).

The AR5 concluded that human influence on the climate system is clear (IPCC, 2013), based on observed increasing greenhouse gas concentrations in the atmosphere, positive radiative forcing, observed warming, and physical understanding of the climate system. The AR5 also assessed that it was virtually certain that internal variability alone could not account for observed warming since 1951 (Bindoff et al., 2013). Evidence has grown since AR5 that observed changes since the 1950s in many parts of the climate system are attributable to anthropogenic influence. So far, this chapter has focused on examining individual aspects of the climate system in separate sections. The results presented in Sections 3.3 to 3.7 substantially strengthen our assessment of the role of human influence on climate since pre-industrial times. In this section we look across the whole climate system to assess to what extent a physically consistent picture of human induced change emerges across the climate system (Figure 3.41).

Figure 3.41 | Summary figure showing simulated and observed changes in key large-scale indicators of climate change across the climate system, for continental, ocean basin and larger scales. Black lines show observations, brown lines and shading show the multi-model mean and 5th–95th percentile ranges for CMIP6 historical simulations including anthropogenic and natural forcing, and blue lines and shading show corresponding ensemble means and 5th–95th percentile ranges for CMIP6 natural-only simulations. Temperature time series are as in Figure 3.9, but with smoothing using a low pass filter. Precipitation time series are as in Figure 3.15 and ocean heat content as in Figure 3.26. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

The observed global surface air temperature warming of 0.9°C to 1.2°C in 2010–2019 is much greater than can be explained by internal variability (likely –0.2°C to +0.2°C) or natural forcings (likely –0.1°C to +0.1°C) alone, but consistent with the assessed anthropogenic warming (likely 0.8°C to 1.3°C; Section 3.3.1.1). It is very likely that human influence is the main driver of warming over land Section 3.3.1.1). Moreover, the atmosphere as a whole has warmed (Table 7.1), and it is very likely that human-induced greenhouse gas increases were the main driver of tropospheric warming since 1979 (Section 3.3.1.2). It is virtually certain that greenhouse gas forcing was the main driver of the observed changes in hot and cold extremes over land at the global scale (Cross-Chapter Box 3.2).

As might be expected from a warming atmosphere, moisture in the troposphere has increased and precipitation patterns have changed. Human influence has likely contributed to the observed changes in humidity and precipitation (Section 3.3.2). It is likely that human influence, in particular due to greenhouse gas forcing, is the main driver of the observed intensification of heavy precipitation in global land regions during recent decades (Cross-Chapter Box 3.2). The pattern of ocean salinity changes indicate that fresh regions are becoming fresher and that salty regions are becoming saltier as a result of changes in ocean-atmosphere fluxes through evaporation and precipitation (high confidence) making it extremely likely that human influence has contributed to observed near-surface and subsurface salinity changes since the mid-20th century (Section 3.5.2.2). Taken together, this evidence indicates a human influence on the water cycle.

It is very likely that human influence was the main driver of Arctic sea ice loss since the late 1970s (Section 3.4.1.1), and very likely that it contributed to the observed reductions in Northern Hemisphere springtime snow cover since 1950 (Section 3.4.2). Human influence was very likely the main driver of the recent global, near-universal retreat of glaciers and it is very likely that it contributed to the observed surface melting of the Greenland Ice Sheet over the past two decades (Section 3.4.3.2.1). It is extremely likely that human influence was the main driver of the ocean heat content increase observed since the 1970s (Section 3.5.1.3), and very likely that human influence was the main driver of the observed GMSL rise since at least 1970 (Section 3.5.3.2).

Combining the evidence from across the climate system (Sections 3.3–3.7) increases the level of confidence in the attribution of observed climate change to human influence and reduces the uncertainties associated with assessments based on a single variable. From this combined evidence, it is unequivocal that human influence has warmed the atmosphere, ocean and land components of the global climate system, taken together.

Similar to the assessment of multivariate attribution of climate change in the previous section, this section covers the performance of the models across different variables (Sections 3.8.2.1) and different classes of models (Section 3.8.2.2). Here the focus is on a system-wide assessment using integrative measures of model performance that characterize model performance using multiple diagnostic fields derived from multi-model ensembles.

3.8.2.1 Integrative Measures of Model Performance

The purpose of this section is to use multivariate analyses to address how well models simulate present-day and historical climate. For every diagnostic field considered, model performance is compared to one or multiple observational references, and the quality of the simulation is expressed as a single number, for example a correlation coefficient or a root mean square difference versus the observational reference. By simultaneously assessing different performance indices, model improvements can be quantified, similarities in behaviour between different models become apparent, and dependencies between various indices become evident (Gleckler et al., 2008; Waugh and Eyring, 2008).

AR5 found significant differences between models in the simulation of mean climate in the CMIP5 ensemble when measured against meteorological reanalyses and observations (Flato et al., 2013), see also Stouffer et al. (2017). The AR5 determined that for the diagnostic fields analysed, the models usually compared similarly against two different reference datasets, suggesting that model errors were generally larger than observational uncertainties or other differences between the observational references. In agreement with previous assessments, the CMIP5 multi-model mean generally performed better than individual models (Annan and Hargreaves, 2011; Rougier, 2016). The AR5 considered 13 atmospheric fields in its assessment for the instrumental period but did not assess multi-variate model performance in other climate domains (e.g., ocean, land, and sea ice). The AR5 found only modest improvement regarding the simulation of climate for two periods of the Earth’s history (the Last Glacial Maximum and the mid-Holocene) between CMIP5 and previous paleoclimate simulations. Similarly, for the modern period only modest, incremental progress was found between CMIP3 and CMIP5 regarding the simulation of precipitation and radiation. The representation of clouds also showed improvement, but remained a key challenge in climate modelling (Flato et al., 2013).

The type of multi-variate analysis of models presented in AR5 remains critical to building confidence for example in projections of climate change. It is expanded here to the previous-generation CMIP3 and present-generation CMIP6 models and also to more variables and more climate domains, covering land and ocean as well as sea ice. The multi-variate evaluation of these three generations of models is performed relative to the observational datasets listed in Annex I, Table AI.1. For many of these datasets, a rigorous characterization of the observational uncertainty is not available, see discussion in Chapter 2. Here, as much as possible, multiple independent observational datasets are used. Disagreements among them would cause differences in model scoring, indicating that observational uncertainties may be substantial compared to model errors. Conversely, similar scores against different observational datasets would suggest model biases may be larger than the observational uncertainty.

An analysis of a basket of 16 atmospheric variables (Figure 3.42a) assessed across CMIP3, CMIP5, and CMIP6 models but excluding high-resolution models participating in HighResMIP, reveals the progress made between these three generations of models (Bock et al., 2020). Progress is evidenced by the increasing prevalence of blue colours (indicating a performance better than the median) for the more recent model versions. Additionally, a few CMIP6 models outperform the best-performing CMIP5 models. Progress is evident across all 16 variables. As noted in AR5, the models typically score similarly against both observational reference datasets, indicating that indeed uncertainties in these reference datasets are smaller than model biases. Several models and model families perform better compared to observational references than the median, across a majority of the climate variables assessed, and conversely some other models or model families compare more poorly against these reference datasets. Such a good correspondence across a range of diagnostic fields probing different aspects of climate enhances confidence that the improved performances reflect progress in the physical realism of these simulations. An alternative explanation, that progress is due to a cancellation of errors achieved by model tuning, appears improbable given the large number of diagnostic fields involved here. However, several instances of poor model performance (red colours in Figure 3.42) still exist in the CMIP6 ensemble. Family relationships (i.e. various degrees of shared formulations; Knutti et al., 2013) between the models are apparent, for example, the GISS, GFDL, CESM, CNRM, and HadGEM/UKESM1/ACCESS families score similarly across all atmospheric variables, both for the CMIP5 and CMIP6 generations. In the cases of CESM2/CESM2(WACCM), CNRM-CM6-1/CNRM-ESM2-1, NorCPM1/NorESM2-LM, and HadGEM3-GC31-LL/UKESM1-0-LL, the high-complexity model scores as well or better than its lower-complexity counterpart, indicating that increasing complexity by adding Earth system features, which by removing constraints could be expected to degrade a model’s performance, does not necessarily do so. Several high climate-sensitivity models (Section 7.5; Meehl et al., 2020), in particular CanESM5, CESM2, CESM2-WACCM, HadGEM3-GC31-LL, and UKESM1-0-LL, score well against the benchmarks. In accordance with AR5 and earlier assessments, the multi-model mean, with some notable exceptions, is better than any individual model (Annan and Hargreaves, 2011; Rougier, 2016).

Figure 3.42 | Relative space–time root-mean-square deviation (RMSD) calculated from the climatological seasonal cycle of the CMIP simulations (1980–1999) compared to observational datasets. (a) CMIP3, CMIP5, and CMIP6 for 16 atmospheric variables (b) CMIP5 and CMIP6 for 10 land variables and four ocean/sea-ice variables. A relative performance measure is displayed, with blue shading indicating better and red shading indicating worse performance than the median of all model results. A diagonal split of a grid square shows the relative error with respect to the reference data set (lower right triangle) and an additional data set (upper left triangle). Reference/additional datasets are from top to bottom in (a): ERA5/NCEP, GPCP-SG/GHCN, CERES-EBAF, CERES-EBAF, CERES-EBAF, CERES-EBAF, JRA-55/ERA5, ESACCI-SST/HadISST, ERA5/NCEP, ERA5/NCEP, ERA5/NCEP, ERA5/NCEP, ERA5/NCEP, ERA5/NCEP, AIRS/ERA5, ERA5/NCEP and in (b): CERES-EBAF, CERES-EBAF, CERES-EBAF, CERES-EBAF, LandFlux-EVAL, Landschuetzer2016/ JMA-TRANSCOM; MTE/FLUXCOM, LAI3g, JMA-TRANSCOM, ESACCI-SOILMOISTURE, HadISST/ATSR, HadISST, HadISST, ERA-Interim. White boxes are used when data are not available for a given model and variable. Figure is updated and expanded from Bock et al. (2020), their Figure 5 CC BY 4.0https://creativecommons.org/licenses/by/4.0/. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Regarding model performance for the ocean and the cryosphere (Figure 3.42b), it is apparent that for many models there are substantial differences between the scores for Arctic and Antarctic sea ice concentration. This might suggest that it is not sea ice physics directly that is driving such differences in performance but rather other influences, such as differences in geography, the role of large ice shelves (which are absent in the Arctic), or large-scale ocean dynamics. As for atmospheric variables, progress is evident also across all four ocean and ten land variables from CMIP5 to CMIP6.

In summary, CMIP6 models perform generally better for a basket of variables covering mean historical climate across the atmosphere, ocean, and land domains than previous-generation and older models (high confidence). Earth System models characterized by additional biogeochemical feedbacks often perform at least as well as related more constrained, lower-complexity models lacking these feedbacks (medium confidence). In many cases, the models score similarly against both observational references, indicating that model errors are usually larger than observational uncertainties (high confidence). Moreover, synthesizing across Sections 3.3–3.7, we assess that the CMIP6 multi-model mean captures most aspects of observed climate change well (high confidence).

Using centred pattern correlations (quantifying pattern similarity on a scale of –1 to 1, with 1 expressing perfect similarity and 0 no relationship) for selected fields, AR5 documented improvements between CMIP3 and CMIP5 in surface air temperature, outgoing longwave radiation, and precipitation (Figure 9.6 of Flato et al., 2013). Little further progress between CMIP3 and CMIP5 was found for fields that were already quite well simulated in CMIP3 (such as surface air temperature and outgoing longwave radiation). For precipitation, the spread reduced because the worst-performing models improved. The shortwave cloud radiative effect remained relatively poorly simulated with significant inter-model spread (e.g., Calisto et al., 2014). This comparison of centred pattern correlations is designed to help determine the quality of simulation of different diagnostics relative to each other, and also to examine progress between generations of models. Figure 3.43 shows the centred pattern correlations for 16 variables for CMIP3, CMIP5 and CMIP6 models. In the ensemble averages, CMIP6 performs better than CMIP5 and CMIP3 for near-surface temperature, precipitation, mean sea-level pressure, and many other variables. For the variables shown, the uncertainties in observational datasets, in particular for precipitation and northward wind at 850 hPa, remain substantial relative to mean model errors (see grey dots in Figure 3.43).

Figure 3.43 | Centred pattern correlations between models and observations for the annual mean climatology over the period 1980–1999. Results are shown for individual CMIP3 (green), CMIP5 (blue) and CMIP6 (red) models (one ensemble member from each model is used) as short lines, along with the corresponding multi-model ensemble averages (long lines). Correlations are shown between the models and the primary reference observational data set (from left to right: ERA5, GPCP-SG, CERES-EBAF, CERES-EBAF, CERES-EBAF, CERES-EBAF, JRA-55, ESACCI-SST, ERA5, ERA5, ERA5, ERA5, ERA5, ERA5, AIRS, ERA5). In addition, the correlation between the primary reference and additional observational datasets (from left to right: NCEP, GHCN, -, -, -, -, ERA5, HadISST, NCEP, NCEP, NCEP, NCEP, NCEP, NCEP, ERA5, NCEP) are shown (solid grey circles) if available. To ensure a fair comparison across a range of model resolutions, the pattern correlations are computed after regridding all datasets to a resolution of 4° in longitude and 5° in latitude. Figure is updated and expanded from Bock et al. (2020), their Figure 7 CC BY 4.0https://creativecommons.org/licenses/by/4.0/. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

In addition to the multivariate assessments of simulations of the recent historical period, simulations of selected periods of the Earth’s more distant history can be used to benchmark climate models by exposing them to climate forcings that are radically different from the present and recent past (Harrison et al., 2015, 2016; Kageyama et al., 2018; Tierney et al., 2020a). These time periods provide an out-of-sample test of models because they are not in general used in the process of model development. They encompass a range of climate drivers, such as volcanic and solar forcing for the Last Millennium, orbital forcing for the mid-Holocene and Last Interglacial, and changes in greenhouse gases and ice sheets for the LGM, mid-Pliocene Warm Period, and early Eocene (Sections 2.2 and 2.3). These drivers led to climate changes, including in surface temperature (Section 2.3.1.1) and the hydrological cycle (Section 2.3.1.3.1), which are described by paleoclimate proxies that have been synthesized to support evaluations of models on a global and regional scale. However, the more sparse, indirect, and regionally incomplete climate information available from paleo-archives motivates a different form of the multivariate analysis of simulations covering these periods versus the equivalent for the historical period, as described below.

AR5 found that reconstructions and simulations of past climates both show similar responses in terms of large-scale patterns of climate change, such as polar amplification (Flato et al., 2013; Masson-Delmotte et al., 2013). However, for several regional signals (e.g., the north–south temperature gradient in Europe and regional precipitation changes), the magnitude of change seen in the proxies relative to the pre-industrial period was underestimated by the models. When benchmarking CMIP5/PMIP3 models against reconstructions of the mid-Holocene and LGM, AR5 found only a slight improvement compared with earlier model versions across a range of variables. For the Last Interglacial, it was noted that the magnitude of observed annual mean warming in the Northern Hemisphere was only reached in summer in the models. For the mid-Pliocene Warm Period, it was noted that both proxies and models showed a polar amplification of temperature compared with the pre-industrial period, but a formal model evaluation was not carried out.

Since AR5, new simulation protocols have been developed in PMIP4 (Kageyama et al., 2018), which are further described for the mid-Holocene and the Last Interglacial by Otto-Bliesner et al. (2017), for the LGM by Kageyama et al. (2017), for the Pliocene by Haywood et al. (2016), and for the early Eocene by Lunt et al. (2017). These have resulted in new model simulations for these time periods (Brierley et al., 2020; Haywood et al., 2020; Kageyama et al., 2021a; Lunt et al., 2021; Otto-Bliesner et al., 2021). These time periods span an assessed temperature range of 20°C (Section 2.3.1.1), and for all periods the PMIP4 multi-model ensemble mean is within 0.5°C of the assessed range of GSAT (Figure 3.44a). Those time periods for which the multi-model ensemble mean is outside the assessed range of GSAT, the mid-Holocene and the Last Interglacial, are primarily forced by orbital changes not greenhouse gas forcing, and as a result the forcing as well as the assessed and modelled response are relatively close to zero in the global annual mean. During these periods, climate change therefore is a consequence of more poorly understood Earth System feedbacks acting on the response to orbital differences versus the present-day, affecting the seasonality of insolation.

Figure 3.44 | Multivariate synopsis of paleoclimate model results compared to observational references. Data-model comparisons for (a) GSAT anomalies for five PMIP4 periods and for regional features for the (b) mid-Holocene and (c) LGM periods, for PMIP3 and PMIP4 models. The results from CMIP6 models are shown as coloured dots. In (a) the light orange shading shows the very likely assessed ranges presented in Section 2.3.1.1. In (b) and (c), the regions and variables are defined as follows: North America (20°N–50°N, 140°W–60°W), Western Europe (35°N–70°N, 10°W–30°E) and West Africa (0°–30°N, 10°W–30°E); mean temperature of the coldest month (MTCO; °C), mean temperature of the warmest month (MTWA; °C), mean annual precipitation (MAP; mm yr^–1). In (b) and (c) the ranges shown for the reconstructions (Bartlein et al. (2011) for mid-Holocene and Cleator et al. (2020) for LGM) are based on the standard error given at each site: the average and associated standard deviation over each area is obtained by computing 1000 times the average of randomly drawn values from the Gaussian distributions defined at each site by the reconstruction mean and standard error; the light orange colour shows the ±1 standard deviation of these 1000 estimates. The dots on (b) and (c) show the average of the model output for grid points for which there are reconstructions. The ranges for the model results are based on an ensemble of 1000 averages over 50 years randomly picked in the model output time series for each region and each variable: the mean ± one standard deviation is plotted for each model. Figure is adapted from Brierley et al. (2020), their Figure S3 for the mid-Holocene; and from Kageyama et al. (2021b), their Figure 12 for the LGM. Further details on data sources and processing are available in the chapter data table (Table 3.SM.1).

Open figure

Polar amplification in the LGM, mid-Pliocene Warm Period, and Early Eocene Climatic Optimum (EECO) simulations is assessed in Section 7.4.4.1.2. Here we focus on the mid-Holocene and the LGM, which have been a part of AMIP or CMIP through several assessment cycles, and as such serve as a reference to quantify regional model-data agreement from one IPCC assessment to another. We compare the results from 15 CMIP6 models using the PMIP4 protocol (CMIP6-PMIP4), with non-CMIP6 models using the PMIP4 protocol, with PMIP3 models, and with regional temperature and precipitation changes from proxies for the mid-Holocene (Figure 3.44b). For six out of seven variables shown, the CMIP6 multi-model mean captures the correct sign of the change. For five out of seven of them the CMIP6 ensemble mean is within the reconstructed range. For the other two variables (changes in the mean temperature of the warmest month over North America and in the mean annual precipitation over West Africa) nearly all PMIP4 and PMIP3 models are outside the reconstructed range. CMIP6 models show regional patterns of temperature changes similar to the PMIP3 ensemble (Brierley et al., 2020), but the slight mid-Holocene cooling in PMIP4 compared with PMIP3, probably associated with lower imposed mid-Holocene carbon dioxide concentrations (Otto-Bliesner et al., 2017), improves the regional model performance for summer and winter temperatures (Figure 3.44b). However, this cooling also results in a CMIP6 mid-Holocene GSAT that lies further from the assessed range (Figure 3.44a). All models show an expansion of the monsoon areas from the pre-industrial to the mid-Holocene simulations in the Northern Hemisphere, but this expansion in some cases is only large enough to cancel out the bias in the pre-industrial control simulations (Section 3.3.3.2; Brierley et al., 2020). There is a slight improvement in representing the northward expansion of the West African monsoon region in PMIP4 compared with PMIP3 (Figures 3.11 and 3.44b).

Fourteen simulations of the LGM climate have been produced following the CMIP6-PMIP4 protocol using 11 models, five of which are from the latest CMIP6 generation. The multi-model-mean global cooling simulated by these models is close to that simulated by the CMIP5-PMIP3 ensemble, but the range of results is larger. The increase in the range is largely due to the inclusion of CESM2 which simulates a much larger cooling than the other PMIP4 models (Figure 3.44a). This is consistent with its larger climate sensitivity (see also (Section 3.3.1.1; Zhu et al., 2021). The other models on average also simulate slightly larger cooling in PMIP4 versus PMIP3 (Kageyama et al., 2017, 2021a). The PMIP4 multi-model mean is within the range of reconstructed regional averages for four out of seven regional variables; this is unchanged from PMIP3 but for different variables (Figure 3.44c). For all fields, the results of many individual models are outside the reconstructed range. For two variables out of seven (changes in the mean temperature of the warmest month and mean annual precipitation over Western Europe) no model is within the range of the reconstructions. This analysis is strengthened compared with the equivalent analysis in AR5 because it is based on larger and improved reconstructions (Cleator et al., 2020). Most CMIP6-PMIP4 models simulate a slightly stronger AMOC in the LGM, but no strong deepening of the AMOC (Kageyama et al., 2021a), while most other PMIP4 models simulate a strengthening and strong deepening of the AMOC, as was the case for the PMIP3 models (Muglia and Schmittner, 2015; Sherriff-Tadano et al., 2018). Only one model (CESM1.2) shows a shoaling of LGM AMOC which is consistent with reconstructions (Marzocchi and Jansen, 2017; Sherriff-Tadano et al., 2018).

17 PMIP4 models completed Last Interglacial simulations (Otto-Bliesner et al., 2021). The comparison to reconstructions is generally good, except for some discrepancies, such as for upwelling systems in the South East Atlantic or discrepancies which may result from local melting of remnant ice sheets absent in the Last Interglacial simulation protocol. All models simulate a decrease in Arctic sea ice in summer, commensurate with increased summer insolation, while some models even simulate a large or complete loss (Guarino et al., 2020; Kageyama et al., 2021b). Sea ice reconstructions for the central Arctic are, however, too uncertain to evaluate this behaviour. The Last Interglacial simulations indicate a clear relationship between simulated sea ice loss and model responses to increased greenhouse gas forcing (Kageyama et al., 2021b; Otto-Bliesner et al., 2021).

Overall, the PMIP multi-model means agree very well (within 0.5°C of the assessed range) with GSAT reconstructed from proxies across multiple time periods, spanning a range from 6°C colder than pre-industrial (Last Glacial Maximum) to 14°C warmer than pre-industrial (Early Eocene Climate Optimum) (high confidence). During the orbitally-forced mid-Holocene, the CMIP6 multi-model mean captures the sign of the regional changes in temperature and precipitation in most regions assessed, and there have been some regional improvements compared to AR5 (medium confidence). The limited number of CMIP6 simulations of the LGM hinders model evaluation of the multi-model mean, but for both LGM and mid-Holocene, models tend to underestimate the magnitude of large changes (high confidence). Some long-standing model-data discrepancies, such as a dry bias in North Africa in the mid-Holocene, have not improved in CMIP6 compared with PMIP3 (high confidence).

The dominant role of humans in driving recent climate change is clear. This conclusion is based on a synthesis of information from multiple lines of evidence, including direct observations of recent changes in Earth’s climate; analyses of tree rings, ice cores, and other long-term records documenting how the climate has changed in the past; and computer simulations based on the fundamental physics that governs the climate system.

Climate is influenced by a range of factors. There are two main natural drivers of variations in climate on time scales of decades to centuries. The first is variations in the sun’s activity, which alter the amount of incoming energy from the sun. The second is large volcanic eruptions, which increase the number of small particles (aerosols) in the upper atmosphere that reflect sunlight and cool the surface–an effect that can last for several years (see also FAQ 3.2). The main human drivers of climate change are increases in the atmospheric concentrations of greenhouse gases and of aerosols from burning fossil fuels, land use and other sources. The greenhouse gases trap infrared radiation near the surface, warming the climate. Aerosols, like those produced naturally by volcanoes, on average cool the climate by increasing the reflection of sunlight. Multiple lines of evidence demonstrate that human drivers are the main cause of recent climate change.

The current rates of increase of the concentration of the major greenhouse gases (carbon dioxide, methane and nitrous oxide) are unprecedented over at least the last 800,000 years. Several lines of evidence clearly show that these increases are the results of human activities. The basic physics underlying the warming effect of greenhouse gases on the climate has been understood for more than a century, and our current understanding has been used to develop the latest generation climate models (see FAQ 3.3). Like weather forecasting models, climate models represent the state of the atmosphere on a grid and simulate its evolution over time based on physical principles. They include a representation of the ocean, sea ice and the main processes important in driving climate and climate change.

Results consistently show that such climate models can only reproduce the observed warming (black line in FAQ 3.1, Figure 1) when including the effects of human activities (grey band in FAQ 3.1, Figure 1), in particular the increasing concentrations of greenhouse gases. These climate models show a dominant warming effect of greenhouse gas increases (red band, which shows the warming effects of greenhouse gases by themselves), which has been partly offset by the cooling effect of increases in atmospheric aerosols (blue band). By contrast, simulations that include only natural processes, including internal variability related to El Niño and other similar variations, as well as variations in the activity of the sun and emissions from large volcanoes (green band in FAQ 3.1, Figure 1), are not able to reproduce the observed warming. The fact that simulations including only natural processes show much smaller temperature increases indicates that natural processes alone cannot explain the strong rate of warming observed. The observed rate can only be reproduced when human influence is added to the simulations.

Moreover, the dominant effect of human activities is apparent not only in the warming of global surface temperature, but also in the pattern of warming in the lower atmosphere and cooling in the stratosphere, warming of the ocean, melting of sea ice, and many other observed changes. An additional line of evidence for the role of humans in driving climate change comes from comparing the rate of warming observed over recent decades with that which occurred prior to human influence on climate. Evidence from tree rings and other paleoclimate records shows that the rate of increase of global surface temperature observed over the past fifty years exceeded that which occurred in any previous 50-year period over the past 2000 years (see FAQ 2.1).

Taken together, this evidence shows that humans are the dominant cause of observed global warming over recent decades.

FAQ 3.1, Figure 1 | Observed warming (1850–2019) is only reproduced in simulations including human influence. Global surface temperature changes in observations, compared to climate model simulations of the response to all human and natural forcings (grey band), greenhouse gases only (red band), aerosols and other human drivers only (blue band) and natural forcings only (green band). Solid coloured lines show the multi-model mean, and coloured bands show the 5–95% range of individual simulations.

Natural variability refers to variations in climate that are caused by processes other than human influence. It includes variability that is internally generated within the climate system and variability that is driven by natural external factors. Natural variability is a major cause of year-to-year changes in global surface climate and can play a prominent role in trends over multiple years or even decades. But the influence of natural variability is typically small when considering trends over periods of multiple decades or longer. When estimated over the entire historical period (1850–2020), the contribution of natural variability to global surface warming of –0.23°C to +0.23°C is small compared to the warming of about 1.1°C observed during the same period, which has been almost entirely attributed to the human influence.

Paleoclimatic records (indirect measurements of climate that can extend back many thousands of years) and climate models all show that global surface temperatures have changed significantly over a wide range of time scales in the past. One of these reasons is natural variability, which refers to variations in climate that are either internally generated within the climate system or externally driven by natural changes. Internal natural variability corresponds to a redistribution of energy within the climate system (for example via atmospheric circulation changes similar to those that drive the daily weather) and is most clearly observed as regional, rather than global, fluctuations in surface temperature. External natural variability can result from changes in the Earth’s orbit, small variations in energy received from the sun, or from major volcanic eruptions. Although large orbital changes are related to global climate changes of the past, they operate on very long time scales (i.e., thousands of years). As such, they have displayed very little change over the past century and have had very little influence on temperature changes observed over that period. On the other hand, volcanic eruptions can strongly cool the Earth, but this effect is short-lived and their influence on surface temperatures typically fades within a decade of the eruption.

To understand how much of observed recent climate change has been caused by natural variability (a process referred to as attribution), scientists use climate model simulations. When only natural factors are used to force climate models, the resulting simulations show variations in climate on a wide range of time scales in response to volcanic eruptions, variations in solar activity, and internal natural variability. However, the influence of natural climate variability typically decreases as the time period gets longer, such that it only has mild effects on multi-decadal and longer trends (FAQ 3.2, Figure 1).

Consequently, over periods of a couple of decades or less, natural climate variability can dominate the human-induced surface warming trend – leading to periods with stronger or weaker warming, and sometimes even cooling (FAQ 3.2, Figure 1, left and centre). Over longer periods, however, the effect of natural variability is relatively small (FAQ 3.2, Figure 1, right). For instance, over the entire historical period (1850–2019), natural variability is estimated to have caused between –0.23°C and +0.23°C of the observed surface warming of about 1.1°C. This means that the bulk of the warming has been almost entirely attributed to human activities, particularly emissions of greenhouse gases (FAQ 3.1).

Another way to picture natural variability and human influence is to think of a person walking a dog. The path of the walker represents the human-induced warming, while their dog represents natural variability. Looking at global surface temperature changes over short periods is akin to focusing on the dog. The dog sometimes moves ahead of the owner and other times behind. This is similar to natural variability that can weaken or amplify warming on the short term. In both cases it is difficult to predict where the dog will be or how the climate will evolve in the near future. However, if we pull back and focus on the slow steady steps of the owner, the path of the dog is much clearer and more predictable, as it follows the path of its owner. Similarly, human influence on the climate is much clearer over longer time periods.

FAQ 3.2, Figure 1 | Annual (left), decadal (middle) and multi-decadal (right) variations in average global surface temperature. The thick black line is an estimate of the human contribution to temperature changes, based on climate models, whereas the green lines show the combined effect of natural variations and human-induced warming, different shadings of green represent different simulations, which can be viewed as showing a range of potential pasts. The influence of natural variability is shown by the green bars, and it decreases on longer time scales. The data is sourced from the CESM1 large ensemble.

Yes, climate models have improved and continue to do so, becoming better at capturing complex and small-scale processes and at simulating present-day mean climate conditions. This improvement can be measured by comparing climate simulations against historical observations. Both the current and previous generations of models show that increases in greenhouse gases cause global warming. While past warming is well simulated by the new generation models as a group, some individual models simulate past warming that is either below or above what is observed. The information about how well models simulate past warming, as well as other insights from observations and theory, are used to refine this Report’s projections of global warming.

Climate models are important tools for understanding past, present and future climate change. They are sophisticated computer programs that are based on fundamental laws of physics of the atmosphere, ocean, ice, and land. Climate models perform their calculations on a three-dimensional grid made of small bricks or ‘gridcells’ of about 100 km across. Processes that occur on scales smaller than the model grid cells (such as the transformation of cloud moisture into rain) are treated in a simplified way. This simplification is done differently in different models. Some models include more processes and complexity than others; some represent processes in finer detail (smaller grid cells) than others. Hence the simulated climate and climate change vary between models.

Climate modelling started in the 1950s and, over the years, models have become increasingly sophisticated as computing power, observations and our understanding of the climate system have advanced. The models used in the IPCC First Assessment Report published in 1990 correctly reproduced many aspects of climate (FAQ 1.1). The actual evolution of the climate since then has confirmed these early projections, when accounting for the differences between the simulated scenarios and actual emissions. Models continue to improve and get better and better at simulating the large variety of important processes that affect climate. For example, many models now simulate complex interactions between different aspects of the Earth system, such as the uptake of carbon dioxide by vegetation on land and by the ocean, and the interaction between clouds and air pollutants. While some models are becoming more comprehensive, others are striving to represent processes at higher resolution, for example to better represent the vortices and swirls in currents responsible for much of the transport of heat in the ocean.

Scientists evaluate the performance of climate models by comparing historical climate model simulations to observations. This evaluation includes comparison of large-scale averages as well as more detailed regional and seasonal variations. There are two important aspects to consider: (i) how models perform individually and (ii) how they perform as a group. The average of many models often compares better against observations than any individual model, since errors in representing detailed processes tend to cancel each other out in multi-model averages.

As an example, FAQ 3.3 Figure 1 compares simulations from the three most recent generations of models (available around 2008, 2013 and 2021) with observations of three climate variables. It shows the correlation between simulated and observed patterns, where a value of 1 represents perfect agreement. Many individual models of the new generation perform significantly better, as indicated by values closer to 1. As a group, each generation out-performs the previous generation: the multi-model average (shown by the longer lines) is progressively closer to 1. The vertical extent of the colored bars indicates the range of model performance across each group. The top of the bar moves up with each generation, indicating improved performance of the best performing models from one generation to the next. In the case of precipitation, the performance of the worst performing models is similar in the two most recent model generations, increasing the spread across models.

Developments in the latest generation of climate models, including new and better representation of physical, chemical and biological processes, as well as higher resolution, have improved the simulation of many aspects of the Earth system. These simulations, along with the evaluation of the ability of the models to simulate past warming as well as the updated assessment of the temperature response to a doubling of CO₂ in the atmosphere, are used to estimate the range of future global warming (FAQ 7.3).

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