Most experiments discussed here (Table 1) employ the Community Earth System Model, version 2 (CESM2), a comprehensive model with coupled land, atmosphere, sea ice and ocean components (Danabasoglu et al., 2020). CESM2 represents Antarctica surface climate and the processes that affect snow accumulation very well, improving upon its predecessor, CESM1 (Dunmire et al., 2022). Here, we use the term “snow accumulation” or simply “accumulation” for consistency with Medley and Thomas (2019) and to emphasize that accumulation is the only mass input to the AIS. “Surface mass balance (SMB)” has the same meaning as our formulation of accumulation, which is explained in Sect. 2.3 below. CESM2 does not have an interactive Antarctic Ice Sheet or interactive ice shelves.

Dunmire et al. (2022) report that CESM2 simulates an unrealistically large, upward trend in Antarctic accumulation for the 1979–2015 period, which they attribute to CESM2's equilibrium climate sensitivity (ECS) of ∼ 5 K (Gettelman et al., 2019a), which is greater than observationally constrained estimates of ECS suggest (Armour et al., 2024). However, when observed SST warming patterns are accounted for, high-ECS models have a lower effective climate sensitivity (Armour et al., 2024). This is explained by the SST pattern effect, which describes the dependence of radiative feedbacks on the spatial pattern of surface warming (e.g. Andrews et al., 2022). In observations, the Southern Ocean and eastern Pacific have cooled in recent decades, in contrast to the warming simulated by CESM2 and many other CMIP6 models (Wills et al., 2022). In CESM1 (Hurrell et al., 2013) at least, introducing Antarctic observational data into the model via enhanced meltwater fluxes or wind nudging brings simulated SST and sea level pressure trends into better agreement with observations (Dong et al., 2022a; Armour et al., 2024). However, it has yet to be demonstrated that reduced Southern Ocean warming lowers the Antarctic accumulation rate. To test this idea, we exploit the many realizations of internal variability within a large ensemble and examine accumulation trends in several experiments constrained by observations.

We compare snow accumulation simulated by CESM2 with the gridded reconstruction of Medley and Thomas (2019) (hereafter, “MT19” or “reconstruction”), favoring this reconstruction because it is accessible, well verified and has been used in previous work with CESM2 (Dunmire et al., 2022). It was calibrated to the spatial signature of precipitation minus evaporation (P minus E) in the MERRA2 atmospheric reanalysis (Gelaro et al., 2017), with a bias correction based on accepted in-situ measurements. We specifically use the version remapped by Dunmire et al. (2022) from its native resolution to the standard grid of CESM2's land and atmosphere components. It spans 1801–2000 at annual resolution.

Our primary metric of snow accumulation is the timeseries of cumulative mass change over the grounded AIS relative to a pre-industrial baseline, as explained in Sect. 2.3 below. This metric de-emphasizes the large temporal variability of precipitation (Previdi and Polvani, 2016) making for a less noisy data-model comparison, and permits the time-integrated signals of climate forcings (Casado et al., 2023) to be detected. The spatial patterns of snow accumulation trends are also indicators of the responsible climate drivers; pattern correlations (Appendix A) between reconstructed and modeled trends are adopted as secondary metrics for interpreting the Antarctic snow accumulation history. Signatures of atmospheric circulation variability are prominent in Antarctic accumulation records (e.g. Medley and Thomas, 2019; Genthon et al., 2003), making it imperative to account for these impacts to uncover the underlying warming-driven trends. As such, we calculate a wind index that captures the dominant circulation variability and determine the trends in accumulation that are congruent with trends in this index (Appendix A).

Our evaluation leverages a suite of previously published CESM2 experiments (Table 1). The first four ensembles comprise the single-forcing large ensemble (Simpson et al., 2023), whose forcings added together are the same external forcings as in the half of the 100-member CESM2 Large Ensemble (CESM2-LE) that has smoothed biomass burning (Rodgers et al., 2021). The other half of the Large Ensemble, denoted as CESM2-LEcmip6, uses the standard CMIP6 historical forcing with a few exceptions (Rodgers et al., 2021; Danabasoglu et al., 2020). Hereafter the text will use “Large Ensemble” to generally refer to the whole ensemble.

The tropical Pacific pacemaker experiment, TPACE, constrains the evolution of internal variability in the model by nudging it towards observed SST anomalies in the tropical Pacific (15° N to 15° S and from the west coast of the Americas to the dateline; west of the dateline, the wedge-shaped nudging mask tapers off to a point at ∼ 0° N, 130° E) (Rosenbloom et al., 2025; O'Connor et al., 2025a). It generally follows the method of Kosaka and Xie (2013). SSTs outside of the nudged region are free to evolve. Given the ubiquitous role of tropical teleconnections in Antarctic climate variability (e.g. Li et al., 2021; Schneider et al., 2012), the pacemaker setup is an important way to constrain simulations with observations. In addition to the synching of internal variability, it can also help account for biases in the model's forced response, but it does not override the model's mean-state biases.

Stratospheric ozone depletion and recovery is represented in the Large Ensemble, TPACE and “everything else” (EE) ensembles, but its role cannot be isolated with these experiments, as was done in experiments with CESM1 (Lenaerts et al., 2018; Chemke et al., 2020; Schneider et al., 2020). There is no CESM2 single-forcing ensemble with stratospheric ozone depletion and recovery as the sole evolving forcing. Stratospheric ozone is however the dominant time-evolving forcing in EE during the latter part of the 20th Century, during which a strong accumulation signal appears. While the lack of an ozone-only ensemble is a limitation of this work, the greenhouse gas only (GHG), industrial aerosol-only (AAER), and biomass burning aerosols only (BMB) ensembles enable the novel analyses of accumulation responses to these forcings, which to our knowledge have not been previously presented.

We use the data-rich period since 1979 to evaluate multiple observation-constrained simulations (Table 2) and their depictions of the spatial patterns of snow accumulation change. These experiments include a coupled wind nudging experiment, uncoupled prescribed SST and sea ice experiments, and a coupled meltwater experiment to explore specific mechanisms of accumulation change.

Wind nudging is a technique that can override circulation biases in a model, for example associated with polar stratospheric vortex dynamics (e.g. Gettelman et al., 2019b), in addition to controlling for internal variability in the highly dynamic polar regions (e.g. Gilbert et al., 2025). In CESM2-WNUDGE the model's winds are nudged to winds from ERA5 (Hersbach et al., 2020) across the middle and high southern latitudes, following a protocol developed with CESM1 (Blanchard-Wrigglesworth et al., 2021) and recently applied to CESM2 (Espinosa et al., 2024). Zonal (U) and meridional (V) winds over January 1950 through November 2023 are nudged to 6-hourly ERA5 U and V from 850 hPa to the top of the model between 55 and 80° S. The model was free to evolve through the end of 2024. As the first few decades of CESM2-WNUDGE show signs of drift in the Antarctic accumulation values we use only the 21st Century portion of this experiment. The difficulty in removing drift and the single ensemble member are the key limitations of this experiment.

Anomalous freshwater fluxes from thinning ice shelves are not included in standard CMIP6 historical experiments nor the CESM2 experiments in Table 1. Recent work argues that ice sheet and ice shelf meltwater is a climate forcing that should be included in such experiments to enhance their accuracy in simulating historical trends and predicting future change (e.g. Rye et al., 2020; Roach et al., 2023; Schmidt et al., 2023; Sadai et al., 2025). Pauling et al. (2016) and Dong et al. (2022a) describe a meltwater hosing ensemble with CESM1 that spans 1980–2013. We utilize this ensemble (using the ensemble mean of 7 members where total anomalous freshwater flux is set to 2000 Gt yr⁻¹, prescribed at the front of ice shelves at depths of 100–200 m) for insight into the possible influence of meltwater on snow accumulation. A caveat is that the magnitude of the freshwater fluxes imposed in this ensemble is much larger than observational estimates (Pauling et al., 2016). In the observational record since 1990, anomalous freshwater fluxes to the ocean from the AIS and its ice shelves have exceeded 2000 Gt yr⁻¹ in a few years (Schmidt et al., 2025), but the time-averaged value is around 500 Gt yr⁻¹ (Swart et al., 2023). In reality, anomalous ice-to-ocean freshwater fluxes are approximately equally distributed between ice shelf basal melting and iceberg discharge (e.g. Coulon et al., 2024; Schmidt et al., 2025). Pauling et al. (2016) present an iceberg freshwater flux experiment, but we do not include it here due to the small ensemble size and the precedent of previous climate dynamics studies in using the larger CESM1 ice shelf basal melt ensemble (Dong et al., 2022a; Armour et al., 2024). Advantages of the CESM1 basal meltwater experiment are that it captures the spatial pattern of anomalous meltwater forcing including its depth, and that it simulates significant responses in SSTs (Dong et al., 2022a), sea ice extent (Pauling et al., 2016) and transient climate sensitivity (Dong et al., 2022a). All CESM experiments conserve ocean mass and volume such that snow accumulating on the AIS in excess of 1 m is returned to the ocean, analogous to river runoff. The freshwater forcing in CESM1-AIS meltwater is implemented as an additional, negative salinity flux to estimate the climate impact of the mass imbalance of the AIS (Pauling et al., 2016).

To circumvent the challenges in modeling the roles of meltwater and winds on SSTs and accumulation, we directly link Antarctic accumulation with the observed SST and sea ice record by considering two uncoupled, atmosphere-land-only simulations (“AMIP” style) with prescribed SSTs and sea ice concentrations from observational datasets. The Global Atmosphere – Global Ocean (GOGA) experiment specifies time-varying SSTs and sea ice concentrations throughout the global ocean domain. The Tropical Ocean – Global Atmosphere (TOGA) uses time-varying SSTs in the tropical latitudes only and specifies monthly climatology for SSTs and sea ice concentrations elsewhere.

The domain of our snow accumulation calculations is the grounded AIS, as defined by Zwally et al. (2012). Unlike Medley and Thomas (2019), we do not include Antarctic islands in our calculation of Antarctic-wide snow accumulation. For Antarctic-wide snow accumulation from MT19, grid-cell values in mm w.e. yr⁻¹ are scaled by the area of the respective grid cell, and summed over the grounded AIS for each year of the dataset. This mass timeseries is converted to a relative mass timeseries by subtracting the long-term mean for 1801–1900 (1927.4 Gt yr⁻¹) from the entire timeseries. Cumulative mass is calculated by integrating the relative mass timeseries with time, starting in 1901. SLE at the year 2000 is obtained by dividing the cumulative mass by 361 Gt mm⁻¹. This procedure was repeated for +1σ and -1σ error timeseries; the gridded error fields are as provided by Medley and Thomas (2019) in their MERRA2 reconstruction. Here, the term “relative mass” has a distinct meaning from “Antarctic-wide accumulation rate”. Changes in the Antarctic-wide accumulation rate are assessed over specific time intervals, for example linear trends over 1901–2000. Relative mass refers specifically to how much the Antarctic-wide snow accumulation rate (expressed in Gt yr⁻¹) in a particular year is anomalous relative to the 1801–1900 baseline value.

From the output of CESM2's land model component, SNOW and RAIN variables are used as the precipitation terms, while QSOIL and QRUNOFF are used as the ablation terms. Monthly data are summed to annual means by weighting each monthly value by the length of the month. The relative and cumulative mass timeseries are determined by the same method as for the MT19 reconstruction, except for the calculation of the preindustrial baseline value.

Crucial for the calculation of cumulative mass and its sea level equivalence from CESM2 is the estimate of the baseline accumulation rate that is representative of nineteenth-century values in a stable climate. We reference each ensemble member to the same baseline value from piControl, averaged over the 452-year period of 1000–1451, yielding 2012.9 Gt yr⁻¹. All ensemble members of all historical experiments in Table 1 had their initial states drawn from this piControl segment, between the years 1001 and 1301, which was chosen because it has a stable control climate (Danabasoglu et al., 2020; Rodgers et al., 2021; Simpson et al., 2023). An ensemble member initialized from the piControl year 1301 at calendar year 1850 would reach calendar year 2000 150 years later, overlapping with piControl through year 1451. This overlap method is a way of ensuring that mass accumulated by year 2000 arises from forcing (or internal variability), minimizing the influence of drift. The same 1000–1451 piControl period was used for accumulating mass over 24, overlapping 100-year segments to determine how much mass gain or loss within a century could occur due to internal variability alone, in the absence of anthropogenic forcing.

To help illustrate and quantify the role of atmospheric circulation and surface temperatures over the entire 20th Century, we employ annually resolved reconstructions of near-surface zonal winds, sea level pressure (SLP) and surface air temperature (SAT) generated by paleoclimate data assimilation (PDA). We first use the reconstruction presented in O'Connor et al. (2021a) that uses members of the CESM1 Large Ensemble (Kay et al., 2015) to form the prior, which is a physically based assumption about the state of the system without accounting for new information from observational or proxy data. The global proxy network includes Antarctic ice core data used in the MT19 reconstruction. This reconstruction, while skillful in the regions of interest to this study, is not fully dynamically consistent with the CESM2 experiments evaluated here. We therefore utilize a newer PDA reconstruction following the same methodology and using the same global proxy network, but with the prior formed by the first seven members of CESM2-LEcmip6 (O'Connor et al., 2025a). Use of both CESM1-LE PDA and CESM2-LE PDA has some advantages. While both reconstructions skillfully estimate the ASL index (Fig. 1a) and other aspects of the dynamics represented by winds and sea level pressure (Fig. A1), there are marked differences in their surface temperature reconstructions (Figs. A2, A3). Based on considerations discussed in Appendix A, we regard the CESM1-LE PDA as the more reliable of the two surface temperature reconstructions.

The broad suite of experiments combined with results from PDA enable a multi-faceted strategy to characterize signals arising from external climate forcing and separate them from the effects of internal variability. For the forced response, the ensemble mean of a large ensemble smooths out the signals of internal variability arising from the initial conditions, revealing the model's response to the external radiative forcings (greenhouse gasses, aerosols, etc.) imposed in the experiment. In this text, the forced response is denoted by square brackets, for example “[GHG]” for the ensemble-mean of the greenhouse gas single forcing ensemble. The internal component of a given ensemble member is found by subtracting the forced response from that member.

Using large ensembles is only the first step towards a comprehensive accounting of internal variability. This step may not be sufficient because (a) models like CESM2 may simulate an unrealistic magnitude of internal variability (e.g. Casado et al., 2023); and (b) internal variability in free-running coupled experiments like the Large Ensemble is not synched with real-world internal variability. The second step is the use of nudged coupled experiments, TPACE and CESM2-WNUDGE, which constrain the coupled model by introducing observed anomalies. Third, we use a long segment of piControl to characterize the evolution of accumulation in the absence of anthropogenic forcing. Fourth, the PDA reconstructions of the major surface climate variables that we use are constrained both by well-calibrated proxy records and by the physical dynamical frameworks of CESM1 and CESM2. Finally, we evaluate accumulation trends in prescribed SST/sea ice experiments using again the same dynamical framework of the atmosphere and land component models of CESM. Importantly, results from all of these approaches, presented below, converge upon a single set of conclusions: The accumulation rate increase during the 20th Century that mitigated sea level rise by ∼ 11 mm can only be explained by external forcing, namely greenhouse gases. Real-world internal variability most likely has dampened the forced response. The follow-up question we discuss but ultimately leave open for future work is the extent to which internal variability is sufficient to account for observation-model trend discrepancies, or if there is an important role for poorly resolved processes such as the response to anomalous ice sheet and ice shelf meltwater and the sensitivity of SST anomalies to wind stress.

From the MT19 reconstruction, we find a cumulative mass gain over the grounded AIS of 10.5 mm SLE over the 20th Century (Fig. 2a), in agreement with the 10.6 mm previously reported (Medley and Thomas, 2019). The cumulative mass gain arises because the relative mass (the annual anomaly relative to the nineteenth century) increases in a stepwise fashion during the 20th Century (Fig. 1c). For 1901–1925, the relative mass is about 20 Gt yr⁻¹; for 1976–2000 it is above 60 Gt yr⁻¹. The cumulative mass timeseries is highly correlated with global-mean temperature anomalies in the CESM1-LE PDA (r=0.92; Fig. 3a). Since some common proxy data were used in MT19 and the PDA, we verify this relationship with the independent SAT dataset from ERA-20C (Poli et al., 2016), which similarly shows a high correlation with cumulative mass. Cumulative mass and global temperature anomalies are also highly correlated in CESM2 (r=0.83 for [TPACE]). On the continental scale, both the reconstructions and the model experiments indicate a significant positive correlation between AIS-wide surface temperature anomalies and relative mass timeseries (Fig. 3b), consistent with previous work on the sensitivity of snow accumulation to temperature (Monaghan et al., 2008; Frieler et al., 2015; Dalaiden et al., 2020). Broadly, these results affirm the expectation from thermodynamics of more snow accumulation with warming (e.g. Frieler et al., 2015). However, the exact sensitivity depends on the choice of averaging period, specific experiment, and/or whether an ensemble mean or individual ensemble member is used. Global- or Antarctic- mean temperatures provide only a weak constraint on the magnitude of the snow accumulation increase and are not diagnostic of the forcings that are driving it. To understand the role of these forcings, we evaluate the suite of ensembles listed in Table 1.

As shown in Fig. 2 and Table 1, [CESM2-LE] has a mass gain of 16.8 mm SLE from 1901 to 2000, well above the MT19 reconstruction. The original members of the Large Ensemble with the CMIP6 biomass burning protocol exhibit a slightly higher ensemble-mean value of 17.8 mm SLE, physically consistent with the spurious warming that arises from CMIP6 biomass burning emissions (Fasullo et al., 2022). Ensemble spread is indicated in the box-whisker plots (Fig. 2b, d). The 5 % lower bound of the CESM2-LE lies within the +1σ error of MT19, while two outlier ensemble members are within MT19's -1σ error. Interpreting the [CESM2-LE] value of 16.8 mm SLE as the forced response to all forcings combined, a mass loss of 6.3 mm SLE is required to match the value of 10.5 mm SLE from MT19. This mass loss can be interpreted as driven by internal variability as 10.5 mm lies within the ensemble spread of CESM2-LE, but falls on the extreme end of the ensemble distribution outside of the 5 % lower whisker (Fig. 2b). A similar distribution occurs in CESM2-LEcmip6, with 10.5 mm SLE lying on the extreme end (Fig. 2d).

Mass accumulated over century-length segments of piControl suggests that internal variability alone cannot account for the observed mass gain (Fig. 2b). Yet, as the ensemble spreads show, internal variability can partially counteract the forced response in [CESM2-LE] or [CESM2-LEcmip6] and bring model results in agreement with MT19. In the constrained TPACE experiment, mass gain is only 11.5 mm SLE, just 1 mm greater than MT19 and within its 1σ error (Fig. 2b). The observed trends and variability captured in TPACE reduced 20th-Century cumulative mass by 6.4 mm SLE, roughly matching the magnitude of internally driven mass loss in the Large Ensemble that is required to fit the observations.

The single-forcing ensembles uncover the two major components of the forced response (Figs. 2b, d). Greenhouse gases are the dominant driver of the cumulative mass gain, with [GHG] giving a value of 21 mm SLE, twice the value from the reconstruction. Aerosols offset this; mass change in [AAER] is -9 mm SLE. The sum of [GHG] and [AAER], 12 mm SLE, is within the +1σ error of MT19 (Fig. 1b). The other forced responses are smaller, at 2.1 mm SLE for [EE] and -0.78 mm SLE for [BMB]. Although the response in [EE] does not strictly isolate the role of stratospheric ozone depletion, the mass gain in [EE] occurs after 1980 (Fig. 2c), consistent with the timing of ozone depletion and qualitatively consistent with ozone-driven accumulation increases found in previous work (Lenaerts et al., 2018; Schneider et al., 2020). The sum of the four separate ensemble means (13 mm SLE) is near the top of the +1σ error range of MT19 and in line with the ∼ 14 mm SLE from the reconstruction of Wang and Xiao (2023). The result that [CESM2-LE] is ∼ 4 mm SLE larger than the sum of the single-forcing ensemble means could be due to aerosol forcing alone causing a large increase in Antarctic sea ice area, but not when it acts in concert with greenhouse gas forcing (Simpson et al., 2023).

The results thus far, while showing general consistency between the CESM2 and the MT19 reconstruction, suggest two interpretations for the cumulative mass gain in MT19. Namely, the single-forcing ensembles suggest that the 10.5 mm SLE from MT19 is only due to external forcing, dominated by opposing responses to GHGs and aerosols. In contrast, the Large Ensemble and TPACE suggest a smaller role for aerosols, but a large role for internal variability in counteracting the mass gain due to GHGs. We therefore turn to the spatial patterns of snow accumulation and atmospheric circulation change to help inform the most likely interpretation of the trends. This analysis will lead to a third possibility: A portion of the mass loss that is internally driven according to the Large Ensemble and TPACE may indicate a missing and/or misrepresented external forcing.

For spatial analysis, we survey the pattern correlations between the 20th-Century trend in MT19 accumulation and each ensemble member of the Large Ensemble and TPACE, along with the ensemble means of all historical experiments from Table 1 (Figs. B1, B3). Similarly, pattern correlations for the sea level pressure (SLP) trend across 40–90° S are computed with the same model experiments and using the CESM2-LE PDA as the benchmark (Figs. B2, B4). For illustration, we select the ensemble member whose trend patterns in snow accumulation and atmospheric circulation exhibit the best pattern correlations to the MT19 and CESM2-LE PDA reconstructions, respectively. Specifically, member 40 from CESM2-LEcmip6 has a pattern correlation with the reconstructed SLP trend over 40–90° S of 0.92 and with the reconstructed accumulation trend on Antarctica of 0.52, giving an average correlation of these two metrics of 0.72. This is the highest average correlation out of 110 individual members from the CESM2-LE, CESM2-LEcmip6, and TPACE ensembles. The following discussion uses member 40 to illustrate spatial patterns but these same relationships are found in several other ensemble members.

Like the reconstructed patterns (Fig. 4a), member 40 (Fig. 4b) exhibits an accumulation trend dipole across the WAIS, consistent with the deepened ASL and stronger onshore flow onto the eastern WAIS and offshore flow from the western WAIS. On a larger scale, both member 40 and the reconstruction feature a pressure dipole between the middle and high southern latitudes, consistent with the strengthening and poleward shift of the westerlies (O'Connor et al., 2021a; Dalaiden et al., 2022). Using the Large Ensemble, member 40 is separated into its forced (Fig. 4c) and internal (Fig. 4d) components. The increased accumulation rate is entirely explained by external forcing, with only a minor offset from internal variability. Using the single-forcing ensembles, the forced response patterns associated with individual forcings are also evaluated (Figs. B3, B4). [GHG] and [EE] are best correlated with the reconstructed SLP and accumulation patterns, but not better than the all-forcings response in [CESM2-LEcmip6].

The accumulation pattern in member 40, including the dipole across the WAIS, is shaped by internal variability (Fig. 4d). In the forced response, snow accumulation increases nearly everywhere on the continent but is especially enhanced on the coasts of the WAIS and Queen Maud Land (Fig. 4c). The forced atmospheric circulation pattern indicates a strengthening and poleward shift of the westerlies, especially in the Indian Ocean sector and Drake Passage. The internally driven circulation pattern is also consistent with a strengthening of the westerlies, displaying the characteristic pressure dipole between the middle and high southern latitudes. Neither of the two circulation patterns are perfectly zonally symmetric. The forced pattern features lobes of low pressure jutting out into the eastern Ross Sea and into the south Atlantic, whereas the internal pattern features a deepened ASL, marked by low pressure anomalies in the southeastern Pacific. As discussed below, these zonally asymmetric pressure patterns are associated with zonally asymmetric SST trends in lower latitudes.

How have the strengthening westerly winds, associated with the positive phase of the SAM, affected accumulation? Medley and Thomas (2019) use the SAM index (Marshall, 2003) to show a reduction in accumulation related to the positive phase of the SAM for 1957–2000. Here, we use the zonal winds directly and for the full century, as provided by CESM2-LE PDA and member 40 (Fig. 1b). In the MT19 reconstruction, there is a large reduction in the Antarctic-wide accumulation rate that is linearly congruent with the significant trend in the zonal wind index (Fig. 4e). Spatially, the accumulation rate is reduced nearly everywhere on the continent expect for the Peninsula. This pattern is more weakly expressed in member 40 and there is only a small change in the Antarctic-wide accumulation rate associated with the winds (Fig. 4f). With wind-congruent trends removed from the reconstructions, the residual pattern exhibits a large increase in the accumulation rate associated with increased accumulation over East Antarctica, accompanied by a weaker ASL (Fig. 4g). This wind correction brings the reconstructed Antarctic-wide accumulation trend closer to the forced response given by [CESM2-LEcmip6].

The effect of tropical SST nudging in TPACE is to weaken the ASL and reduce the magnitude of the westerly wind trend (Figs. B2, B5a), resulting in a worse fit of modeled and reconstructed trend patterns than in [CESM2-LE*] (Figs. B1, B2). Nudging also drives a reduction in the Antarctic-wide accumulation rate, bringing the cumulative mass gain in [TPACE] closer to the reconstruction than in [CESM2-LE*]. This suggests that it is not the winds or deepened ASL per se which reduce the accumulation rate. A key differentiator of the reconstructed (Fig. 4a) and modeled trend patterns (Fig. 4b) is their surface temperature trends. According to regression, reconstructed and modeled accumulation have similar sensitivities to temperature of 5 %–6 % °C⁻¹ on an Antarctic-wide basis (Fig. 3b), consistent with CMIP6 multi-model consensus (Nicola et al., 2023). In the reconstruction, there is a surface cooling trend in the Pacific Sector of the Southern Ocean, especially in the Amundsen Sea near the center of the ASL. The cooling region covers the major moisture source region for snowfall on the WAIS (Sodemann and Stohl, 2009). In member 40, the Pacific Sector warms less than other regions of the Southern Ocean, but only slightly less. The wind patterns are also associated with surface cooling in the mid-latitude Indian Ocean, the major moisture source region for East Antarctic precipitation (Sodemann and Stohl, 2009). [TPACE] is also associated with muted Southern Ocean warming, but it is maximized in the Atlantic and Indian Ocean sectors rather than the Pacific (Fig. 5o). Comparing Fig. 4g and h, even after the wind corrections, there are stronger surface warming trends in the free-running model than in the reconstruction; TPACE appears to account for some of this difference.

To provide context for the polar wind and surface temperature trends, we zoom out to the entire globe (Fig. 5). The observed SST pattern from the independent ERSSTv5 dataset (Fig. 5a) is not obviously La Niña – like or El Niño – like, but it does feature less warming in the central-eastern Pacific than in the western Pacific, consistent with the strengthening Walker circulation during this time period (Lee et al., 2022) that is reflected in the zonal gradient in the SLP trend simulated by [CESM2-GOGA]. The zonal wind index from CESM2-LE PDA is congruent with cooling in the central-Eastern Pacific (Fig. 5b), in an overall pattern that closely resembles “Pattern 1” associated with observed-modeled SST trend discrepancies since 1979 (Wills et al., 2022, their Fig. 3). The wind-residual pattern is associated with enhanced eastern equatorial Pacific warming (Fig. 5c). The PDA reconstructions capture these large-scale patterns, showing muted warming in the central-eastern Pacific and enhanced warming in the western Pacific (Fig. 5d). This zonal temperature contrast is enhanced by the winds (Fig. 5e). Like in ERSSTv5, if the wind-congruent trends are removed from the reconstruction, the residual pattern shows an El Niño – like equatorial warming pattern with a weakened ASL (Fig. 5f). Considering the dynamics of poleward propagating Rossby waves, a weakened ASL is typically associated with an El Niño – like SST pattern in the tropics, while a deepened ASL is typically associated with a La Niña – like SST pattern (e.g. Raphael et al., 2016).

The observed and reconstructed patterns can be compared to the forced responses to all forcings (Fig. 5j), aerosols only (Fig. 5k), and GHGs only (Fig. 5l) ([EE] exhibits a similar pattern to [GHG] but of much weaker magnitude, not shown). [GHG], with its El Niño – like warming pattern and north Atlantic cooling, is a good fit to the wind-residual patterns in the observations and reconstructions. Cooling in [AAER] is concentrated in the north Pacific and eastern equatorial Pacific, in disagreement with the cooling patterns associated with winds. Aggressive aerosol-driven cooling is evident in the all-forcings forced response, [CESM2-LEcmip6], which has too weak of a warming trend compared to observations and reconstructions. The Pacific zonal SST gradient in is improved in member 40 (Fig. 5g), along with the representation of the deepened ASL. However, in member 40, the winds are only associated with weak SST anomalies (Fig. 5h), and the wind-residual pattern is not El Niño like (Fig. 5i), possibly because the aerosol response in the tropics overwhelms the GHG response (Heede and Fedorov, 2021). The stronger zonal SST contrast and deeper ASL in member 40 are driven by a pattern of internal variability that resembles the negative phase of the Interdecadal Pacific Oscillation (IPO; Fig. 5n) defined by Henley et al. (2015). The response to nudging in TPACE (Fig. 5o) is a near mirror image to the internal component of member 40 (Fig. 5n) except that both have a patch of cooling in the central tropical Pacific, consistent with the muted warming there in observations (Fig. 5a).

In summary, modeling a good fit to the reconstructed accumulation trend pattern requires (a) A deepened ASL; (b) Stronger circumpolar westerlies; and (c) A strong zonal SST gradient in the subtropical Pacific, with more warming in the western than eastern part of the basin. Holland et al. (2022) similarly find that western subtropical Pacific warming accompanies ASL deepening, which they hypothesize is part of a negative IPO-like pattern of internal variability. Consistently, member 40's internal component is a negative IPO pattern. Note that the pattern congruent with reconstructed winds (Fig. 5e) differs from the canonical negative IPO in that it has its strongest loadings in the Amundsen Sea region.

For 1979–2019, we adopt [CESM2-GOGA] as the reference pattern (Fig. 6a). [CESM2-GOGA] has no trend in the Antarctic-wide accumulation rate, consistent with results from reanalysis and regional models (e.g. Clem et al., 2023; Mottram et al., 2021). Like the pattern during the 20th Century, a trend dipole occurs across the WAIS, consistent with the deepened ASL (Fig. 1a). Significant negative accumulation trends occur in Wilkes Land over the Totten Glacier and in West Antarctica over the Thwaites and Pine Island Glaciers, two regions where independent observational studies show strong evidence for circulation-driven variability in snow accumulation that affects year-to-year variations in the mass balance of these basins (King and Christoffersen, 2024). Nonetheless, long-term mass loss in these basins has been predominantly driven by ice-ocean interactions and grounding line discharge, not by surface processes (e.g. Shepherd et al., 2019; Lu et al., 2025). The accumulation trend magnitudes are weaker and less significant in [CESM2-TOGA] (Fig. 6b), which does not simulate the negative trend over Totten. This suggests that regional accumulation trends are responsive to extratropical SST and sea ice trends, consistent with mid-to-high latitude moisture sources of precipitation for near-coastal Antarctic regions (Sodemann and Stohl, 2009).

The coupled experiments, [CESM2-LEcmip6] (Fig. 6c) and [TPACE] (Fig. 6d) exhibit significant upward accumulation trends across much of the continent, inconsistent with the observation-constrained [CESM2-GOGA]. Spatially, [TPACE] fits the reference pattern given by [CESM2-GOGA] better than does [CESM2-LEcmip6]. Although tropical pacemaker experiments can simulate ASL deepening for the period since 1979, they do not simulate Antarctic sea ice expansion and Southern Ocean surface cooling (Schneider and Deser, 2017), which explains why [TPACE] simulates too strong of an accumulation trend. A temporal perspective on these recent trends is provided in Fig. 6e. In MT19 and the prescribed SST experiments, most of the accumulation increase occurs between about 1970 and 1990, during a period of Southern Ocean warming (Zhang and Deser, 2024). In the coupled experiments, the accumulation rate steadily increases from about 1970 through the end of the timeseries, physically consistent with stronger-than-observed ocean surface warming and sea ice loss.

The ASL has continued to deepen in the 21st Century (Fig. 1a), as reflected in the 2001-2022 SLP trend pattern in ERA5 (Fig. 7a). The ASL deepening is embedded within a larger-scale trend pattern associated with the positive phase of the SAM. The ERA5 SLP and snow accumulation patterns are well captured in CESM2-WNUDGE (Fig. 7b). Underlying the nudged pattern is a steady upward trend in the forced response (Fig. 7c) associated with background warming. Figure 7d isolates the effect of the wind nudging, which drives cooling in the Amundsen Sea but strong warming in the western Ross Sea, reminiscent of the pattern associated with wind-congruent trends in the reconstructions (Fig. 4e). While the forced response drives significant accumulation trends in the interior of East Antarctica and Queen Maud Land (Fig. 7c), the wind pattern is associated with significant drying over Wilkes Land and a significant accumulation increase in the Peninsula region (Fig. 7b, d).

Large-scale orographic effects are evident in the wind-driven accumulation pattern. The WAIS divide separates the positive and negative snow accumulation trends, with Elsworth Land on the windward side and Marie Byrd Land on the lee side with respect to the ASL. In East Antarctica, Queen Maud Land receives high snowfall as it faces the winds, and its climatological moisture transport pathway (Naakka et al., 2021) is enhanced by the trend pattern. The positive snow accumulation trend extends inland to the South Pole, qualitatively consistent with observations there (Zhai et al., 2023). At least some of the increased cyclonic activity in the Amundsen Sea and south Atlantic, associated with positive snow accumulation trends in the Peninsula region and Queen Maud Land, respectively, is driven by the forced response (Fig. 7c). A negative snow accumulation trends occurs across most of the terrain that lies downwind of the main East Antarctic ice divide. The largest negative trends occur in Wilkes Land. A recent observational study (Wang et al., 2025a) found a significant decrease in the 2005–2020 accumulation rate along a transect from Zhongshan Station (69° S, 76° E) to Dome A (80° S, 77° E), associated with trends in the large-scale atmospheric circulation and consistent with the accumulation trend pattern in CESM2-WNUDGE.

While atmospheric circulation explains spatial accumulation patterns, temperature in and around Antarctica is the main driver of the accumulation rate. The most pronounced period of Southern Ocean surface cooling and increasing sea ice extent in recent years was 1980–2013, which corresponds with the period covered by the meltwater experiment. For comparison, we include [CESM2-GOGA] (Fig. 8a) and the older [CESM1-GOGA] (Fig. 8b), which uses the same CAM5 atmospheric model as [CESM1-AIS meltwater] (Fig. 8c). The snow accumulation pattern in the meltwater experiment exhibits good agreement with the pattern in [CESM1-GOGA] (r=0.62), while having a less than 1 % increase in the Antarctic-wide accumulation rate during 1979–2013. Thus, the accumulation response to meltwater is consistent with the response to the observed SSTs. This result of an accumulation decrease in response to Antarctic meltwater is qualitatively supported by Zhang et al. (2026), who find meltwater-driven precipitation decreases on the Antarctic continent across multiple CMIP6-era models. Additionally, Sadai et al. (2025) find a similar precipitation decrease in a set of CESM1 experiments driven by meltwater fluxes from an offline ice sheet model. Observational uncertainties and model limitations make a highly realistic meltwater simulation very challenging, but there is a multi-model consensus that meltwater leads to Southern Ocean surface cooling and a precipitation decrease over Antarctica (Zhang et al., 2026).

As discussed in previous work (Dong et al., 2022a; Armour et al., 2024), Antarctic meltwater reduces global-mean warming and reduces effective climate sensitivity via the SST pattern effect. A key piece of the SST pattern effect is the zonal SST gradient in the subtropical south Pacific, which CMIP6 models systematically do not replicate (Wills et al., 2022). In the reconstructions for 1901–2000, there is less warming in the subtropical and mid-latitude eastern Pacific than in the western part of the basin even if the wind-congruent trends are removed (Fig. 9a). This is a stronger gradient than can be explained by GHGs alone (Fig. 5l), suggesting that another signal is present. The meltwater-induced pattern (Fig. 9b) could be this missing signal, as it features more cooling in the eastern than western Pacific. If the meltwater signal is added to [GHG], the resulting pattern (Fig. 9c) is strongly correlated (r=0.91 over 30–70° S) with the wind-residual pattern in the PDA reconstruction (Fig. 9a).

The meltwater-induced SST anomaly pattern in Fig. 9b was originally estimated by Dong et al. (2022a) as a decadal trend; here we multiply their decadal trend by five to account for 50 years during which Antarctic meltwater has plausibly been a relevant forcing. Ice shelf thinning and grounding line retreat in West Antarctica commenced in the mid-20th Century, several decades prior to the modern satellite era (Smith et al., 2016; Clark et al., 2024). Since the mid-20th Century, an anomalous contribution of freshwater to the ocean has been sustained by ice shelf thinning and glacier retreat driven by northerly wind anomalies in the Amundsen Sea Embayment associated with the ASL deepening trend (O'Connor et al., 2025a). For snow accumulation, it is around 1950 when the cumulative mass gain in [CESM2-LE] begins to outpace the gain in the MT19 reconstruction (Fig. 2a), and when the signal in [TPACE] noticeably counteracts the signal in [CESM2-LE*] (Fig. 2c). It is plausible that a meltwater-driven cooling signal could be propagated towards the subtropics via atmospheric advection and a chain of amplifying wind-SST-cloud feedbacks (Dong et al., 2022a, b; Kim et al., 2022; Kang et al., 2023) and be embedded in the observed SST record and the proxy records. ERA5 trends over 1950–2024 show cooling around 60° S, 180° W (Fig. C1) that is consistent with the meltwater response in Fig. 9b–c. The pattern due to tropical SST nudging in TPACE (Fig. 9d) shows a central Pacific El Niño-like state (resembling “El Niño Modoki”, e.g. Ashok and Yamagata, 2009) that is consistent with the CESM1 response to meltwater (Fig. 9b). As discussed above, the Southern Ocean cooling response to meltwater is supported by a well-established multi-model consensus. However, the tropical Pacific response depends upon model physics such as the strength of subtropical shortwave cloud feedbacks and the structure of the intertropical convergence zone (ITCZ) (Zhang et al., 2026; Dong et al., 2026; Kim et al., 2022).

The foregoing results are summarized in graphical form in Fig. 10, which compares accumulation trends and their large-scale contexts across the reconstructions (Fig. 10a), the GHG-only scenario (Fig. 10b), and associated with the proxy reconstructed wind timeseries (Fig. 10c). The maps use a pseudo Spilhaus projection (e.g. Chen et al., 2023) to place Antarctica and the Southern Ocean near the center of the map and emphasize connections across the ocean basins. In Fig. 10a, the 10.5 mm of sea level mitigation occurs in the context of a deepened ASL and strong warming around most of Antarctica except for the Amundsen Sea. The ice-albedo feedback makes these warming/cooling trends larger than anywhere else except for the Arctic. Cold-air advection and the wind-evaporation-SST (WES) feedback propagate the Amundsen Sea cooling signal towards the tropics as the south Pacific subtropical high strengthens. However, this signal is blocked from reaching the eastern equatorial Pacific by the ITCZ and by the competing effects of greenhouse gases (Fig. 10b). Although the eastern tropical Pacific SSTs look El Niño like, the overall SST anomaly pattern results in Rossby waves propagating towards Antarctica that deepen the ASL, reinforcing this pattern (see [CESM2-GOGA] simulated SLP pattern, Fig. 5a), which promotes ice shelf thinning and glacier retreat in the Thwaites and Pine Island region (King and Christoffersen, 2024; O'Connor et al., 2025a). Under greenhouse gases only (Fig. 10b), the tropics are decidedly El Niño like and poleward propagating Rossby waves act to weaken rather than reinforce the ASL deepening trend. The south Pacific subtropical high strengthens, promoting the WES feedback, but it does not extend to the coast of South America where it could help stabilize the stratocumulus cloud deck, enhance coastal upwelling, and alter the trajectory of the tropics.

Greenhouse gases explain the widespread warming and accumulation increase evident in the proxy reconstructions but do not explain the important dynamical details that suppress the accumulation rate increase, give rise to a dipole in snow accumulation across the WAIS, and reshape the warming patterns. The Southern Hemisphere westerly winds are the major source of momentum forcing of the global ocean (e.g. Wunsch, 1998). Their potential impact on surface climate is depicted in Fig. 10c. In this depiction, the cooling signal originating in the Amundsen Sea is communicated all the way to the equatorial Pacific along the reverse teleconnection pathway. The subtropical shortwave cloud feedback (Kim et al., 2022) and the ITCZ (Dong et al., 2026) are the key mediators of this connection. Recent studies argue that strengthening winds forced by stratospheric ozone depletion contribute to the La Niña like SST trend in the tropics (Hartmann, 2022; Dong et al., 2025). Similarly, Antarctic meltwater can contribute to a La Niña like SST trend pattern (Dong et al., 2022a; Zhang et al., 2026; Dong et al., 2026). However, none have argued that these high latitude signals are the sole cause of the observed La Niña like trend. It is important to keep in mind that the wind timeseries (Fig. 1b) used in Fig. 10c includes signals arising from internal variability. During the 20th Century, a trend towards the negative phase of the IPO (which favors La Niña) preceded strong forcing from greenhouse gas increases and stratospheric ozone depletion (Dalaiden et al., 2024). As such, this IPO trend is part of the wind-congruent map in Fig. 10c. Nonetheless, the presence of internal signals within this IPO-like pattern does not rule out the possibility of anthropogenic forcing driving a multidecadal shift towards the negative phase of the IPO (Klavans et al., 2025). Regardless of the exact origins of this pattern, as shown it exhibits the mutually reinforcing two-way teleconnections discussed in Dong et al. (2022b). This reinforced circulation pattern enhances meltwater production by the mechanism of O'Connor et al. (2025a). To the extent that the meltwater-driven cooling signal reaches the tropics, it is communicated back to the Southern Ocean via the two-way teleconnection.

For brevity, Fig. 10 omits the cooling patterns explicitly associated with aerosols and meltwater. Figure 5k shows that aerosols drive Southern Ocean cooling, which reduces accumulation across Antarctica (Fig. B3). Meltwater also cools the Southern Ocean and reduces the accumulation rate (Figs. 8c, 9b). Both the aerosol-only and meltwater experiments have limited utility as explanations for the 20th-Century history. The strong cooling in [AAER] may be partially spurious because of the experimental design in which aerosol forcing acts alone on a cold pre-industrial background state, triggering ice-albedo feedbacks to amplify the cooling more than it would be when greenhouse gases increase at the same time (Simpson et al., 2023). The meltwater experiment suffers from its unrealistic total magnitude of forcing and from its distribution of this forcing all around the continent. More realistically, 20th Century meltwater forcing was small and concentrated in the Amundsen Sea, where it could locally alter sea ice formation and SST anomalies, ultimately strengthening the westerlies in the southeast Pacific, to which the ocean would have to adjust (e.g. Kang et al., 2026). In turn, this cooling signal could be propagated towards the eastern Pacific tropics along the reverse teleconnection pathway. Some combination of high-latitude winds and meltwater signals have likely contributed to the wedge-shaped pattern of cooling or muted warming in the proxy reconstruction that extends into the central tropical Pacific (Fig. 10a); this pattern is remarkably similar to recent results from high-resolution model simulations forced by the standard suite of external forcings (greenhouse gases, aerosols, stratospheric ozone, etc.) without the meltwater feedback (Kang et al., 2026).

The first part of our analysis focuses on attributing the 20th-Century history of Antarctic snow accumulation using the single-forcing large ensemble and the all-forcings Large Ensemble from CESM2. We evaluate what combination of external forcing and internal variability explains the historical trends and therefore may be applied towards more accurate projections of snow accumulation into the future. Greenhouse gases have been the primary driver of increased snow accumulation on the AIS, leading to a cumulative mass gain of 21 mm SLE for 1901–2000 due to their role in warming the atmosphere and ocean surface. Internal variability over the 20th Century cannot explain the observed cumulative mass gain, but internal variability could combine with the forced response to greenhouse gases and aerosols (when used together in the same Large Ensemble experiment) to reduce the magnitude of the cumulative mass gain. This contrasts with Previdi and Polvani (2016), who argue that internal variability has completely masked the anthropogenic increase in Antarctic snow accumulation. Their study did not have the long observational perspective of the MT19 reconstruction to work with, nor the single-forcing large ensemble. By employing these datasets and applying the cumulative mass gain metric, the anthropogenic signal in snow accumulation can be detected.

Table 3 summarizes our interpretation of the cumulative mass gain results, and how a missing meltwater forcing in conjunction with the “reverse” tropical teleconnection could reconcile results among the experiments and the reconstructions. What this does not resolve is the portion of the wind-congruent accumulation trends that are attributable to external forcing vs. internal variability, nor the question of whether West Antarctic ice shelf thinning (the source of the meltwater forcing) was naturally triggered. These questions are explored in Holland et al. (2022). From their study, the pattern of internal variability that fits PDA-reconstructed atmospheric circulation trends is the negative phase of the IPO, which concurs with our analysis. However, the pattern we uncover differs from the canonical IPO pattern in that it is not symmetrical about the equator; it has stronger loadings in the Southern Hemisphere. This is consistent with the idea of the Southern Ocean rather than the tropics as the true pacemaker (e.g. Kang et al., 2023, 2026; Zhang and Deser, 2024). The ASL is part of the negative IPO pattern; Dalaiden et al. (2024) attribute ASL deepening trends to a combination of internal variability and anthropogenic forcing. Holland et al. (2022) suggest that while ice-shelf thinning may have been naturally triggered in the mid-20th Century, it has been sustained by anthropogenic forcing. The ASL deepening trend is consistent with northerly wind anomalies in the Amundsen Sea Embayment that are driving ongoing ice shelf thinning and WAIS retreat, supplying the meltwater forcing (O'Connor et al., 2025a).

Results presented here agree with the previous attribution work. Much of the reconstructed circulation trend can be explained by the forced responses to greenhouse gases and “everything else” but there is considerable spread across the Large Ensemble arising from internal variability (Fig. B4). The ensemble member that best matches the reconstructed circulation history and accumulation spatial pattern, member 40, does not exhibit the strong relationship between winds and surface temperature trends that is evident in reconstructions and observations (Figs. 4, 5); this is also seen in other members (Fig. A4). This decoupling has major consequences for the simulated Antarctic-wide accumulation rate and the corresponding SST trends. It suggests either misrepresented processes or a missing forcing.

In the context of anthropogenic warming, the lack of an observed increase in Antarctic-wide accumulation for 1980 to near present has been puzzling. Surface cooling (or muted warming) of the Southern Ocean is the solution to this puzzle. In a prescribed SST and sea ice experiment, CESM2-GOGA, no trend in the Antarctic-wide accumulation rate occurs, in contrast to the coupled Large Ensemble in which accumulation steadily increases. In the coupled CESM1, the snow accumulation trend is reduced by prescribing ice shelf meltwater fluxes. Qualitatively similar results are obtained from a CESM2 experiment nudged to observed Southern Ocean SST anomalies (Fig. B6; Kang et al., 2023).

Accurate projection of future changes in snow accumulation is hampered by uncertainty in the evolution of SST and atmospheric circulation trends. To tentatively project potential sea level mitigation from increased accumulation to 2050, we use the lower bound of the GHG + AAER ensemble at 2050 in reference to the value given by [GHG] + [AAER] at the year 2000 (Fig. 11d–e). This is based on the idea that [AAER] has too strong of a cooling and effectively does the work of both aerosols and omitted meltwater to suppress the GHG-driven accumulation increase. Using the 5 % lower bound at 2050 gives an average rate of 0.5 mm yr⁻¹ of snowfall-related sea level mitigation for 2001–2050. This is less than the widely used CESM2 Large Ensemble projects, and less than the ice sheets' current rate of contribution to sea level rise (Stokes et al., 2025). If meltwater continues to dampen surface warming (Sadai et al., 2020, 2025) snow accumulation will have a hard time keeping up with ocean-driven dynamic mass loss. While the uptick in intense atmospheric river events, recent sea ice loss and warm Southern Ocean conditions could indicate that increased snow accumulation will provide significant sea level mitigation, strong warming will mean more mass lost to surface melting and runoff (Gilbert and Kittel, 2021) especially after 2050 (e.g. Kittel et al., 2021; Jourdain et al., 2025). As warming continues, some precipitation will fall as rain (Vignon et al., 2021). The observed uptick in atmospheric river events that bring intense heat and moisture to the ice sheet (e.g. Blanchard-Wrigglesworth et al., 2023; Wille et al., 2025) is a worrying trend that may mean that rainfall and mass loss processes become important sooner than model simulations would suggest (Mottram et al., 2025).

Spatially, the wind and snow accumulation patterns that were dominant in the recent past should continue, marked by relatively more snow in the Peninsula and Queen Maud Land, and relatively less snow in Wilkes Land and a positive SAM pattern in the pressure field (Fig. 11a–c). Although these maps (and the all-forcings CESM2-LE, not shown) indicate that the westerlies will continue to strengthen through the mid- 21st Century due to continued greenhouse gas increases, this could be a model-dependent result stemming from CESM2's high climate sensitivity and a scenario-dependent result associated with the high-emissions SSP3.7 scenario. Under a lower emissions scenario and/or in a model with lower climate sensitivity we might expect stratospheric ozone recovery to lead to a weakening of the westerlies by the mid- 21st Century (Arblaster et al., 2011). However, on the annual-mean basis the positive SAM trend is projected to continue across many models in intermediate- to high- emissions scenarios (Purich et al., 2025).

This work has several limitations, the most prominent being that it features a single low-resolution model, CESM2. The advantage of relying on CESM2 is that it enables a robust assessment of the roles of anthropogenic forcing and internal variability across a hierarchy of experiments from free-running fully coupled to nudged coupled to prescribed SST/ sea ice experiments. We are not aware of comparable studies that have used CMIP6 or other models to make similar assessments of Antarctic accumulation trends across a hierarchy of experiments. As such, we encourage future coordinated efforts like CMIP7 and beyond to facilitate this type of study across multiple models.

Higher resolution modeling might change some details of our results and the relative importance of meltwater and winds, but should not change our overall interpretation. Regional high resolution would resolve finer details of the orographic patterns in accumulation, better represent snowfall in coastal regions and may change the response of processes like surface melting to warming (Datta et al., 2023; Noël et al., 2023; Yin et al., 2025). Regional atmospheric grid refinement over the recent historical period in an atmosphere-only configuration produces differences in the simulated snow accumulation conforming strongly to topography and enhanced poleward moisture transport (Datta et al., 2023). Global high resolution enhances the linkage of the Southern Ocean with the tropical Pacific (Chang et al., 2020; Yeager et al., 2023; Kang et al., 2026), which is too weak in most low-resolution models (Chung et al., 2022; Dong et al., 2026). The linkage is relatively strong in CESM2 due to its strong shortwave cloud feedbacks (Kim et al., 2022; Zhang et al., 2026) and improved representation of the ITCZ (Dong et al., 2026). Low resolution and/or weak shortwave cloud feedbacks could be reasons why CESM1 is not very responsive to meltwater fluxes, requiring a perturbation of 2000 Gt yr⁻¹ to produce a significant response (Pauling et al., 2016). Alternatively, if the model's SSTs are more sensitive to winds at high resolution, meltwater may not need to be invoked to explain observation-model discrepancies, as suggested by Kang et al. (2026). Our results imply that the ability of a model to reproduce Antarctic accumulation trends is an important benchmark for any high-resolution experiment that purports to solve observation-model trend discrepancies. We propose adding Antarctic evaluation metrics to high-resolution model intercomparisons such as Dhame et al. (2025) and meltwater intercomparisons like Zhang et al. (2026).

Observations and/or reconstructions underpin any model evaluation study. We have based our model evaluation on the MT19 reconstruction, which is similar to the kriging reconstruction of Wang and Xiao (2023). Applying two distinct methodologies to the same proxy network, Wang and Xiao (2023) find significant differences between reconstructions. This methodological dependence might be alleviated if more ice core data were available, especially on the East Antarctic plateau (Eswaran et al., 2024) and West Antarctic coastal margin (Neff, 2020). More observations would avoid the need to indirectly infer what trends have occurred in these regions. We have also used CESM2-LE PDA, which is dynamically consistent with the CESM2. Long-term (1950–2024) trends in the ERA5 reanalysis qualitatively support the overarching interpretation we have made from these datasets (Fig. C1).