Effect of uncertainties of Southern Ocean surface temperature and sea-ice change on Antarctic climate projections

In this study, the atmospheric model ARPEGE is used with a stretched grid in order to reach a average horizontal resolution of 35 kilometers over Antarctica. Over the historical period (1981-2010), ARPEGE is forced by the historical sea surface conditions (SSC, i.e. sea surface temperature and sea-ice concentration) from MIROC and NorESM1-M CMIP5 historical runs and by observed SSC (AMIP-experiment). These three simulations are evaluated against ERA-Interim for atmospheric general circulation and against MAR regional climate model and in-situ observations for surface climate. As lower 5 boundary conditions for simulations for the period 2071-2100, we use SSC from coupled climate model CMIP5 simulations of the same models following the RCP8.5 emission scenario. We use these output both directly and with an anomaly method based on quantile mapping. We assess the uncertainties linked to the choice of the coupled model and the impact of the method (direct output and anomalies). For the simulation using SSC from NorESM1-M, we do not find significant changes in climate change signals over Antarctica when using bias-corrected SSC. For the simulation using MIROC-ESM output, an additional increase 10 of +185 Gt yr−1 in precipitation and of +0.8 K in winter temperatures for the grounded Antarctic ice-sheet was obtained when using bias-corrected SSC.

katabatic winds that blow at the ice sheet surface is also generally improved with a better representation of the topography (e.g., van Lipzig et al., 2004).
In this study, we use CRNM-ARPEGE, the atmosphere general circulation model (AGCM) from Météo-France, with a stretched grid allowing a horizontal resolution of about 40 km over the whole Antarctic continent, to dynamically downscale different climate scenarios. As a global atmospheric model, ARPEGE is driven by prescribed SSC, but does not require any 5 lateral boundary conditions. More details on the model setup are given in section 2.1. This method has some advantages over the more commonly use of RCM. It is possible to use observed SSC at the present and model-generated SSC anomalies for projections (e.g., Krinner et al., 2008). When such an anomaly method is used, the results do not absolutely require the AOGCM used as driver for sea surface conditions to represent the atmospheric general circulation and its variability in the region of interest realistically in every respect. Using a stretched grid GCM also allows better taking into account potential 10 feedbacks and teleconnections between the high-resolution region which the focus lies on, and other regions of the world.
Rather unsurprisingly, several studies showed that AGCMs produce a better representation of atmospheric general circulation and a better repartition of precipitation when forced by observed instead of simulated SSC (Krinner et al., 2008;Ashfaq et al., 2011;Hernández-Díaz et al., 2017). These studies also showed that bias correction of SSC before the downscaling of future climate scenarios gives significantly different results with respect to original scenarios. For these reasons, we performed a 15 bias-correction of SSC using a quantile mapping method for SST and an analog method for SIC following the methods and recommendations described in Beaumet et al. (2018). We reduced our ensemble of possible simulations to the choice of two AOGCMs from the CMIP5 experiment : MIROC-ESM and NorESM1-M. As they are bias-corrected in a second step, the main criterion was the amplitude of the climate change signal in the oceanic forcings coming from these two models, not the realism of the simulated present-day SSC. The short analysis 20 on which we based our model choice is described in section 2.2. We also performed an AMIP-style control simulation for the period 1981-2010 in which CNRM-ARPEGE is forced by observed SST and SIC coming from PCMDI data set (Taylor et al., 2000). CNRM-ARPEGE was also forced by the original oceanic SSC coming from the historical simulations of MIROC-ESM andNorESM1-M (1981-2010) and from projections under the radiative concentration pathway RCP8.5 (Moss et al., 2010) carried out with the same two models (2071-2100). In section 3.1, we present the ability and limitations of CNRM-ARPEGE 25 to represent current Antarctic climate as well as the differences between the AMIP experiment and the experiments forced by oceanic forcings coming from historical simulations of CMIP5 GCMs. In section 3.2, we present modelled climate at the end of the 21 st century by CNRM-ARPEGE and the differences in climate change signal between scenarios realized with bias-corrected and original SSC from the RCP8.5 scenarios of MIROC-ESM and NorESM1-M.

Sea Surface Conditions in CMIP5 AOGCMs
SSC forcings have been identified as key forcings for the evolution of the Antarctic climate of the continent (Krinner et al., 2014;Agosta et al., 2015). In this study, SSC obtained from CMIP5 projections are bias-corrected using recommendations and 3 The Cryosphere Discuss., https://doi. org/10.5194/tc-2018-231 Manuscript under review for journal The Cryosphere Discussion started: 3 December 2018 c Author(s) 2018. CC BY 4.0 License. methods from Beaumet et al. (2018) before being used as surface boundary conditions for the atmospheric model. Therefore, the importance of the bias of each CMIP5 model for the reconstruction of oceanic conditions around Antarctica in their historical simulation is reduced. There is however a limitation in the previous statement, as the analog method used to bias-correct SIC runs into trouble when the bias is so large that sea ice completely disappears over wide areas for too long. Besides this caveat, however, the choice of CMIP5 AOGCMs used in this study was guided by compliance to desired characteristics of the 5 climate change signal rather than by the skills of the models in reproducing SSC in the historical periods.
Therefore, we identified CMIP5 models with the highest and lowest climate change signal by the end of the 21 st century considering only SSC in the Southern Ocean, in order to span the uncertainty range associated with model response. We computed the relative evolution of integrated winter SIE over the whole Southern Ocean between the historical simulation (reference period: 1971-2000) and the RCP8.5 scenario (reference period: 2071-2100) for 21 AOGCMs from CMIP5 experiment. The 10 CMIP5 ensemble was reduced to 21 because some models sharing the same history of development and high code comparability as others have been discarded. The model list is the same as in Krinner and Flanner (2017) and can be seen in the Fig. 1 legend. We also looked at the mean summer SST increase South of 60°S for the same reference periods. In order to be consistent with periods of maximum (minimum) SIE, seasons considered in this analysis are slightly shifted, and winter (summer) correspond here to the period August-September-October, ASO (February-March-April, FMA). 15 The results of the computation can be seen in Figure 1, which displays the relative decrease of SIE in winter (ASO) in the RCP8.5 scenario as a function of the value of the mean winter SIE in the historical simulation. The four models with the highest decrease in SIE are CNRM-CM5 (-62.4%), GISS-E2-H (-53.4%), inmcm4 (-47.9%) and MIROC-ESM (-45,2%). Because of the above-mentioned limitation of the bias-correction method, the first three GCMs cannot be selected due to a large negative bias of winter and spring SIE. We therefore selected MIROC-ESM as representative for models projecting a large 20 climate change signal for sea ice around Antarctica. If we consider weak climate change signals, MIROC5 shows the lowest decrease (-1,5%) followed by NorESM1-M (-13,6%). For the same reasons of limitations of the bias correction method, we dismissed MIROC5 and kept NorESM1-M as representative for a weak climate change signals in the SSC around Antarctica.
The impact of primarily considering changes in winter SIE rather than in summer SST is limited as the climate change signal for these two variables are strongly linked (R 2 =0.96). For summer SSTs, MIROC-ESM shows the 6 th largest increase(+1.8K) 25 while NorESM1-M exhibits the second lowest (+0.4K).

CNRM-ARPEGE set-up
We use version 6.2.4 of AGCM ARPEGE, a spectral primitive equation model from Météo-France, CNRM (Déqué et al., 1994). The model is run at T255 truncation with a 2.5 zoom factor and a pole of stretching at 80°S and 90°E. With this setting, the horizontal resolution in the Antarctic ranges from 35 km near the stretching pole on the Antarctic Plateau to 45 km at Historical Winter (ASO) SIE, 1971SIE, -2000 km 2 ) PCMDI (1971PCMDI ( -2000 CNRM Historical Antarctic Winter (August-September-October: ASO) Sea-Ice Extent (SIE, in millions of km 2 ) as function of the relative decrease of Winter SIE in the RCP8.5 scenario for the period 2071-2100 with respect to the reference period 1971-2000. The mean Winter SIE in the observations for the historical reference period is indicated by the horizontal black line (PCMDI 1971(PCMDI -2000 layers and for the percolation and refreezing of liquid water in the snow pack. Over the ocean, we use a 1D version of sea-ice model GELATO (Salas y Mélia, 2002) which means that no advection of sea-ice is possible. The sea-ice thickness is prescribed following the empirical parametrization used in Krinner et al. (1997Krinner et al. ( , 2010 and described in Beaumet et al. (2018). The use of GELATO is therefore limited to the computation of heat and moist fluxes in sea-ice covered regions and also allows taking into account for the accumulation of snow on top of sea-ice.

5
In each ARPEGE simulation, the first two years are considered as a spin-up phase for the atmosphere and the soil, and are therefore discarded from the analysis. The characteristics of the different ARPEGE simulations presented in this paper are summarized in the table1.

Evaluation for Present Climate
The ability of ARPEGE model to reproduce atmospheric general circulation of the Southern Hemisphere is assessed by comparing sea level pressure (SLP) and 500 hPa geopotential (Z500) South of 20°S to those of ERA-Interim reanalysis (ERA-I).
For surface climate of the Antarctic continent, several studies have shown that (near) surface temperatures from ERA-I are not 5 reliable (Bracegirdle and Marshall, 2012;Jones and Harpham, 2013;Fréville et al., 2014), as the reanalysis is not constrained by enough observations and because the boundary layer physics of the model fails to successfully reproduce strong temperature inversions near the surface that characterize the climate of the Antarctic Plateau. As a consequence, near-surface temperatures in Antarctica from ARPEGE simulations are evaluated using observations from the SCAR READER data base (Turner et al., 2004) as well as temperatures from a MAR RCM simulation in order to increase the spatial coverage of the model evaluation.

10
Modèle Atmosphérique Régional (MAR, Gallée and Schayes (1994)) has been one of the most successful RCMs in reproducing the surface climate of large ice-sheets such as Greenland (Fettweis et al., 2005;Lefebre et al., 2005) and Antarctica Amory et al., 2015;Agosta et al., 2018). In this evaluation, we compare ARPEGE near surface temperatures to those of an ERA-I driven MAR simulation (hereafter MAR-ERA-I) at a similar horizontal resolution of 35 kilometres (Agosta et al., 2018). SMB of the grounded Antarctic Ice Sheet and its components from ARPEGE simulations are compared to outputs  -20) historical SSC can be seen respectively in Fig. 2b and Fig. 2c. The pattern and the magnitude of the errors are similar to those of the ARP-AMIP simulation in summer (DJF). The root mean square errors (RMSE) per seasons for each simulations are summarized in Table 2. In winter (JJA), spring (SON) and autumn (MAM) the errors are substantially larger in ARP-NOR-20 and ARP-MIR-20 than in ARP-AMIP. The patterns of the errors and the ranking of simulations scores are similar for 500hPa geopotential height than for SLP (not shown).

5
The mean atmospheric general circulation in each simulation has also been compared and evaluated against ERA-I by analyzing the 850 hPa eastwards wind component (referred to as westerly winds in the following) latitudinal profile, as well as the strength (m/s) and position (°Southern latitude) of the zonal mean westerly wind maximum or westerly "jet" (Fig. 3). In this figure, results are only presented for the annual average, as the differences between simulations or with ERA-I do not depend much on the season considered (not shown). When compared with ERA-I, ARP-AMIP and ARP-MIR-20 are closer to ERA-I when 10 the westerly winds maximum strength is considered, and ARP-NOR-20 when it is its position. With respect to ARP-AMIP, ARP-NOR-20 displays a much weaker and poleward surface westerly jet in all seasons, while ARP-MIR-20 is characterized by a lower latitude westerly wind maximum of comparable strength.

Near Surface Temperatures
Screen level (2m) air temperatures (T 2m ) from ARP-AMIP simulation are compared to those of the ERA-Interim driven MAR Considering errors on surface temperatures of the Antarctic Plateau as large as 3 to 6K for ERA-I reanalysis in all seasons (Fréville et al., 2014), the magnitude of the errors in this region in ARP-AMIP simulation is encouraging. The error for Amundsen Scott station is even insignificant at p=0.05 level in all seasons but autumn (MAM). The warm bias on the large 5 ice-shelves is due to the fact that ice-shelves are not considered as land in the ARPEGE version used. In order to correct this weakness, we prescribed an SIC of 100% and a thickness of 40 m in grid points corresponding to ice-shelves. Even if this reduced the initial bias by about 5K, it did not prevent the warm bias from still being as high as 12K in the center of the Ross Ice-Shelf in Winter. Part of the errors on this ice-shelf are also likely due atmospheric general circulation errors, but this issue will require further investigation. As a consequence of these large biases in temperatures and surface climate over large  level. For the ARP-NOR-20, differences ranging from 0,4K to 1,2K in autumn, winter and spring are significant as well. 10 The Cryosphere Discuss.

Southern Ocean
Gough

Surface Mass Balance
In this study, SMB from ARPEGE simulations is defined as the total precipitation minus the surface snow sublimation/evaporation minus run-off. Differences between ARP-AMIP and MAR-ERA-I total precipitation, snow sublimation and SMB (in mm of water equivalent per year) for the reference period 1981-2010 can be seen in Fig. 5. As differences in run-off are restricted to the ice-shelfs and some very localized coastal areas, their spatial distribution is not displayed in this figure. Yearly mean SMB, 5 total precipitation, surface sublimation, run-off, rainfall and melt, integrated over the whole Antarctic GIS for the different ARPEGE simulations, for MAR-ERA-I and from other studies are presented in Table 5.
At the continental scale, we can see that estimates of the SMB of the ice-sheet from the ARP-AMIP simulation resemble those from state of the art polar-oriented RCM MAR and RACMO2. However, higher total precipitation values in ARPEGE-AMIP are compensated for by much higher values of sublimation/evaporation of surface snow and, to a lesser extent, higher run- This is consistent with ARP-AMIP being systematically 1 to 3K warmer than MAR-ERA-I in summer in those areas. Run-off at the continent scale is eight times higher in ARP-AMIP than in MAR-ERA-I, which is also most likely a consequence of warmer coastal areas in ARPEGE in summer. However, inter-annual variability is very high in the simulated ARPEGE run-off, 20 and so it is in MAR-ERA-I (σ is at least 50% of the mean). If we have a closer look at the values of rainfall, surface sow melt and run-offs in the three present-day ARPEGE simulations in Table 5, the ratio between inputs of liquid water into the snow pack (rainfall + surface snow melt) and the water run-off that finally leaves the snow-pack is about 1/4. In MAR-ERA-I and in RACMO2-ERA-I, this ratio is about 1/20. This means that although the snow surface scheme SURFEX-ISBA used in ARPEGE is able to model storage and refreezing of liquid water in the snow-pack, the retention capacity of the Antarctic snow 25 pack underestimated with respect to MAR and RACMO2.
In ARP-MIR-20 simulation, snow sublimation and evaporation, run-off and melt were found significantly lower than in ARP-AMIP, which is consistent with this simulation being 1.5K cooler in summer (DJF). The effect of driving ARPEGE by biased SSC for the modelling of Antarctic precipitation is discussed in the supplementary material (see Sec. C).

Climate change signal
(Vaughan et al., 1999) 1811 Table 5 Earth System Model historical simulation , values for the total ice-sheet from Lenaerts et al. (2016) from the respective coupled model . For scenarios realized with bias correction of the SSC (ARP-NOR-21-OC and ARP-MIR-21-OC), the reference simulation for the historical period is ARP-AMIP (observed SSC).
The primary goal here is to evaluate the uncertainty in climate change signals for Antarctica associated with oceanic forcing coming from coupled models and the changes coming from the bias correction of the SSC.

5
Climate change signals in mean SLP for the different RCP8.5 scenarios realized with ARPEGE can be seen in Fig. 6. All scenarios show a pressure increase at mid-latitudes (30-50°S) and a decrease around Antarctica. This corresponds to a shift of the circum-antarctic low pressure belt towards the continent (positive phase of the SAM) and is a generally expected consequence of 21 st century climate forcing (Kushner et al., 2001;Arblaster and Meehl, 2006). This pattern (increase at mid-latitude, decrease around Antarctica) is sharper in scenarios realized with MIROC-ESM SSC.

10
Differences in the climate change signal for ARP-NOR-21-OC and in ARP-NOR-21 with respect to their corresponding references in historical climate are small. The ASL deepens more in the scenario realized with non bias-corrected SSC (ARP-NOR-21) in winter while it is the opposite in summer. Differences in SLP climate change signal are more obvious in the scenarios realized with MIROC-ESM SSC. In the scenario realized with non bias-corrected SSC (ARP-MIR-21), the intensification of the low pressure systems around Antarctica in winter is clearly organized in a 3-wave pattern (Fig. 6b). In ARP-MIR-21-OC, 15 the JJA pressure decrease is rather organized in a dipole with one maximum of pressure decrease centered the Eastern side of the Ross Sea and another one West of the Weddell Sea. As a result, the 3-wave pattern is clearly noticeable in the difference between the two scenarios climate change signals (Fig. 6b, right). In summer, the differences between the two simulations are  weaker and mainly consist of a sharper pressure increase at mid latitudes in ARP-MIR-21-OC.
Regarding the changes in westerly wind maximum strength (Table 6), the differences between the two scenarios using NorESM1-M SSC are once again limited. We can however mention a -1.4°higher decrease in westerly winds maximum position in the scenario using bias-corrected SSC. Differences in changes in position and strength are not substantial between ARP-MIR-21 and APR-MIR-21-OC. Compared to scenarios realized with SSC from NorESM1-M, these scenarios show a slightly larger 5 increase in jet strength and a much larger poleward shift, although this difference is reduced when comparing scenarios with bias corrected SSC.

Near-surface Temperatures
The mean yearly T 2m increase for the Antarctic GIS using SSC from NorESM1-M rcp8.5 scenario is 2.9±1.0 K using original SSC (ARP-NOR-21) and 2.8±0.8 K using bias-corrected SSC (ARP-NOR-21-OC). For scenarios using SSC from MIROC-10 ESM, these temperatures increases are respectively 3.8±0.7 K and 4.2±1.0 K. The differences in yearly T 2m increase using bias-corrected SSC are found non significant at p=0.05 level in both cases. T 2m increase per season can be seen in Tab. 7.
Only a +0.8 K difference in winter temperature increase in ARP-MIR-21-OC with respect to the scenario with original SSC is found significant. At the regional scale ( For scenarios using SSC from NorESM1-M, no seasonal difference was found significant at the scale of the ice-sheet although a 0.5K weaker temperature increase in summer for ARP-NOR-21 is close to the significance threshold. However, if we look at regional warming (Fig. 7a), we can see that for large areas covering the center of East Antarctic Plateau and coastal areas, the regional warming is 0.5 to 1K higher in winter and 0.5 to 1K lower in summer in the scenario with bias-corrected SSC.   In each scenario, the sublimation increases by about 20 to 25% with respect to the corresponding references in the historical 10 period. Run-off and melt increase by about a factor 4 in scenarios with NorESM1-M SSC and by factors ranging from 6 to 10 in scenarios with MIROC-ESM SSC. This, however, does not prevent these components to remain one order of magnitude smaller than total precipitation. As a consequence, increases in SMB are essentially determined by the increases in total precipitation.
In future climate simulated by ARPEGE, the ratio between liquid water inputs (rainfall + melt) and liquid water leaving the snow-pack (run-off) remains around 1/3. As the change in SMB is mainly the result of change in total precipitation, we only  differences in precipitation changes between the simulations with MIROC-ESM original SSC. The relative mean precipitation changes (in%) and associated standard deviation for four RCP8.5 scenarios realized in this study can be seen in Fig. 9.

Reconstruction of the historical climate
The atmospheric model ARPEGE correctly captures the main features of the atmospheric circulation around Antarctica. The 5 three local minima in SLP and 500hPa geopotential, generally present around 60°W, 90°E and 180°E, are well reproduced in the ARPEGE-amip simulation (see Fig. D1a). However, there is a positive SLP bias in the seas around Antarctica, particularly in the ASL sector, and a negative bias in mid-latitudes (30-40°S), especially in the Pacific sector. This bias structure in the Southern Hemisphere is present in many coupled and atmosphere-only GCMs. Its consequence is an equator-ward bias in the position of the surface jet associated with westerly winds (Bracegirdle et al., 2013). The use of observed SSC (ARP- Regarding surface climate, ARPEGE also correctly reproduces Antarctic T 2m except over the large ice-shelfs. The T 2m error with respect to MAR is generally lower than 3K over most of the GIS. There is a warm bias on the ridge of the Antarctic Plateau in winter. However, the magnitude of these errors (+1.5K at Amundsen-Scott, +3.4K at Vostok) is to be compared with 5 much larger biases in other GCMs or even in reanalysis, as most models usually fail to capture the strength of the near-surface temperature inversion. The cold bias of ARPEGE on the Antarctic Peninsula, especially in winter, can largely be explained by atmospheric circulation errors, as an underestimate of the depth and/or recurrence of the ASL leads to an underestimate of mild and moist flux from the North-West onto the Peninsula.
The GIS SMB in ARP-AMIP simulation (2191±106 Gt yr−1) falls within the 1-standard deviation (1σ) uncertainty range with 10 respect to estimates using the MAR RCM, and concurs with studies using other RCMs and GCMs or independent estimates.
However, it has to be mentioned that higher precipitation rates in the ARP-AMIP simulation than in MAR and RACMO2 (about 2.5σ) are compensated for by a much stronger surface snow sublimation in the ARPEGE simulations. Some of the differences with MAR-ERA-I in the spatial distribution of precipitation rates in the ARP-AMIP simulation can also be linked to errors in atmospheric general circulations. These errors are certainly part of the explanation for ARPEGE being wetter in is noteworthy that all scenarios agree on a (slight) precipitation decrease in Marie-Byrd Land and western Ross ice shelf (see Fig. 9). Victoria, Adélie, and Wilkes Land as well as the eastern side of the AP are also regions of lower precipitation increase compared to the rest of the continent. All these regions show high uncertainty in future changes in precipitation estimated in this study (Fig. 9, high value of standard deviation when compared to mean change). Lower increase or slight decrease in precipitation in Marie-Byrd Land are also present in other studies (Krinner et al., 2008;Lenaerts et al., 2016). These results from MIROC-ESM, the SMB (precipitation) increase obtained with ARPEGE range between 5.2 and 6.3 %.K −1 (6 and 7.4 %.K −1 ). This is in the range of values obtained in previous studies (Agosta et al., 2013;Ligtenberg et al., 2013;Krinner et al., 2014;Frieler et al., 2015;Palerme et al., 2017). Only the SMB increase obtained with bias-corrected SSC from MIROC-ESM, 7.9%.K −1 is above the higher end values of previous studies. As in Krinner et al. (2014), we found that regional precipitation increases depend on the AOGCM chosen as SSC source and on their bias-correction or not. For a weaker climate change signal 5 such as the one coming from NorESM1-M SSC, we found no significant difference in climate change signals at the continent scale over Antarctica using bias corrected or original SSC to drive ARPEGE. However, for a more dramatic change in SSC such has the one coming from MIROC-ESM, we found a +14% higher precipitation increase using bias-corrected SSC.
Finally, we draw the attention on the fact that when considering absolute values rather than climate change signals, both annual total precipitation rates and SMB are significantly different than when using bias corrected SSC. In the scenarios with original 10 SSC, the annual GIS SMB at the end of current century is slightly higher in ARP-NOR-21 than in ARP-MIR-21, which is a bit surprising considering the very weak decrease in sea-ice around Antarctica in NorESM1-M RCP8.5 scenario. When using bias-corrected SSC, the order is reversed and SMB values are respectively 2585 Gt yr −1 and 2914 Gt yr −1 , which is more intuitive considering much larger decrease in sea-ice in MIROC-ESM RCP8.5 scenario.

15
The late winter (ASO) and late summer (FMA) differences between historical SST and SIC from NorESM1-M and MIROC-ESM and the observations, as well as the same differences between SSC of their RCP8.5 scenario and their bias-corrected equivalent are displayed in the annex ( Fig. A1 and A2). The differences in SSC used to drive the atmospheric model are, unsurprisingly, extremely similar between historical and future climate experiments. For the SST, the similarity is almost perfect and for SIC, the patterns are the same, but given the decrease in SIC in future climate, they are shifted poleward. 20 Has the introduction of the same SSC "biases" with respect to the observed or bias-corrected references yielded the same responses of the atmospheric model in the historical and future climates? This consistency of the response of the atmospheric model is considered here as being the key for having the same climate change signals between experiments using original SSC from the CMIP5 model and experiments considering the climate change signal between the AMIP experiment and the corresponding bias-corrected projected SSC. 25 For simulations using SSC from the NorESM1-M model, the consistency of the response of the atmospheric model is obvious.
The similarity in the differences between ARP-NOR-20 and ARP-AMIP with differences between ARP-NOR-21 and ARP-NOR-21-OC is strong for most climate variables, e.g. SLP, 500 hPa geopotentials, stratospheric temperatures, 500hPa zonal wind, near-surface atmospheric temperatures...(an example for SLP can be seen in Fig. D2). The most interesting feature in this perspective is that in both historical and future climate, the ARPEGE simulations forced by NorESM1-M original SSC For the simulations realized with oceanic forcings from MIROC-ESM, the consistency of the response of the atmospheric 10 model is less generalized. Some changes in the differences between simulations forced with original SSC and those forced by their bias-corrected references are noticeable in winter and autumn SLP and zonal wind speed (an example for SLP can be seen in Fig. D3). The main result here, as a consequence of these differences, is a total precipitation difference in the RCP8.

Implication of Sea Surface Conditions selection
In many cases, it has been reported that selecting best skilled models for a given aspect of the climate system helped in better constraining the associated uncertainties on climate change signal (e.g., Massonnet et al., 2012). Here, because we use biascorrection of the SSC, this aspect has reduced importance. Our aim is to cover as much as possible, while performing a limited number of climate projections, the range of uncertainties associated with the evolution of the Southern Ocean surface condition 30 for the Antarctic climate projection as it was shown to be its primary driver (Krinner et al., 2014). The fact that biases on largescale atmospheric circulation of coupled climate models were shown to be highly stationary under strong climate change (Krinner and Flanner, 2017) and that the response of ARPEGE atmospheric model to the introduction of the same SSC "bias" was shown to be mostly unchanged in future climate support this approach.
The warming signal for the Antarctic ice-sheet in the CMIP5 model ensemble RCP8.5 scenario is evaluated to be 4±1°C (Palerme et al., 2017). Interestingly, by picking NorESM1-M and MIROC-ESM which show some of the more opposite climate change signal on Southern Hemisphere SIE among the CMIP5 ensemble, we cover in our scenario (2.8 to 4.2°C) mostly the lower half of this uncertainty range on Antarctic warming. Bracegirdle et al. (2015) found that about half of the variance of the CMIP5 projection in RCP8.5 scenario for Antarctic temperature and precipitation is explained by historical biases and sea-ice 5 decrease by late 21 st century. Obviously, a non negligible part of the uncertainties on Antarctic climate changes is linked to the representation of general circulation in the atmospheric model (Bracegirdle et al., 2013) and these should be assessed in future work.

Summary and Conclusion
In this study, we present a first general evaluation of the capability of the AGCM ARPEGE to reproduce the atmospheric 10 general circulation and the surface climate of the Antarctic ice-sheet. ARPEGE is able to correctly represent the main features of atmospheric general circulation, although a negative bias in sea-level pressures at mid-latitudes and a positive bias around Antarctica especially in the Amundsen sector is to be reported. Unsurprisingly, the use of observed sea surface conditions (ARP-AMIP simulation) rather than SSC from NorESM1-M and MIROC-ESM helped to improve the representation of sealevel temperatures in the southern latitudes in all seasons but summer. ARPEGE is also able to correctly reproduce surface 15 climate of Antarctica except for large ice-shelves. The differences in T 2m with polar RCM MAR and in-situ observations is encouraging, especially given the large biases that can exhibit other GCMs or even reanalysis when surface climate of Antarctica is considered (Fréville et al., 2014;Bracegirdle and Marshall, 2012). Regarding SMB, our estimates at the continental scale concur with estimates from other studies such as those using polar RCM MAR or RACMO2, even though higher precipitation rates in ARPEGE tend to be compensated for by higher surface snow sublimation rates (+200 Gt yr −1 ). Concerning 20 regional patterns, the distribution of precipitation in ARP-AMIP simulation differs from the one in the MAR RCM, mainly as a consequence of errors in atmospheric general circulation.
Concerning climate change signals, we evaluate the impact of using original and bias-corrected sea surface conditions from MIROC-ESM and NorESM1-M, which display very different changes in winter SIE in their RCP8.5 scenario : respectively -45% and -14% at the end of the 21 st century (2071-2100). Using SSC from NorESM1-M model, we found a T 2m increase 25 of +2.8K and a precipitation increase of about 20%. No significant differences in yearly or seasonal mean T 2m increase, in precipitation or SMB changes were found when using bias-corrected SSC. When using SSC from MIROC-ESM model, the increase in T 2m is around +4K in both cases, but the increase in precipitation is +23% when using directly SSC from MIROC-ESM, while it reaches +37% when using corresponding bias-corrected SSC. This difference is found significant and is to be linked with clearly different dynamical and thermodynamical changes in SLP around Antarctica, mainly in winter and spring 30 when using bias-corrected SSC. At the regional scale, large differences in T 2m and precipitations increase are found when using bias-corrected SSC both from NorESM1-M and MIROC-ESM.
In this study, we have shown the potential of the ARPEGE model for the study of Antarctic climate and climate change. Un- surprisingly, the representation of present climate, especially atmospheric general circulation is improved when using observed SSC. When using SSC from NorESM1-M, we found a 10% higher precipitation accumulation at the Antarctic-continent scale with respect to the bias-corrected reference in both historical and future climate. With respect to the observations, NorESM1-M SST are characterized by a weaker meridional gradient in the Southern Ocean, which decreases the strength of Westerlies around Antarctica and favor the meridional transport of moisture towards the Pole.

5
Concerning climate change signals, we confirm the importance of the choice of the coupled model from which SSC scenario is taken. By performing bias correction of SSC, we showed that not only the regional pattern of temperature and precipitation changes can be different. Indeed, in the case of MIROC-ESM SSC, we found significantly higher precipitation increase and larger increase in winter T 2m when using bias-corrected SSC. These results are another argument in favor of the bias correction of SSC when performing future climate scenarios, as it reduces the uncertainty of the baseline (historical) climate and the need 10 for computational resources as only one historical simulation using observed SSC in needed. However, this method still bears some uncertainties for the study of the climate change in Antarctica, mainly coming from the errors of the atmospheric model ARPEGE. We have seen that the errors on atmospheric general circulation remain substantial even when using observed SSC.
Therefore, in future work, we will assess the uncertainties associated with the errors of the atmospheric model by performing an ARPEGE simulation nudged towards the reanalysis and use the statistics of the model drift in this nudged simulation 15 such as done in Guldberg et al. (2005) to perform an atmosphere bias-corrected ARPEGE historical simulation. Bias-corrected projections such as in Krinner et al. (2018) can then also be assessed using the methods presented here.

26
The We also thank the Scientific Committee on Antarctic Research, SCAR and the British Antarctic Survey for the availability of the MET READER data base.

Appendix A: Sea Surface Conditions
In this section, the historical bias in SSC in MIROC-ESM and NorESM1-M (Fig. A1) used to force ARPEGE model as well as the differences between SSC in rcp8.5 scenarios in these model and their bias-correction (Fig. A2). These first two figures illustrate the efficiency of the bias-correction methods fro SSC as the similarity between differences in futures SST is striking.
For SIC, the patterns of the model bias in historical climates can easily be identified in the differences between original and 15 bias-corrected SSC (Fig. A2), but because there is a decrease of SIE, these patterns are shifted poleward. Yearly and seasonal Southern hemisphere SIE in MIROC-ESM, NorESM1-M and observations (Table A1) and in the two AOGCM original and bias-corrected rcp8.5 scenario (Table A2) are also presented in this supplementary material. Here again, the efficiency of the bias-correction methods to reproduce the climate change signal in hemispheric SIE from the coupled model can be confirmed.
In Figure A3, SST historical bias for both coupled model for each seasons on the whole Southern hemisphere are displayed 20 in order to support the discussion on how the atmospheric model has responded to the same SST biases or perturbations in present and future climate.  Fig. B1.   Change (10 6 km 2 ) -1.6 -0.8 -1.5 -2.3 -1.8 NorESM1-M-rcp85-bc 7.9 3.5 4.2 11.1 12.7 Change (10 6 km 2 ) -1.6 -0. The effect of introducing biased SSC on the modelling of Antarctic T 2m with ARPEGE AGCM is also presented in Fig. B2.
For ARP-NOR-20 (Fig. B2a), the introduction of biased SSC increase the warm bias on the East Antarctic Plateau with respect to MAR and weather stations already present in ARP-AMIP (Fig. 4). The same statement can be made for the winter cold bias over the Peninsula. In summer, there are relatively few differences in the skills of the latter two simulation, which is consistent with similar errors on large-scale atmospheric circulation (Fig. 2).

5
For ARP-MIR-20 (Fig. B2a), the cold bias over the Peninsula is also larger than ARP-AMIP in both seasons. The winter warm bias over the EAP is similar than in ARP-AMIP. In summer, the general tendency of ARP-MIR-20 to be cooler than ARP-AMIP over the continent leads to a decrease of the warm bias with respect to MAR over the margins of the EAIS and WAIS on one hand, but increase the cold bias on the EAP on the other hand, which can be seen in the differences with MAR and weather stations. Appendix D: Atmospheric general circulation

D1 Present climate
In this section, we present and discuss the ability of ARPEGE atmospheric model to represent the broad features of the atmospheric general circulation around Antarctica. The winter (JJA) and summer (DJF) 500 hPa geopotentials and sea-level pressures (SLP) for ERA-I reanalyses and the ARP-AMIP simulation are presented in Fig. D1. In winter, it can be seen than 5 ARPEGE reproduces quite correctly the 3 climatological minimum in SLP and the localization of the maximum of the South Polar vortex above the Ross Sea rather than on the South Pole. However, as already mentioned, the depth of the three SLP minimum and the meridional gradient around 50 to 60°S is underestimated. This remark is also valid in summer. It can also be noted that ARPEGE reproduces relatively correctly the displacement of the third SLP minima (Amundsen Sea Low) from eastern Ross Sea in winter to the Bellingshausen Sea, west of the Peninsula in summer.

D2 Consistency of the atmospheric model response
In this section, we briefly discuss the consistency of the response of the atmospheric model ARPEGE when forced by similar SSC between present and future climate mentioned in the discussion. For the similarity of the SSC bias, see Fig. A1 and Fig. A2. This consistency of the atmospheric model response is considered as being the key for having similar climate signals between climate projections realized with or without bias corrected SSC. In Fig. D2, the difference in SLP between ARP-15 NOR-20 and ARP-AMIP for the four climatological seasons in shown on the upper part, and the corresponding difference for future climate (ARP-NOR-21-ARP-NOR-21-OC) is shown on the lower part. It can be seen that there are few changes in the differences pattern between present and future climate which is to be related with the minor differences in climate changes