Anthropogenic Weakening of the Atmospheric Circulation During the Satellite Era

Climate models predict a slowing of the atmospheric overturning circulation with warming. In models, this slowing manifests primarily as a weakening of the Walker Circulation (WC). However, observational studies indicate a strengthened Pacific WC over the past several decades, raising questions about the models' ability to represent critical energetic and hydrologic constraints responsible for the predicted weakening. This discrepancy is closely tied to differences in the warming pattern over the Pacific during this period. We show that model simulations with either observed or model‐projected warming patterns predict a robust weakening of atmospheric overturning circulation, despite having opposing changes in the Pacific WC strength. This weakening occurs in the zonally asymmetric circulation, rather than in the zonal‐mean Hadley cell. Weakening inferred from satellite observations is reproduced in coupled models only when anthropogenic forcing is included, suggesting that a human‐induced weakening of the global atmospheric circulation is already detectable in observations.


Plain Language Summary
The atmospheric circulation transfers moisture and energy from the tropics to the polar regions and regulates the distribution of rainfall in the tropics which is home to around 40% of the world's population.In recent years, observations show a strengthening of the Pacific Walker Circulation (WC), a regional-level circulation, which contrasts the weakening of the circulation predicted by climate models.This discrepancy questions the climate models' ability to predict future changes in overturning circulations.Using various strength indices, we find a consistent "weakening with warming" of the global overturning circulation in both observations and climate models for the common period, regardless of the behavior of regional-level circulations.Moreover, we find that it is highly unlikely that this observed weakening is due to natural variations in the climate.

Anthropogenic Weakening of the Atmospheric Circulation During the Satellite Era
Sisam Shrestha 1 and Brian J. Soden 1 1 Rosenstiel School of Marine, Atmospheric, and Earth Science, University of Miami, Coral Gables, FL, USA Key Points: • The global atmospheric overturning circulation has weakened in recent decades despite a strengthening of the Walker Circulation • Climate models predict a circulation weakening consistent with observations, manifested primarily as a weakening of the zonally asymmetric overturning circulation • The observed weakening is reproduced in coupled climate models only when anthropogenic forcing is included

Supporting Information:
Supporting Information may be found in the online version of this article.et al., 2021).At a millennium scale, paleo-proxy reconstructions show insignificant shift in long-term Pacific WC trend between the pre-and post-industrial onset era (Falster et al., 2023).
The Pacific WC strength is closely tied to the east-west sea surface temperature (SST) gradient over the equatorial Pacific (Sandeep et al., 2014).Observed SST trends over recent decades indicate warming in the western Pacific and cooling in the eastern Pacific, enhancing the east-west SST gradient (Figure 1a and Figure S1 in Supporting Information S1).However, coupled climate models forced with observed changes in greenhouse gases project a reduction in the east-west SST gradient as response to historical radiative forcings (Figure 1b and Figure S1 in Supporting Information S1).These contrasting patterns of surface warming drive contrasting responses of the atmospheric circulation; coupled ocean-atmosphere models simulate a weakening of the Pacific WC over this period, while observations indicate a strengthening (Figure S2 in Supporting Information S1).This robust discrepancy between the coupled model simulations and the observed trends thus reflects either a fundamental problem in model projection to the forced response or a systematic underestimate in their simulation of the Pacific decadal variability (Lee et al., 2022).
Early attempts to reconcile this discrepancy pointed to internal variability on decadal time scales as a potential explanation (Chung et al., 2019;Power et al., 2021;Watanabe et al., 2021;Wu et al., 2021).This would imply that the strengthening of the Pacific WC in recent decades reflects the appearance of low-frequency modes of natural variability, such as the Interdecadal Pacific Oscillation (IPO) or Atlantic Multidecadal Oscillation (Bordbar et al., 2017;Murphy et al., 2017).
However, as the divergence between observed and model-projected trends has persisted, this explanation has become increasingly unlikely (Seager et al., 2022;Wills et al., 2022).Strengthened Pacific trade winds during the early 2000s global warming hiatus was stronger than that associated with a negative IPO phase and outside the range of model simulations (England et al., 2014).Recent studies that examine large ensembles of model simulations indicate that the persistence of observed warming pattern over the Pacific lies at the extreme limit of 10.1029/2023GL104784 3 of 10 model behavior and is therefore unlikely to result from internal variability (Seager et al., 2022).This would imply that the observed trends in SST and the Pacific WC are actually a forced response to increasing greenhouse gases.However, due to their coupled nature, it's unclear if a strengthened Pacific WC drives the asymmetric pattern of SST change, or vice-versa.This, in turn, raises questions about the validity of model projections of a weakened atmospheric circulation.
Changes in the atmospheric overturning circulation have important implications for the pattern of precipitation change, as well as the frequency of droughts and extreme precipitation events (Chadwick et al., 2013;Lau & Kim, 2015;Su et al., 2014Su et al., , 2019;;Wodzicki & Rapp, 2020).Atmospheric overturning circulation and SST pattern are closely tied to the climate feedback from tropical low cloud covers and thus has a strong impact on climate sensitivity (Andrews et al., 2022;Loeb et al., 2020;Sherwood et al., 2014).Understanding how atmospheric overturning circulations have behaved in recent decades is thus imperative to predict their evolution under future conditions and to formulate mitigation and adaptation policies related to climate change.

Strength Indices for Large-Scale Atmospheric Circulation
We employed three different estimates to quantify the strength of atmospheric overturning circulations.The first index used is the change in the convective mass flux (ΔM c ) which is also the preferred value to detect changes in atmospheric circulation as it is a direct measure of convection.In a warming world, the differential rate of increase between boundary-layer moisture and precipitation requires a weakening of large-scale atmospheric circulations at a rate of ∼5%/K (Held & Soden, 2006).However, most CMIP6 models used in this study have not archived their convective mass flux output.Following Held and Soden (2006), we estimated convective mass flux as where P is the precipitation and q is the boundary-layer specific humidity.For models that archived the convective mass flux,  ΔM * c agrees well with the model-simulated ΔM c at 500 hPa (Figure S3 in Supporting Information S1).We find similar results for the vertically integrated ΔM c (surface to 30 hPa) since the vertical profiles of convective mass flux don't show a significant change during this period.
The second index used is the fractional change in the upward component of the mid-tropospheric vertical pressure velocity (ω 500 + ) (Bony et al., 2013;Vecchi & Soden, 2007).As the circulation weakens, following the simple mass-balance theory, tropical-mean ω 500 + is also expected to weaken (Vecchi & Soden, 2007).We used a daily ω 500 data set to construct a monthly data set of only ω 500 + .
The third index we employed is the change in the spatial variance for ω 500 which should decrease at twice the rate of the mean, provided the weakening of the vertical velocity is proportional to the preexisting vertical velocity field (Held & Soden, 2006).The spatial variance index can further be separated into a component related to the zonal-mean circulation, for example, Hadley circulation, and the zonally asymmetric circulation, like the WC.Since changes in the atmospheric circulation are expected to manifest in both ascent and descent regions (Su et al., 2014), we consider changes in the spatial variance of ω 500 instead of just ω 500 + as in previous studies (Held & Soden, 2006;Vecchi & Soden, 2007).Time series were deseasonalized using a base period of 1990-2010.Fractional changes are computed with reference to the first 10 years of the time series.We define the tropics as the region bound by the 30° latitude band.Trends were computed using least squares linear regression.A negative trend in the timeseries implies a weakening of the circulation.Uncertainty was calculated using a two-sided Student's t-test, considering the autocorrelation of noise in a time series (Weatherhead et al., 1998), and can be understood as a reflection of the interannual variability.Observed trends are reported with ±2 Standard Errors (SE) and can be considered significant at the 95% confidence interval if it is greater than ±2 SE.

Climate Models
To analyze changes in recent decades, we employed 25 coupled ocean-atmosphere models forced with historical radiative forcings (1850-2014) from the Coupled Model Intercomparison Project Phase 6 (CMIP6, Eyring et al., 2016) along with the corresponding atmosphere-only simulations integrated using observed SSTs as a lower boundary condition from the Atmospheric Model Intercomparison Project (AMIP, Gates et al., 1999Gates et al., ) (1979Gates et al., -2014) ) (Table S1 in Supporting Information S1).To analyze the contribution of anthropogenic forcings on the observed trends in atmospheric overturning circulation, we extended the coupled historical experiments to 2021 using the ssp5-8.5ScenarioMIP (O'Neill et al., 2016), which represents a high-emission future pathway that produces a radiative forcing of 8.5 W m −2 in 2100.We also used 90 members from the CESM2 Large Ensemble Project (CESM2-LENS; Rodgers et al., 2021) and the 10-member Community Atmosphere Model version 6 prescribed SST Global AMIP (CAM6 AMIP) (Danabasoglu et al., 2020) ensemble that extends the prescribed SST AMIP experiments to 2021.To analyze the contribution of unforced climate variability on the observed trend, we used preindustrial control (piControl) experiment that is representative of conditions before the onset of large-scale industrialization in 1850.We only used the first ensemble member (r1i1p1f1) from each model, except for the large ensembles, to ensure all the models are weighted equally in the multi-model mean (MMM) calculation.

Observations
For the estimation of changes in the observed atmospheric overturning circulation, we used precipitation observations from Global Precipitation Climatology Project v2.3 (GPCP v2.3, Adler et al., 2003) which combines precipitation from satellites, rain gauges and soundings to provide a global rainfall record.GPCP data before 1988 is, however, limited to IR-based estimates.The monthly data set runs from 1979 to the present day and is provided in a 2.5° × 2.5° grid.For boundary-layer moisture, we used total precipitable water (TPW) from the Merged Precipitable Water 1° Monthly Climate Product from Remote Sensing Systems (Remote Sensing Systems, 2016).The TPW data set has been constructed using intercalibrated microwave radiometers that provide data from 1988 to the present day over global ice-free oceans.We used European Centre for Medium-Range Weather Forecasts v5 reanalysis (ERA5) data set (Hersbach et al., 2023) to fill in the data gaps and construct a global TPW data set.Both the TPW data sets have a similar interannual trend over ice-free oceans for the common period (1988-2021) (Figure S4 in Supporting Information S1; Allan et al., 2022).TPW provides a good estimate of the near-surface specific humidity since much of the atmospheric water content is concentrated in the boundary layer, as seen from the strong positive correlation between the fractional changes in near-surface specific humidity and TPW (Figure S5 in Supporting Information S1).However, since TPW is a column integral, it tends to overestimate the change in boundary-layer specific humidity (and thus ΔM c ) with warming since the fractional increase in saturation vapor pressure is inversely proportional to temperature, that is, dln(e s )/dT ∼ 1/T 2 .Thus, we use a linear regression between monthly anomalies in near-surface specific humidity and TPW from the AMIP MMM to derive an observational estimate for near-surface specific humidity (Figure S5 in Supporting Information S1).We used NASA GISS Surface Temperature Analysis version 4 (GISTEMP v4) to estimate global surface temperature change (GISTEMP Team, 2023;Lenssen et al., 2019).

Model Projections of Pacific Warming and Large-Scale Atmospheric Circulation Change
Difference in the SST trend patterns between the coupled and the AMIP simulations is reflected in their respective ω 500 trends (Figures 1c and 1d).The AMIP models simulate an increased ascending motion in the western Pacific and an increased subsidence in the eastern Pacific, consistent with a strengthened Pacific WC.This increase in the equatorial Pacific ω 500 gradient is consistent with observations of SLP, water vapor transports and reanalysis data sets (Chung et al., 2019;Sohn & Park, 2010).In contrast, the CMIP6 coupled models simulate an opposing pattern, that is, a weakening of the equatorial Pacific ω 500 gradient, reflecting a weakened Pacific WC.
To investigate the role of SST patterns on the atmospheric overturning circulation, we analyze trends in three different indices used to quantify the strength of the global-mean atmospheric circulation.For brevity, we focus on the  ΔM * c index here, but similar results are found for the other two strength indices (Figure S6 in Supporting Information S1).For the coupled historical experiments, all three indices indicate a weakening of the atmospheric overturning circulation over the 1979-2014 period (Figure 2 right), consistent with the expected "weakening with warming" mechanism.Despite a large intermodel spread in the historical ensemble (attributable to the different interannual variability amongst models), all the members capture the temporary cooling associated with the aerosols released into the stratosphere following the Mt.Pinatubo eruption in 1991.This period of cooling corresponds with a brief strengthening of the atmospheric overturning circulation in all the models.
Although the AMIP simulations markedly differ from the coupled historical experiments in terms of the pattern of circulation response (Figure 1c), all models simulate a weakening atmospheric overturning circulation for this period (Figure 2 left).The AMIP MMM weakening falls within the historical intermodel spread, however, the magnitude of weakening is smaller for the AMIP experiments, in part due to the smaller global-mean temperature change in AMIP compared to the coupled model simulations.The MMM time series shows interannual variations associated with ENSO.Short-term weakening (strengthening) of global-mean circulation corresponds with warming (cooling) associated with major El Niño (La Niña) events like 1982-1983 and 1997-1998 (1988-1989 and 2000).

Changes in ω 500
+ index and the spatial variance of ω 500 also indicate a weakening of the global-mean atmospheric circulation for this period in both the simulations.Changes in the spatial variance of ω 500 are dominated by the weakening of the zonally asymmetric component of the circulation for both experiments (Figures 2c and 2d).The result that models preferentially weaken the asymmetric circulation rather than the zonal-mean circulation was first pointed out by Held and Soden (2006) and Vecchi and Soden (2007).Although not demonstrated, the possible explanation given by those studies was that zonal-mean circulation is constrained by momentum and meridional energy transport constraints, whereas the zonally asymmetric circulation is not.Thus, despite different warming patterns and contrasting responses of the Pacific WC, both the atmosphere-only and coupled historical experiments simulate a weakening of the global-mean atmospheric circulation, almost exclusively by a reduction in the zonally asymmetric component of the overturning circulation rather than the zonal-mean.In the coupled historical experiments, this weakening of the zonally asymmetric overturning circulation primarily manifests as a reduction in the Pacific WC.However, in the AMIP simulations, the weakening of the zonally asymmetric circulation occurs despite a strengthening of the Pacific WC.Thus, it is not necessary for the Pacific WC to weaken in order for models to simulate a weakening of the zonally asymmetric overturning circulation; other modes of the zonally asymmetric circulations (e.g., monsoons) are presumably responsible for this weakening.

Changes Inferred From Observations
Due to the lack of a global observational record for ω 500, we quantify observed changes in atmospheric overturning circulation using satellite measurements of precipitation and water vapor to compute  ΔM * c following the same approach used for the models.The observation analysis begins in 1988 when microwave satellite measurements of TPW and precipitation first became available.
Consistent with the model simulations, observations also indicate a weakening of the global-mean atmospheric circulation.The magnitude of weakening (−0.51 ± 0.3%/decade) is comparable to the AMIP MMM weakening (−0.65 ± 0.37%/decade) but noticeably smaller than the weakening projected in the MMM of the coupled model simulations (−1.11 ± 0.51%/decade) for the 1988-2014 period.This discrepancy arises partly from the different rates of warming between observations and coupled model simulations.To account for the different warming rates, we compute the change in convective mass flux per unit temperature change, by dividing the trends in convective mass flux by trends in global-mean surface air temperature.The normalized changes in convective mass flux are comparable between the observed (−3.02%/K),AMIP (−4.64 ± 2.64%/K), and coupled model (−3.7 ± 0.83%/K) simulations.The larger spread in the AMIP is mainly due to an outlier, FGOALS-g3 which has a near-zero global-mean temperature trend and simulates a large weakening of the global overturning circulation (∼−10%/K) and an unusually large strengthening of the tropical overturning circulation (∼25%/K).Removing the FGOALS-g3 model results in an AMIP weakening of −4.58 ± 1.13%/K.
To better understand the contributions of anthropogenic forcing to the observed trend in ΔM c , we extend the historical coupled simulations to 2021 using the ssp5-8.5ScenarioMIP.Using an alternative ScenarioMIP to extend the data set does not produce a significantly different result, as the ScenarioMIPs do not diverge significantly during the first few years.The extended observational record indicates a robust weakening of both the global-(−0.95± 0.21%/decade) and the tropical-mean (−0.63 ± 0.28%/decade)  ΔM * c (Figure 3).Compared to the CMIP6 ensemble, where differing model physics contributes to the ensemble spread, large ensembles like CESM2-LENS have the advantage of sharing the same model physics.Thus, the CESM2-LENS average and spread can be considered as the forced response and internal variability, respectively.Despite their different phases of internal variability, all the CESM2-LENS members exhibit a robust weakening of the atmospheric overturning circulation (Figure 3).While the CESM2-LENS spread captures the observational estimate of the weakening, the ensemble-mean weakening is stronger than the observational estimate (Figure S7 in Supporting Information S1), which may reflect the higher sensitivity of CESM2-LENS ensemble average (Figure 4).The historical MMM has a similar magnitude of weakening for both the 1988-2021 and 1979-2014 period (blue and pale blue lines, Figure 4).
The 10-member CAM6 AMIP ensemble also shows a robust weakening of the atmospheric overturning circulation (Figure 3).The magnitude of weakening for the CAM6 AMIP average is comparable to the observed global-mean weakening, highlighting the robustness of the weakening signal of atmospheric circulations in recent decades (Figure S7 in Supporting Information S1).The observed tropical-mean weakening, however, lies on the extreme of simulated trends, suggesting models' inability to simulate observed weakening in the tropics.We find a correlation of 0.7 and 0.4 between the observational estimate and CAM6 AMIP average for global and tropical convective mass flux change, respectively.The lower correlation in the tropics (also present for the shorter AMIP experiment) primarily stems from the low correlation between the observed and simulated precipitation, which presumably reflects the role of internal atmospheric variability on year-to-year changes in precipitation.
To demonstrate that internal variability alone cannot explain the observed weakening of large-scale atmospheric circulations in recent decades, we analyze 34-year composites from the CMIP6 piControl runs (Figure 3 and Figure S7 in Supporting Information S1).The piControl ensemble is equally divided into strengthening and weakening trends, suggesting internal variability alone can result in either weakening or strengthening of the overturning circulation but cannot explain the magnitude of weakening observed in recent decades.

Conclusions and Discussion
In this work, we utilize satellite-based observations and the latest version of climate models to analyze trends in atmospheric overturning circulation during the satellite era.For all three indices used, we find a robust weakening  of the atmospheric overturning circulation for both the AMIP and the coupled historical experiments.Further analysis of single-model large ensembles also shows a weakening of the atmospheric overturning circulation with warming.A comparable magnitude of global-mean overturning circulation weakening is seen in observations, within the range of CESM2-LENS, suggesting a role of anthropogenic forcing in the observed weakening in recent decades.In contrast, the piControl runs suggest that internal variability alone cannot explain the magnitude of the observed atmospheric overturning circulation weakening.
In both AMIP and the historical experiments, this weakening is dominated by a weakening of the zonally asymmetric component of the circulation.Despite having a similar thermodynamic response to global-mean surface warming for atmospheric overturning circulation, the AMIP and the coupled historical experiments show opposing trends in the regional-level Pacific WC.A recent study shows that the Pacific WC does weaken in accordance with the hydrologic constraint, but this effect is small compared to the effect of the recent SST pattern (Watanabe et al., 2023).This complex coupling between large-scale atmospheric circulation and regional-level circulations requires further analysis.Attempts to understand the mechanisms of WC weakening with warming have pointed at an inverse relation with gross moist stability, related to an increase in the vertical depth of deep convection (Duffy & O'Gorman, 2023;Wills et al., 2017).
Although the weakening of atmospheric overturning circulation is robust across climate models (except for the piControl experiments) and observations, individual models/ensemble members exhibit a large spread in the weakening per K global-mean warming.The weaker sensitivity of the observational estimate raises the question of models possibly overestimating the response to external forcing, especially in the case of CESM2-LENS.
Past works have shown contributions of fast response to radiative forcings, land-sea contrast and SST patterns on the hydrological cycle and consequently, regional circulation changes (Allan et al., 2020;Bony et al., 2013;He & Soden, 2015;Plesca et al., 2018).Our study shows a robust weakening of the atmospheric overturning circulation with global-mean surface warming, during the satellite era, independent of the SST warming pattern.However, questions remain about the coupling between large-scale atmospheric circulation and regional-level circulations and the influence of forcing biases within the models.

Figure 1 .
Figure 1.Differing sea surface temperature and ω 500 patterns between CMIP6 AMIP and coupled historical experiments: multi-model mean decadal trend (1979-2014) in surface air temperature (a, b) and mid-tropospheric vertical pressure velocity, ω 500 (c, d) for AMIP (left) and historical (right) simulations from CMIP6.Decadal trend for the global-mean surface air temperature is reported on top of each plot.

Figure 2 .
Figure 2. Weakening of large-scale atmospheric circulation in CMIP6 experiments: Global-mean fractional change in the convective mass flux (a, b), the spatial variance in ω 500 (c, d) and the upward component of ω 500 (e, f) for AMIP (left) and coupled historical (right) simulations.Multi-model mean (MMM) is denoted by the thick blue line.Decadal trend for MMM along with the intermodel spread (±2σ) is shown in blue insets for each plot.Changes in the total spatial variance in ω 500 are further decomposed into changes in the zonal-mean variance (black dashed) and zonally asymmetric variance (red dotted).Individual time series are smoothed using a 13-month running average.

Figure 3 .
Figure 3. Implications of internal variability on the weakening of large-scale atmospheric circulation: Time series of fractional change in (a) global-and (b) tropicalmean convective mass flux for the ensemble mean of CESM2-LENS historical (green), CAM6 AMIP ensemble (pink), CMIP6 coupled historical + ssp5-8.5 (blue) and observational estimate (red).Thin green, pink, and black lines represent individual members of the CESM2-LENS, CAM6 AMIP, and piControl ensemble, respectively.Individual time series are smoothed using a 13-month running average.

Figure 4 .
Figure 4. Long-term trend in large-scale atmospheric circulation per K global warming: Trend (1988-2021) in the fractional change in the convective mass flux for historical + ssp5-8.5 (blue), CESM2-LENS (green), CAM6 AMIP (pink), and piControl (black) experiments normalized by global-mean surface air temperature trend.The red line represents the observational estimate.Multi-model mean along with the ±2σ intermodel spread for AMIP and historical experiment (1979-2014) are represented by the orange and pale blue lines, respectively.Bin size = 0.2%/K.