Diverging future surface mass balance between the Antarctic ice shelves and grounded ice sheet

. The future surface mass balance (SMB) will inﬂuence the ice dynamics and the contribution of the Antarctic ice sheet (AIS) to the sea-level rise. Most of recent Antarctic SMB projections were based on the 5th phase of the Coupled Model Intercomparison Project (CMIP5). However, new CMIP6 results have revealed a +1.3° C higher mean Antarctic near-surface temperature than in CMIP5 at the end of the 21st century enabling estimations of future SMB in warmer climates. Here, we investigate the AIS sensitivity to different warmings with an ensemble of four simulations performed with the polar regional 5 climate model MAR forced by two CMIP5 and two CMIP6 models over 1981–2100. Statistical extrapolation allows us to expand our results to the whole CMIP5 and CMIP6 ensembles. Our results highlight a contrasting effect on the future grounded ice sheet and the ice shelves. The SMB over grounded ice is projected to increase as a response to stronger snowfall,


S1 Evaluation of MAR forced by ERA5
Although an exhaustive assessment of MAR is not the main purpose of this study, we present here a short evaluation of MARv3.11 forced by ERA5. The SMB and near-surface climate modelled by MAR have already been evaluated over Antarctica (Kittel et al., 2018;Agosta et al., 2019;Mottram et al., 2020), but with previous versions of the model and with the ERA-Interim reanalysis as forcings. We therefore compare MARv3.11 using the same methods to the same 184 AWS and 983 5 averaged-SMB observations used in Mottram et al. (2020) to evaluate the new version of MAR and its results forced by the new ERA5 reanalysis, which replaced ERA-Interim, hitherto considered to be the best reanalysis in Antarctica (Agosta et al., 2015). We refer to Mottram et al. (2020) for a more detailed descriptions of the observation dataset and the method) MARv3.10 forced by ERA-Interim (as used in Mottram et al. (2020)) is presented for comparison. Since the two versions of MAR are not forced by the same reanalyses, the purpose of this section is to demonstrate that the latest version of MAR forced by ERA5 10 performs equally well (not to discuss differences between the two versions in detail) and can be considered as a reference to evaluate our simulations forced by ESMs over the historical period.

S1.1 Comparison against near-surface climate observations
Using either MARv3.11 (forced by ERA5; MAR(ERA5)) or MARv3.10 (forced by ERA-Interim) yields very equivalent statistical performance when evaluated against near-surface meteorological observations over 1987 -2015 (Tab. S1). MAR(ERA5) 15 displays a non significantly-improved correlation with surface pressure as a result of a stronger upper nudging and/or the more sophisticated (i.e, more recent) forcing. The representation of the near-surface wind speed is improved while MAR still underestimates the mean wind speed. The model tends to overestimate low wind-speed values but underestimates high wind-speed values resulting in this negative bias. Regarding the near-surface temperature, the mean annual (summer) positive bias is increased by +1.01°C (+0.52°C). This bias reflects a difference between low and high temperatures (and is therefore linked to 20 elevation and seasons). MAR overestimates low temperatures (especially on the plateau or in winter) while it slightly underestimates the high temperatures (close to 0°C). This likely results from a bias in the radiative scheme itself resulting from the version of the radiative scheme (from the ERA-40 reanalysis) and/or the low-sophisticated one-moment cloud scheme in MAR.

25
Both versions of MAR also have equivalent statistics for the SMB comparison (Fig. S12). The few slight improvements in one indicator are compensated for by alterations in other statistics, meaning that MAR(ERA5) is close to MARv3.10, the latter being assessed as one of the best models to represent the Antarctic SMB (Mottram et al., 2020). In general, MAR can reproduce the strong coast-to-plateau SMB gradient with high SMB values over low-elevations margins and weak accumulation over the high plateau.
Annual and summer statistics are given for surface pressure, near-surface temperature and wind speed.

S1.3 Comparison against melt estimates
The amount of melt water is low over the present climate, but is expected to increase in the future and may therefore become a major surface process. This is why we also compared the melt simulated by MAR(ERA5) to the annual AWS-forced estimates from Jakobs et al. (2020). MAR melt values were computed using a four-nearest inverse-distance-weighted method for all the AWS whose the elevation difference with MAR does not exceed 250m. Figure S2 shows that MAR(ERA5) can correctly 35 reproduce the annual meltwater estimates. MAR underestimates the relatively-weak melt at AWS14 (located over Larsen C North). The 35km resolution of our simulations is probably still too coarse to correctly reproduce the Foehn winds inducing surface melting (Datta et al., 2019).  S4). This however leads to reduced surface melting (-72 Gt yr −1 ).
MAR(CNRM-CM6-1) simulates nearly the same snowfall amount as MAR(ERA5) but has a higher SMB RMSE due to a 60 less accurate spatial representation of the precipitation. This results from an overestimation of the precipitable water combined to a higher mean sea level pressure in CNRM-CM6-1 potentially reducing cyclonic activity. MAR(CNRM-CM6-1) underes-  However, these differences are non significant over the margins, the Ronne ice shelf excepted (Fig.S4).
As it simulates lower snowfall amounts, MAR(CESM2) slightly underestimates the mean integrated SMB. However, MAR(CESM2) represents a stronger accumulation over the area between the Peninsula, Queen Maud Land and Enderby Land (Fig. S3). This results from the significant overestimation of the precipitable water and the sea level pressure in CESM2 over this area. On the contrary, MAR(CESM2) simulates a lower accumulation over Wilkes Land and the Amundsen sector. CESM2 is colder than 70 ERA5 but the difference is reduced in summer (Agosta et al., in preparation), leading to mostly non significant temperature anomalies (Fig.S4) and lower melt in MAR(CESM2).
In general, the SMB downcalled by MAR forced by the 4 ESMs is close to MAR(ERA5). The anomalies of the annual mean SMB are lower than the interannual variability of the SMB over the historical period. It is also important to note that 6 Figure  the spatial and integrated anomalies are close to (or even lower than) the differences between several RCMs all forced by 75 ERA-Interim (Mottram et al., 2020). This suggests a good ability of the different simulations to closely reproduce the SMB over the present climate and gives some confidence in results of the projections.        Figure S12 compares the modelled MAR SMB anomalies to the ESM reconstructed SMB anomalies. The reconstruction based on the temperature anomaly accurately reconstructs the modeled SMB over the grounded ice (RMSE = 68 Gt yr −1 , i.e 80 lower by 99 Gt yr −1 than the present interannual grounded-SMB variabily in our MAR(ERA5) reference simulations), but is slightly less precise over the ice shelves (RMSE = 38 Gt yr −1 , i.e larger by 22 Gt yr −1 than the present interannual ice-shelf SMB variability). This is mainly due to the large decrease in SMB for MAR(CNRM-CM6-1) that is not fully represented by the regression (also valid for the grounded fit). The projected changes with the strongest warming are however much larger than the error of the regression. Moreover, among all CMIP5 and CMIP6 models, CNRM-CM6-1 projects the strongest warming 85 in 2100, thus allowing us to use our regression to reconstruct the SMB for all CMIP5 and CMIP6 models. Note that using a similar method based on the link between individual SMB components (snowfall, rainfall and ablation) and ESM near-surface anomalies also yields a similar reconstruction. This demonstrates that the future response of the surface of both ice shelves and grounded ice can be mainly determined using the temperature warming until the end of the 21st century.