CMIP6 Models Predict Significant 21st Century Decline of the Atlantic Meridional Overturning Circulation

We explore the representation of the Atlantic Meridional Overturning Circulation (AMOC) in 27 models from the CMIP6 multimodel ensemble. Comparison with RAPID and SAMBA observations suggests that the ensemble mean represents the AMOC strength and vertical profile reasonably well. Linear trends over the entire historical period (1850–2014) are generally neutral, but many models exhibit an AMOC peak around the 1980s. Ensemble mean AMOC decline in future (SSP) scenarios is stronger in CMIP6 than CMIP5 models. In fact, AMOC decline in CMIP6 is surprisingly insensitive to the scenario at least up to 2060. We find an emergent relationship among a majority of models between AMOC strength and 21st century AMOC decline. Constraining this relationship with RAPID observations suggests that the AMOC might decline between 6 and 8 Sv (34–45%) by 2100. A smaller group of models projects much less AMOC weakening of only up to 30%.


Introduction
The Atlantic Meridional Overturning Circulation (AMOC) is an important regulator of the climate system. It plays a key role in the global redistribution of heat, as it accounts for roughly 2/3 of the oceanic northward heat transport at northern midlatitudes (Johns et al., 2011;Trenberth et al., 2019) and is responsible for net equatorward ocean heat transport in the South Atlantic (estimated to be of the order of 0.5 PW Dong et al., 2009;Trenberth et al., 2019), a unique feature among ocean basins. Not only does the AMOC influence the subpolar North Atlantic (e.g., Palter, 2015) but its impacts include the position of the Intertropical Convergence Zone (ITCZ; e.g., Marshall et al., 2014); Arctic sea ice (e.g., Day et al., 2012;Mahajan et al., 2011;Yeager & Robson, 2017); and the global response to Arctic sea ice loss (e.g., Tomas et al., 2016). The AMOC is also a significant driver of carbon (e.g., Carlson et al., 2010;Fontela et al., 2016) and heat (e.g., Häkkinen et al., 2015;Kostov et al., 2014) sequestration in the subsurface ocean. Clearly, the response of the AMOC to continued anthropogenic forcing will strongly determine the trajectory of Earth's climate in the 21st century and beyond.
An accurate representation of the mean state, variability, and sensitivity to anthropogenic forcing of AMOC is a critical test for climate models and is essential for building confidence in the projections they produce. The various phases of the Coupled Model Intercomparison Project (CMIP) have given us an excellent These observations provide useful reference values for evaluating the representation of the AMOC in models. Furthermore, we will examine whether the multimodel data exhibit any "emergent constraint" relationships (Hall et al., 2019), in particular, whether the modeled future decline of the AMOC in the 21st century is related to their historical mean states as suggested by previous generation CMIP models (Gregory et al., 2005;Weaver et al., 2012). Such a relationship would allow us to calibrate the AMOC decline in the 21st century using the present-day observations, thus providing a best-guess quantification of the AMOC strength into the future.

CMIP6 Models Used in This Analysis
We analyze CMIP historical simulations  and four scenarios from the Scenario Model Intercomparison Project (ScenarioMIP; O'Neill et al., 2016). The protocol for the historical simulations (experiment name: "historical") calls for 165-year simulations (from 1850 to 2014), with forcing fields that are based on observations whenever possible (Hoesly et al., 2018;Meinshausen et al., 2017). The scenario simulations are run for 86 years, from 2015 to 2100. ScenarioMIP recommends four tier-1 simulation protocols, reflecting different Shared Socioeconomic Pathways (SSPs) that result in different radiative forcing magnitudes by 2100. These experiments represent low-end (SSP1-2.6: "ssp126"), medium-end (SSP2-4.5: "ssp245"), and high-end (SSP3-7.0: "ssp370"; SSP5-8.5: "ssp585") forcing scenarios. We also use 4xCO 2 ("abrupt4xCO2") simulations and historical simulations in which only the greenhouse gas concentrations ("hist_GHG") or aerosols ("hist_aer") were varied. We use output from 27 different models, as listed in Table 1 and Tables S1 and S2 in the supporting information. These models had submitted the stream function variable (msftmz and msftyz) for the historical simulations to the ESGF archive by the time of our data freeze (15 January 2020). Eighteen models contributed SSP simulations, four of which contributed only a subset of the tier-1 simulations.
We define AMOC strength as the maximum of the overturning stream function in the Atlantic Ocean and evaluate AMOC characteristics at 26 • N and 34 • S. The term "historical period" is understood to be the 1850-2014 period of the CMIP6 historical simulations, while the "RAPID period" is defined here as the 2005-2014 period of overlap between the historical period and the years for which full year's worth of RAPID data is available (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017).
In this paper we use standard error of the mean to quantify the uncertainty in the estimation of the mean based on a limited number of samples (ensemble members or time series points). It is defined as ∕ √ n,  where n is the number of samples and is the standard deviation of the sample. There is roughly a 68% (95%) probability that the true mean falls within the range ± ( ± 2 ).
The AMOC at 34 • S in the South Atlantic has been measured by the South Atlantic MOC Basin-wide Array Comparison of model-simulated overturning stream function with OSNAP observations is not attempted here. OSNAP reports its overturning estimates in density coordinates (Lozier et al., 2019), and only a few models provided their stream function as such. In addition, OSNAP measures the overturning circulation in two sections that are far from zonal, and it is not straightforward to compare these observations with zonally integrated transports reported by the models.

Validating the Historical AMOC
First, we assess the representation of the AMOC in the CMIP6 models for the historical period. All 27 models simulate a well-defined overturning circulation, associated with North Atlantic Deep Water (NADW) formation in the subpolar North Atlantic Ocean (Figures 1 and S1). At 26 • N, several models capture the observed depth of the AMOC maximum (z max ) at 1.03 km quite well (Figure 1a), but most have a shallow bias that puts the ensemble-mean depth of the AMOC maximum at 0.84 ± 0.03 km (Table S3). A shallow bias is seen at 34 • as well, when compared to estimates from the SAMBA array ( Figure 1b). In most models, the depth of the NADW cell at 26 • N (defined as the depth z 0 of the zero crossing of the overturning stream function) is between 2.5 and 3.5 km, and the ensemble mean of 3.12 ± 0.14 km is too shallow compared to the RAPID observations (z 0 = 4.40 km). All models display a zero crossing at 34 • S, indicating a clear separation between the NADW and AABW cells. Some models display another reversed circulation cell at depth (e.g., INM-CM4-8 at 26 • N and HadGEM3-GC31-MM at 34 • S).
The ensemble mean overturning strength at 26 • N is 17.7 ± 0.8 Sv, within one standard error of the RAPID mean (Table S3), but the spread among models ranges from 9.6 Sv (E3SM-1-1; −1.8 , where is the standard deviation of the ensemble spread) to 23.04 Sv (FGOALS-f3-L; +1.1 ). The data-assimilated model NorCPM1 is the extreme outlier (+3.0 ). With an ensemble mean of 16.8 ± 0.7 Sv, CMIP6 models simulate a South Atlantic AMOC that is stronger than the 14.6 Sv that has been inferred from the SAMBA array at 34 • S (Meinen et al., 2018) but weaker than some other estimates in the 18-19 Sv range (Dong et al., 2009(Dong et al., , 2015Garzoli et al., 2013). But again, there are significant outliers on both the weak (E3SM; −1.7 ) and strong (NorCPM1; +2.2 ) side of the mean. The strength of the Antarctic Bottom Water (AABW) cell in the CMIP6 models at 26 • N ranges considerably between 0.0 (SAM0-UNICON; −1.1 ) and −8.5 Sv (FGOALS-f3-L; +4.2 ), but the ensemble mean is only slightly stronger than observations (−1.8 ± 0.3 Sv compared to −1.0 Sv).

AMOC Trends in Historical Simulations and 21st Century Projections
The time series of AMOC strength at 26 • N and 34 • S illustrate the long-term variability ( Figure S2). Many models-including the ensemble mean-show a gradual increase in AMOC strength over the earlier part of 20th century, followed by a rapid decline starting in the 1980s, a feature that was not seen in the CMIP5 ensemble (Cheng et al., 2013;Weaver et al., 2012). Even though the ensemble mean AMOC trend over the historical period at 26 • N is indistinguishable from zero (Figure 2a), there is a clear compensation between positive AMOC trends up to the 1980s (1.3 ± 0.2 Sv) and negative trends in the final decades of the historical period (−1.2±0.2 Sv; Figure S3). In fact, only four out of 26 models show a weakening trend between 1850 and 1980. Figure S4 shows that this behavior reflects a compensation between AMOC response to a continuing increase in greenhouse gasses, and to a late-20th century peak in sulfate aerosol forcing (Delworth & Dixon, 2006;Menary et al., 2013).
The 17 models that submitted at least two of the four tier-1 ScenarioMIP simulations all show a decline in AMOC strength over the 21st century ( Figure 3). Surprisingly, all models show a rate of decline that is independent of the scenario at least through 2060, while even after 2060 only a few models show an appreciable divergence of the trajectories. This behavior is despite the significant differences among the SSPs in emission scenarios for the radiatively active species (Gidden et al., 2019) and will need more careful evaluation. We note that it is consistent with the emerging notion that many climate variables are proportional to the cumulative carbon emissions (Herrington & Zickfeld, 2014;Notz & Stroeve, 2016;Steinacher & Joos, 2016), implying that the AMOC in the first decades of the SSP scenarios is mostly determined by historical CO 2 emissions. Only after the cumulative CO 2 contributions of the SSP scenarios start to become significantly different can the models be expected to diverge. Whether they diverge or not depends on the relative sensitivity of AMOC to CO 2 and the other forcing agents.
Previous model intercomparison studies found that AMOC decline under CO 2 forcing scenarios is stronger in models that have a stronger mean state (e.g., Gregory et al., 2005;Gregory & Tailleux, 2011;Levermann et al., 2007;Weaver et al., 2012;Winton et al., 2014). Indeed, there is weak but statistically significant tendency for models with a weak (strong) mean AMOC to display a positive (negative) trend over the historical period ( Figure 2a). However, at face value, this relationship does not seem to be reproduced by the full complement of CMIP6 models considered for SSP simulations (Figures 2c-2f). A closer examination of these plots reveals the presence of outliers. Indeed, a majority of models display a statistically significant linear relationship between projected 21st century AMOC decline and the AMOC mean strength, and similar behavior is also seen in an analysis of AMOC trends in abrupt 4xCO 2 experiments (Figure 2b). A formal cluster analysis presented in section S5 confirms both the existence of outliers and the statistical significance of the linear relationships.

AMOC and Effective Climate Sensitivity
Recent studies show that the effective climate sensitivity (ECS) of CMIP6 models (defined as the amount of warming resulting from a doubling of the atmospheric CO 2 concentration) is higher than in CMIP5 (Zelinka et al., 2020). Here we document how different AMOC metrics are related to ECS, as determined by  (Table S3). Figure 4a shows a significant inverse relationship between ECS and mean AMOC strength, as models with lower (higher) ECS tend to have stronger (weaker) AMOC mean strength. As could be expected from the linear relationships between mean AMOC and historical AMOC change on the one hand (Figure 2a), and between mean AMOC and ECS on the other hand (Figure 4a), the historical AMOC trends are also proportional to ECS (Figure 4c), with models with low ECS generally displaying negative trends (CO 2 dominated), while models with high ECS tend to display positive trends (aerosol dominated). We expected a similar relationship between ECS and AMOC decline for the 4xCO 2 and SSP experiments, but no such relationship is apparent (Figures 4b and 4d), even if we limit ourselves to the majority cluster.

Discussion and Conclusions
We compare the representation of the AMOC in historical simulations from the CMIP6 multimodel archive to observations from the North (RAPID) and South (SAMBA) Atlantic and study projected AMOC changes in scenario simulations. Our analysis shows that the ensemble mean AMOC strength at 26 • N of 27 CMIP6 models agrees within error with RAPID observations, while the AMOC in the South Atlantic is biased strong compared to SAMBA (by 2 Sv). However, model biases persist, including a multimodel-mean NADW cell that is too shallow. Besides, there is still a significant spread in the representation of the AMOC among the models.
The CMIP6 ensemble does not show a significant AMOC trend over the historical period from 1850 to 2014, but many models (and the ensemble mean) display a local maximum in the later part of the 2th century. This behavior reflects a combined effect of the accelerating increase in CO 2 concentrations, and a sulfate aerosol load that peaks in the 1980s, before declining in the 1990s and 2000s (Delworth & Dixon, 2006;Menary et al., 2013). This is at odds with CMIP5 models (Cheng et al., 2013;Weaver et al., 2012), which generally show a weak decline over the 20th century, and with reconstructions based on proxies (Caesar et al., 2018;Rahmstorf et al., 2015;Thornalley et al., 2018).
All models show a decline in the AMOC in the 21st century, with a rate of decline that is independent of emission scenario at least through 2060; only a few models show appreciable divergence between the SSPs thereafter. On average, AMOC decline is 24%, 29%, 32%, and 39% for SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5, respectively. Cheng et al. (2013) and Schleussner et al. (2014) found that CMIP5 models on average displayed a 9% decline for the RCP2.6 scenario (comparable to SSP1-2.6); 21% for RCP4.5 (SSP2-4.5); and 36% for ). The CMIP6 model generation hence projects a stronger decline than CMIP5, with the discrepancy being most pronounced for the scenarios with weakest radiative forcing.
We show that there are two groups of contrasting model behavior, with the majority of models (75% of the 16 models that provided both ssp585 and abrupt4xCO2 simulations), displaying a strong and significant linear relationship between mean present-day AMOC strength and projected 21st century AMOC decline. We can use this "emergent constraint" relationship to calibrate this subensemble with observed mean AMOC state from the RAPID/MOCHA array. For this group, even the low-end scenario (SSP1-2.6, aggressive mitigation) is projected to have a 6 Sv (34%) weaker AMOC by the end of this century (with a range of 2.5 to 9 Sv based on the 95% prediction interval of the linear fit), while the high-end SSP5-8.5 scenario projects a 8 Sv (45%) decline (with a range of 6 to 11 Sv). However, the smaller group of models projects a more moderate AMOC decline of at most 5 Sv (28%) by the end of the 21st century.
Our results support the emerging body of literature that suggests relationships between AMOC mean strength and the role of the AMOC in the climate response to radiative forcing (e.g., Garuba et al., 2018;Rugenstein et al., 2013;Trossman et al., 2016;Winton et al., 2013Winton et al., , 2014. Levermann et al. (2007) argue that the positive relationship between mean AMOC and AMOC decline under CO 2 forcing ( Figure 2) is mediated by sea ice. In their view, models with weaker AMOC have generally more Arctic sea ice than models with a stronger AMOC, as a result of the weaker northward heat transport. As the atmosphere warms under enhanced radiative forcing, models with weaker AMOC experience the strongest sea ice retreat, exposing more ocean to the atmosphere. The resulting ocean heat loss counteracts the AMOC weakening induced by atmospheric warming. At the same time, the models with stronger AMOC decline experience stronger reduction in northward heat transport, cooling the subpolar North Atlantic, with implications for sea ice and cloud feedbacks (

10.1029/2019GL086075
This provides a negative feedback on Arctic warming and lowers ECS, consistent with Figures 4a and 4c. It is not obvious why the AMOC decline in 4xCO 2 and SSP simulations does not conform to this picture. However, we speculate that the framework describing the subtle interplay between AMOC and high-latitude feedbacks is no longer valid when large-amplitude forcing perturbations are applied and sea ice is no longer a dominant element in the subpolar North Atlantic and Nordic Seas.
Most models investigated in this study have a nominal ocean resolution of 1 • . It is well known that several aspects of the formation (e.g., Pickart et al., 2005), downwelling (e.g., Katsman et al., 2018), and upwelling (e.g., Tamsitt et al., 2017) of deep waters depend on small-scale processes that cannot be explicitly represented by these models. Whether this has biased our estimate of the sensitivity of the AMOC to climate forcing is an important question that will need to be addressed using century-long scenario experiments with high-resolution models. Here we note that the few models that submitted simulations at higher resolutions (CNRM-CM6, MPI-ESM1-2, and HadGEM3-GC31) do not stand out from the rest of the ensemble, consistent with the conclusion from Winton et al. (2014) that spatial resolution in itself does not impact AMOC sensitivity to radiative forcing. its Working Group on Coupled Modelling, coordinated and promoted CMIP6. We thank the Earth System Grid Federation (ESGF) for archiving the data and providing access; the multiple funding agencies who support CMIP6 and ESGF; DOE's RGMA program area, the Data Management program, and NERSC for making this coordinated CMIP6 analysis activity possible; and the modeling centers for generously sharing their CMIP6 model output. CMIP6 model data are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License (https:// creativecommons.org/licenses) and can be accessed through the ESGF nodes (for instance esgf-node.llnl.gov/ search/cmip6). We like to thank three anonymous reviewers for their constructive comments and Jian Lu (PNNL) and Christopher Meinen (NOAA/AOML) for useful discussions. Data from the RAPID/MOCHA program and the RAPID-WATCH MOC monitoring project are funded by the Natural Environment Research Council and the U.S. National Science Foundation and are freely available online from (www.rapid.ac.uk/ rapidmoc). Data from the Southwest Atlantic MOC project (SAM) are funded by the DOC-NOAA Climate Program Office-Ocean Observation and Monitoring Division and the Atlantic Oceanographic and Meteorological Laboratory. These data are freely available online (via www. aoml.noaa.gov/phod/research/moc/ samoc/sam). This research was supported by the Regional