Volume 49, Issue 19 e2022GL100653
Research Letter
Free Access

Cloud and Surface Albedo Feedbacks Reshape 21st Century Warming in Successive Generations of An Earth System Model

David P. Schneider

Corresponding Author

David P. Schneider

CIRES, University of Colorado Boulder, Boulder, CO, USA

Climate and Global Dynamics Laboratory, NCAR, Boulder, CO, USA

Correspondence to:

D. P. Schneider,

[email protected]

Contribution: Conceptualization, Methodology, Formal analysis, ​Investigation, Data curation, Writing - original draft, Writing - review & editing, Project administration, Funding acquisition

Search for more papers by this author
Jennifer E. Kay

Jennifer E. Kay

CIRES, University of Colorado Boulder, Boulder, CO, USA

ATOC, University of Colorado Boulder, Boulder, CO, USA

Contribution: Conceptualization, Methodology, Formal analysis, ​Investigation, Resources, Writing - original draft, Writing - review & editing, Supervision, Project administration, Funding acquisition

Search for more papers by this author
Cecile Hannay

Cecile Hannay

Climate and Global Dynamics Laboratory, NCAR, Boulder, CO, USA

Contribution: Software, Formal analysis, ​Investigation, Data curation, Writing - review & editing

Search for more papers by this author
First published: 20 September 2022
Citations: 1

Abstract

The relative importance of radiative feedbacks and emissions scenarios in controlling surface warming patterns is challenging to quantify across model generations. We analyze three variants of the Community Earth System Model (CESM) with differing equilibrium climate sensitivities under identical CMIP5 historical and high-emissions scenarios. CESM1, our base model, exhibits Arctic-amplified warming with the least warming in the Southern Hemisphere middle latitudes. A variant of CESM1 with enhanced extratropical shortwave cloud feedbacks shows slightly increased late-21st century warming at all latitudes. In the next-generation model, CESM2, global-mean warming is also slightly greater, but the warming is zonally redistributed in a pattern mirroring cloud and surface albedo feedbacks. However, if the nominally equivalent CMIP6 scenario is applied to CESM2, the redistributed warming pattern is preserved, but global-mean warming is significantly greater. These results demonstrate how model structural differences and scenario differences combine to produce differences in climate projections across model generations.

Key Points

  • Higher equilibrium climate sensitivity, from more positive cloud feedbacks, only slightly increases 21st century transient warming

  • Cloud and surface albedo feedbacks redistribute warming from the middle and high northern latitudes to lower latitudes

  • These warming pattern differences result from feedback differences, not CMIP5 to CMIP6 forcing scenario differences

Plain Language Summary

Earth System Models are used to project climate change through the 21st century. Such projections are constrained by a range of assumptions about emissions of greenhouse gases as well as other climate-relevant human activities like deforestation and emissions of aerosols. These assumptions are carefully evaluated and represented in an emissions scenario. In a coordinated set of experiments, a given emissions scenario is applied to a number of different Earth System Models to test differences in climate projections that arise from physical (or “structural”) differences among the models. However, over time, models are updated and improved, and emissions scenarios are updated as well. This continuous updating makes it difficult to track whether a new set of climate projections is different from its predecessor due to model changes or due to scenario changes. In this paper, we disentangle these two factors, clarifying the role of each. For the Earth System Model that we use, the emissions scenario largely controls the amount of global-mean warming projected for the 21st century, but the model or structural differences matter greatly for the regional patterns of warming.

1 Introduction

While the largest uncertainty in long-term climate projections comes from the choices that humanity makes with respect to mitigating or increasing greenhouse gas emissions (Hawkins & Sutton, 2009), substantial physical uncertainty arises from cloud feedbacks, and their interaction with other parts of the Earth system (e.g., Zelinka et al., 2020). Recognizing these physical uncertainties, many modeling groups have made concerted efforts to improve the representation of cloud processes, with important consequences for modeled values of equilibrium climate sensitivity (ECS; the global-mean surface warming for a doubling of atmospheric CO2 concentration above pre-industrial levels). For example, both the United Kingdom Earth System Model (UKESM) and CESM groups recently updated their models' representation of clouds (Bodas-Salcedo et al., 2019; Gettelman et al., 2019), and both groups' latest models now are among the CMIP6 models with the highest ECS values (Meehl et al., 2020). Approximately one-third of the Earth System Models participating in the Coupled Model Intercomparison Project, Phase 6 (CMIP6; Eyring et al., 2016) exhibit higher ECS than the previous generation of models (Forster et al., 2020; Meehl et al., 2020; Zelinka et al., 2020).

Recent efforts have focused on understanding the realism of these high ECS values, their relationship to the rate of 21st century warming, and their policy implications. While ECS is widely considered a robust and easily implemented metric for comparing models, its value for constraining transient warming under realistic 21st century forcing scenarios is unclear (e.g., Knutti et al., 2017). Although some high-ECS models do exhibit more 21st century warming than their predecessors (Forster et al., 2020), some do not. As a case in point, a variant of the CESM1 with notable cloud phase improvements (Kay, Wall, et al., 2016) exhibited a 1.5 K increase in ECS, but only a minor increase in the rate of 21st century warming (Frey et al., 2017). This disconnect between ECS and transient warming resulted from the co-location of strong positive shortwave cloud feedbacks and strong oceanic heat uptake. On the policy front, the low probability but high risk that these high ECS models may be right and may lead to substantially more 21st century warming has led some scientists to increase the urgency for mitigation of anthropogenic greenhouse gas emissions (Forster et al., 2020). Yet on the realism front, recent evidence from cloud observations and more constrained paleoclimate records suggests that the high ECS values are unlikely (e.g., Cesana & Del Genio, 2021; Myers et al., 2021; Zhu et al., 2021).

One of the challenges in narrowing structural uncertainty and assessing model realism is the difficulty in tracing model behavior from one generation to the next to specific changes in the model code. This challenge is especially acute when working within a multi-model framework such as CMIP6, as model differences arise for several reasons, including internal variability and the details of how emissions scenarios are developed and implemented. As a result, it is nearly impossible for the authors of a multi-model assessment study to be familiar with all aspects of every model in the assessment and how they have changed from one generation to the next. While cross-model correlations can be the basis of a compelling story (e.g., high ECS models have more positive shortwave cloud feedbacks and greater transient warming), they cannot provide causal explanations of the underlying specific mechanisms.

These interwoven issues of structural uncertainty, ECS, and cloud feedbacks motivate our approach of using a single model framework to isolate the influence of cloud and other radiative feedbacks on transient patterns of 21st century warming. Specifically, three variants of the CESM provide a unique opportunity to diagnose and document radiative feedbacks and their relationship to transient warming. CESM1 is a well-documented model with a very good representation of the mean climate and its variability, and an ECS of 4.1 K (Hurrell et al., 2013). Nonetheless, CESM1 and its atmospheric component, the Community Atmosphere Model version 5 (CAM5), suffer from some major mean-state biases, including large absorbed shortwave radiation (ASR) biases (Kay et al., 2012) and a lack of supercooled cloud liquid at middle and high latitudes (Kay, Bourdages, et al., 2016; Tan et al., 2016). Modifications to CAM5 clouds (Kay, Wall, et al., 2016) substantially reduced mean ASR biases especially over the Southern Ocean and increased the ECS by 1.5–5.6 K (Frey & Kay, 2018). Here, we extend previous work with this cloud-improved CESM1 by running more ensemble members. Since Frey et al. (2017), CESM2 has been released, which has an ECS of 5.3 K (Bacmeister et al., 2020; Danabasoglu et al., 2020). Due to differences in transient emissions scenarios across CMIP generations, it is not possible to directly compare transient 21st-century warming patterns in CESM2 under a CMIP6 scenario with CESM1 under a nominally equivalent CMIP5 scenario and relate them to their ECS differences. Here, we address this scenario uncertainty by comparing transient experiments with the same CMIP5 scenario applied to all three CESM variants (CESM1, cloud-improved CESM1, and CESM2). This paper also distinguishes model structural uncertainty from scenario uncertainty by contrasting CESM2 with CMIP5 forcing and CESM2 with the nominally equivalent CMIP6 forcing.

2 Methods

We employ three variants of the CESM, and apply the same CMIP5-era historical (1850–2005) and RCP8.5 (2006–2099; Taylor et al., 2012) forcing to each version, as summarized in Table S1 in Supporting Information S1. The forcing is slightly different from the default CMIP5 forcing protocol; see Kay et al., 2015 for details. The first variant is the large ensemble (LE) version of CESM1 (Kay et al., 2015). The 40-member CESM1-LE enables a robust estimate of the forced response from the ensemble mean and gives a probable range of outcomes due to internal variability from the ensemble spread. By comparing the other experiments against the CESM1-LE, we can evaluate the significance of structural uncertainty in 21st Century transient warming. The second variant is the cloud-modified variant of CESM1-LE, which we refer to as “CESM1-LE*.” With CESM1-LE*, we use the one historical and RCP8.5 member from Frey et al. (2017), with four new additional ensemble members run for this work. The advantage of using CESM1-LE* in this work is the cloud and radiation improvements in CESM1-LE* are well documented and are the only changes that have been made (Kay, Wall, et al., 2016). Thus, differences in the mean climate state and the transient response to external forcing between CESM1-LE and CESM1-LE* can be attributed to these cloud parameterization changes alone. The third model variant, CESM2, described in Danabasoglu et al. (2020), is a major evolution from CESM1. Many of the changes between CESM1 and CESM2 are from iterating the atmosphere model from CAM5 to CAM6. CAM6 is the primary reason why CESM2 has higher ECS than CESM1; Gettelman et al. (2019) diagnose how CAM6 got its high ECS. We include a new ensemble of transient experiments with CESM2 under CMIP5 forcing, evaluating five ensemble members for the 1850–2100 period. Finally, we also include results from a 3-member CESM2 ensemble with CMIP6 historical (1850–2014) and SSP5-8.5 (2015–2100) forcings (Eyring et al., 2016; O’Neill et al., 2013).

We evaluate radiative feedbacks with both the Approximate Radiative Perturbation (APRP) (Taylor et al., 2007) and radiative kernel (Soden et al., 2008) methods applied to the first five ensemble members of CESM1-LE, the 5 members of CESM1-LE*, the first 5 members of CESM2 with CMIP5 forcing and the 3 members of CESM2 with CMIP6 forcing. Based on Morrison et al., 2018 and Shell et al., 2008, we expect APRP to yield more accurate results for shortwave feedbacks in the polar regions, especially around the sea ice edge. We therefore report results for shortwave feedbacks from the APRP method. Longwave feedbacks are from the kernel method, using the radiative kernels developed for CAM5 by Pendergrass et al. (2018).

3 Results

We first consider the global, annual mean rates of warming over the 20th and 21st centuries. All four ensembles (Figure 1a) show similar evolutions during the 20th century. Nonetheless, the CESM2-CMIP6 ensemble shows weak 20th-century warming until about 1980, which is more noticeable in the absolute surface temperature timeseries (Figure S1 in Supporting Information S1). This may be due to the relatively weak greenhouse gas forcing and strong aerosol forcing during this time period (e.g., Smith & Forster, 2021). By the mid-21st Century, the ensembles begin to separate, with CESM1-LE* and CESM2-CMIP5 taking a slightly warmer path than CESM1-LE. After 2070, CESM2-CMIP6 takes a considerably warmer path. By the late 21st century (2080–2099), the CESM1-LE global-mean warming ranges from 3.49 to 3.73 K above the early 21st century baseline period (2000–2019). The CESM1-LE* and CESM2-CMIP5 ensembles are different from the CESM1-LE at the 90% confidence level (Figure S2 in Supporting Information S1). In contrast, the CESM2-CMIP6 ensemble-mean warms about 0.80 K more than the ensemble mean of CESM1-LE, and the entire ensemble is well outside the range of CESM1-LE.

Details are in the caption following the image

(a) Global surface temperature timeseries for each of four experiments with historical and high-emissions scenarios, for 1900 or 1920–2099. The thicker lines are the ensemble means and the thinner lines are individual ensemble members; (b) Zonal-mean surface temperature change by latitude bands for each of the four experiments (ensemble means shown); (c) Zonal-mean amplification of surface temperature change (relative to global-mean warming) by latitude bands (ensemble means shown).

The small additional late 21st-century warming in the CESM1-LE* ensemble over the CESM1-LE ensemble occurs at all latitude bands except for 50–70°S (Figure 1b). In contrast, the large additional warming in the CESM2-CMIP6 ensemble occurs at all latitude bands except for 70–90°N. The greater warming is especially noticeable in the tropics, subtropics, and Antarctic. Considering that CESM2-CMIP6 warms almost 1°C more globally than CESM1-LE, it is remarkable that it warms less in the high Arctic (70–90°N). Comparing the CESM2 under CMIP5 and CMIP6 forcing, the differences in warming rates are similar across all latitude bands.

All three variants of CESM show the maximum amplification of global-mean warming in the Arctic, and the minimum in amplification in the SH middle latitudes (Figure 1c). There are only minor differences in the amplification patterns of CESM1-LE and CESM1-LE*. However, in both experiments with CESM2, amplification at 70–90°N is about 25% weaker relative to both variants of CESM1. It is also somewhat weaker at 50–70°N and 30–50°N. However, amplification in CESM2 is slightly stronger than in CESM1 in the tropics and subtropics (30–30°N) as well as the Antarctic (70–90°S). There are no significant differences in the amplification patterns of CESM2 under CMIP5 versus CMIP6 forcing.

In map view (Figure 2), striking warming pattern differences are evident among the three CESM variants. Compared with CESM1-LE, CESM1-LE* warms more over the Southern, North Pacific, and central Arctic oceans, and across parts of Antarctica, Australia, and northern North America. In contrast, CESM1-LE* warms less than CESM1-LE in the North Atlantic and high latitudes of the Southern Ocean. In the maps of CESM2 minus CESM1-LE and CESM2 minus CESM1-LE*, several “hot spots” stand out. CESM2 exhibits greater than 1°C more warming than CESM1-LE in the eastern ocean basins of the SH (east Indian, east Pacific, east Atlantic), across parts of the Australian and African continents, and around Hudson Bay. CESM2 warms less than CESM1-LE across the Arctic Ocean and North Atlantic, and across northern Europe and Scandinavia. CESM2 minus CESM1-LE* is broadly similar to CESM2 minus CESM1-LE, but there are some differences. For example, CESM2 doesn't warm as much as CESM1-LE* in the middle latitudes of the SH.

Details are in the caption following the image

Global maps of ensemble-mean surface air temperature difference (left column) and shortwave cloud feedback differences (W/m2 K) (right column) for each of the four experiments. Ensemble mean differences (2080–2099) are used for the feedbacks, while the temperature differences are based on ensemble-mean linear trends over 2000–2099. The globally integrated values are shown at the upper right corner of each map plot.

The warming pattern difference between CESM2-CMIP6 and CESM2-CMIP5 (Figure 2) mimics the mean warming pattern (Figure S3 in Supporting Information S1), with the largest differences in the Arctic and Antarctic, followed by the continental interiors. A few recent studies have compared warming under CMIP5 and CMIP6 forcing in the historical period, finding some important regional differences due to aerosols (Fasullo et al., 2022; Fyfe et al., 2021), and, on the global scale, a complex interplay among aerosols, greenhouse gasses, and ECS (Smith & Forster, 2021). In the 21st century, although RCP8.5 and SSP5-8.5 have nominally the same end-of-century radiative forcing, SSP5.8–5 has 199 ppmv higher CO2 (Fyfe et al., 2021). The global nature of the late 21st century warming pattern difference seen here suggests that CO2 is the main reason for it, and is consistent with the results of Fyfe et al. (2021), who used the Canadian Earth System Model (CanESM) and found more global-mean warming with the SSP5-8.5 scenario than with RCP8.5.

Turning to radiative feedback analysis, results for the CESM1-LE* and CESM1-LE ensembles are consistent with those of Frey et al. (2017), showing that the globally integrated shortwave cloud feedback is more positive in CESM1-LE* than in CESM1-LE (Figure 2 and Table S1 in Supporting Information S1). With the same CMIP5 forcing, CESM2 shows about the same global-mean value as CESM1-LE*. However, the maximum values of the feedback are shifted northwards in the SH, with a broad maximum from about 40 to 10°S instead of sharp peaks at ∼42°S and ∼10°S (Figure 3a). For the sake of direct comparison with the CESM-LE results, we report radiative feedback results from the CESM2 experiment with CMIP5 forcing in the remainder of this section; results from CESM2-CMIP6 are shown in Figure 2 and in Supporting Information S1.

Details are in the caption following the image

Zonal-mean radiative feedbacks (W/m2 K) and warming in the three experiments forced by CMIP5 historical and RCP8.5 emissions scenarios. Ensemble means shown. The feedbacks and warming are quantified using the difference of the 2080–2099 and 2000–2019 periods. (a) Shortwave cloud feedback; (b) Surface albedo feedback; (c) Longwave feedbacks from the radiative kernel method; (d) Zonal-mean warming. In all panels, the solid lines are CESM1-LE, the large-dashed lines are CESM1-LE*, and the small-dashed lines are CESM2-CMIP5.

In map view of shortwave feedback differences (Figure 2), CESM1-LE* minus CESM1-LE exhibits a relatively simple pattern of more positive feedback in the middle latitudes of the SH, less positive feedback in the tropics, and more positive feedback in the north Pacific basin. On the other hand, CESM2 minus CESM1-LE and CESM2 minus CESM1-LE* show complex spatial patterns. In CESM2, the region of negative shortwave cloud feedback across the Southern Ocean is expanded relative to both variants of CESM1, occurring as far northwards as 45S in the zonal mean (Figure 3a). This creates a broad region in the SH mid latitudes where the shortwave cloud feedback is less positive in CESM2 than in CESM1-LE*. A similar feature occurs in the North Pacific region. In the tropics and subtropics, CESM2 shows more positive shortwave cloud feedback off the west coasts of Australia, South America, and Africa, and more negative shortwave cloud feedback in the central Pacific and over the South American continent. To summarize, the more positive shortwave cloud feedback in CESM1-LE* compared to CESM1-LE comes almost entirely from the SH middle latitudes (particularly 35–55°S); the more positive shortwave cloud feedback in CESM2 compared to CESM1-LE or CESM1-LE* comes largely from the subtropical to middle latitudes (especially 10–40S°). While the shortwave feedback difference in CESM1-LE versus CESM1-LE* comes largely from the cloud phase effect (e.g., Frey et al., 2017), the enhanced shortwave cloud feedback in CESM2 versus CESM1-LE and CESM-LE* is associated with reduced low cloud amount (Figure S4 in Supporting Information S1).

The longwave cloud feedback sums to slightly less than zero in both variants of CESM1, but to 0.20 W M−2 K−1 in CESM2-CMIP5 (Table S1 and Figure S5 in Supporting Information S1). More positive longwave cloud feedback in CESM2 occurs in the central tropical Pacific, the tropical Atlantic and Indian Oceans, and in parts of the south Pacific Ocean basin. This more positive longwave cloud feedback closely matches the patterns of precipitation change (Figure S5 in Supporting Information S1): Generally, positive longwave cloud feedback is spatially correlated with increased precipitation. In the zonal mean, CESM2 has nearly twice the precipitation increase along the equator as both variants of CESM1.

Increased longwave cloud feedback in CESM2 makes the total cloud feedback more positive than in both variants of CESM1. This greater total cloud feedback occurs in the low latitudes, specifically from about ∼5°N to 30°S, a region that is over 25% of the global surface area. An additional positive feedback in this region is the water vapor feedback (Figure 3b), which is also greater in CESM2 than in both variants of CESM1. The lapse rate and Planck feedbacks in this region are virtually equal in all three versions of the CESM. Therefore, in the low latitudes, all of the major radiative feedbacks in CESM2 are either the same or more positive than the corresponding feedbacks in both variants of CESM1. We note that these feedback differences are not symmetric about the equator, but are rather skewed toward the SH.

Globally, the more positive water vapor feedback in CESM2 is compensated by a more negative lapse rate feedback (Table S1 in Supporting Information S1). The offset largely occurs polewards of 30°N (Figure 3b), and corresponds (along with the surface albedo feedback) with the zonal-mean temperature difference (Figure 3c). Polewards of ∼60°N, the absolute value of the lapse rate feedback is positive in all three versions of CESM, however, it is less positive in CESM2 than in both variants of CESM1. The strengthened lapse rate and water vapor feedbacks in CESM2 are consistent with the strengthened cloud feedback. As illustrated in the CESM cloud locking doubled CO2 experiments of Middlemas et al. (2020), these feedbacks tend to weaken or strengthen together: If the global cloud feedbacks are eliminated (or “locked”), then the water vapor and lapse rate feedbacks are substantially weakened as well, in line with the synergy among cloud, water vapor and lapse rate feedbacks discussed in previous work (e.g., Mauritsen et al., 2013). Also consistent with Middlemas et al. (2020), we find that the high-Arctic cloud feedbacks (70–90°N) are small in all variants of the CESM considered here, and likely do not significantly impact the local high-Arctic temperature response.

Polewards of ∼65°N, the zonal-mean surface albedo feedback is much weaker in CESM2 than in CESM1 (Figure 3d). Spatially, there are a few regions where the surface albedo feedback is stronger in CESM2, including the snow-covered regions of the Tibetan Plateau and eastern Canada (Figure S6 in Supporting Information S1). Globally, the surface albedo feedback in CESM2 is about 35% weaker than in CESM1-LE (Table S1 in Supporting Information S1). The most likely culprit for the much weaker surface albedo feedback in the Arctic in CESM2 is the thin sea ice mean state, and rapid sea ice loss, in the late 20th and early 21st centuries (DeRepentigny et al., 2020). With less sea ice to melt under increased radiative forcing and a darker initial surface, the Arctic surface albedo feedback in the 21st Century has less potency in CESM2 than in CESM1.

Is the shifted late-21st century warming pattern in CESM2—with relatively more warming in the low latitudes and relatively less in the Arctic compared to CESM1—explained by the radiative feedbacks? The results strongly suggest “yes.” The main pieces of evidence in support of this include (a) The correlation of the zonal-mean profile of warming with the zonal-mean profiles of feedbacks (Figure 3) and (b) The close resemblance between the patterns of surface temperature difference (Figure 2a) and the patterns of shortwave cloud feedback difference (Figure 2b) and surface albedo feedback difference (Figure S5 in Supporting Information S1). In the zonal mean, CESM2 has a latitudinally flatter profile of warming (i.e., less Arctic amplification) than either variant of CESM1 (Figure 3d). The enhanced warming in the low latitudes, especially in the SH, is consistent with the enhanced shortwave and longwave feedbacks in these same latitudes. In the high-latitude NH, the weaker surface albedo and lapse rate feedbacks in CESM2 reduce Arctic amplification. These relationships are confirmed in the difference maps, which show, for example, that the regions of strong positive shortwave cloud feedback difference off the western coasts of Australia, South America, and Africa (Figure 2b), are the same regions of strong temperature difference (Figure 2a). Similarly, regions where CESM2 has less positive shortwave cloud feedback than CESM1, such as the SO and the North Pacific Ocean, show the least surface temperature differences. Weaker surface albedo feedback in the Arctic corresponds closely to weaker surface warming, while stronger surface albedo feedback over the Tibetan Plateau and eastern Canada corresponds to stronger surface warming in these regions.

4 Discussion and Conclusions

Are the high values of ECS in the latest generation of Earth System Models (Zelinka et al., 2020), and the cloud feedbacks that cause them (Meehl et al., 2020; Zelinka et al., 2020), relevant for 21st-century transient warming? This study has provided a unique perspective on this question by evaluating radiative feedbacks in the transient context under realistic forcing scenarios, and by adopting a single model framework. Applying the same forcing conditions to each variant of the CESM, we conclude that enhanced shortwave cloud feedbacks alone, especially over the mid-latitudes, only result in slightly greater 21st century warming in a high-emissions scenario.

On the regional scale, we find that model structural differences have dramatic implications for the transient spatial patterns of warming. In CESM2 compared to both variants of CESM1, 21st Century warming shifts from the Arctic toward the low latitudes of both hemispheres, muting amplification in the Arctic and increasing it in the tropics and subtropics. Arctic warming is significantly weaker and tropical warming significantly stronger in CESM2 than in CESM1-LE (Figure S2 in Supporting Information S1). This pattern shift closely corresponds with enhanced low-latitude cloud feedbacks and weaker surface albedo and lapse rate feedbacks in CESM2. This evidence of structural differences is consistent with results from experiments with both models in which CO2 increases by 1% per year (Bacmeister et al., 2020). In 1% CO2 experiments, which are used to calculate the models' transient climate response (TCR, surface temperature increase at the time of CO2 doubling in a 1% per year CO2 increase experiment), CESM2 exhibits slower warming than might be expected from its high ECS.

If the enhanced low-latitude warming in CESM2 is realistic, it would exacerbate issues of climate justice, placing low-latitude human populations at disproportionate risk of climate impacts relative to their contribution to emissions. Furthermore, these risks are even more amplified in SSP5-8.5 compared with RCP8.5. Our results suggest a large scenario uncertainty, as illustrated by CESM2 warming ∼0.7 K more under SSP5-8.5 than under RCP8.5. These results are consistent with other recent studies, for example, Fyfe et al. (2021) who found similar CMIP5 versus CMIP6 differences with the CanESM. On the global scale, the scenario uncertainty is larger than the model structural uncertainty, and much larger than uncertainty due to internal variability.

We conclude that the relatively high ECS values in CMIP6 models and the cloud feedbacks that cause them do matter for 21st century warming, but the global temperature response is not the factor that matters the most. Rather, the high ECS values are indicators of model structural differences, which may matter greatly for the regional patterns of warming. Therefore, a research focus on narrowing ECS or TCR estimates may not be the best path forward to constraining 21st Century climate projections. Instead, process-based studies are needed to improve models' representation of clouds and cloud feedbacks. In this regard, a few recent studies show promise, and are relevant to the question of whether the regional warming patterns in CESM2, particularly in the low latitudes, are realistic. Cesana and Del Genio (2021) present observational evidence that shallow cumulus clouds in the tropics are insensitive to warming, unlike many models (including CESM2) simulate, implying that models overestimate the tropical low-cloud feedback. While they do find that marine stratocumulous clouds (which are the dominant type offshore of the west coasts of South America, Africa, and Australia where CESM2 exhibits reduction in cloud fraction, strong shortwave cloud feedback, and warming) are sensitive to warming, they note that models do not correctly simulate SST trends or changes in inversion strength in these regions. Myers et al. (2021) reach similar conclusions with a hybrid observational-modeling approach. Mülmenstädt et al. (2021) report on an underestimated negative cloud lifetime feedback, which could account for the difference in ECS between CMIP5 and CMIP6 generation models. Another line of research has reconsidered longwave cloud feedbacks and the original fixed anvil temperature hypothesis (Seeley et al., 2019; Yoshimori et al., 2020). All of this raises the possibility that positive low-latitude shortwave cloud feedbacks and perhaps longwave feedbacks in CESM2 are overestimated. In addition, high-latitude warming in CESM2 shows sensitivity to the local surface albedo feedback. We suggest that this feedback may be underestimated in CESM2 due to biases in Arctic sea ice, which is the main driver of the surface albedo feedback in the high Arctic (Danabasoglu et al., 2020; DeRepentigny et al., 2020).

A better understanding of the mechanisms behind the spatial patterns of surface temperature change in models is crucial for resolving uncertainty in the “pattern effect,” namely the dependence of climate sensitivity estimates on sea surface temperature trends (e.g., Rugenstein et al., 2016; Xie, 2020). Over the past ∼30 years, observed and modeled SST patterns generally do not agree. While internal variability or missing/mis-represented forcings might be primary causes of the mismatch (e.g., Xie, 2020), these results suggest that there is still a large structural uncertainty in current-generation models, owing at least in large part to cloud feedbacks.

Acknowledgments

D.P.S. and J.E.K. were supported at the University of Colorado/CIRES by the National Science Foundation (NSF) grants 1643493 and CAREER 1554659. D.P.S. and C.H. were supported by the National Center for Atmospheric Research (NCAR), which is a major facility sponsored by the NSF under Cooperative Agreement no. 1852977; D.P.S. received additional support from NSF grant 1952199. We thank Marika Holland for leading the experiments with CESM2 and CMIP5 forcing. Computing and data storage resources were provided by the Computational and Information Systems Laboratory (CISL) at NCAR. The CESM project is primarily supported by NSF. We thank all the scientists, software engineers, and administrators who contributed to the development of CESM1 and CESM2, as well as two anonymous reviewers who provided thoughtful and constructive comments.

    Data Availability Statement

    The CESM1-LE data used in this study are available through https://doi.org/10.5065/d6j101d1. The CESM1-LE* data used in this study are available on the Climate Data Gateway or through https://doi.org/10.26024/3nvz-vt03. The CESM2-CMIP5 data are available at the Climate Data Gateway or through https://doi.org/10.26024/4zgv-rt74. The CESM2-CMIP6 historical data are available through https://doi.org/10.22033/ESGF/CMIP6.7627. The CESM2-CMIP6 SSP5-8.5 data are available through https://doi.org/10.22033/ESGF/CMIP6.7768.