Interannual variability of summer surface mass balance and surface melting in the Amundsen sector, West Antarctica

. Understanding the interannual variability of surface mass balance (SMB) and surface melting in Antarctica is key to quantify the signal-to-noise ratio in climate trends, identify opportunities for multi-year climate predictions and assess the ability of climate models to respond to climate variability. Here we simulate summer SMB and surface melting from 1979 to 2017 using the Regional Atmosphere Model (MAR) at 10 km resolution over the drainage basins of the Amundsen Sea glaciers in West Antarctica. Our simulations reproduce the mean present-day climate in terms of near-surface temperature (mean overestimation of 0.10 ◦ C), near-surface wind speed (mean underestimation of 0.42 m s − 1 ), and SMB (relative bias < 20 % over Thwaites

Abstract. Understanding the interannual variability of surface mass balance (SMB) and surface melting in Antarctica is key to quantify the signal-to-noise ratio in climate trends, identify opportunities for multi-year climate predictions and assess the ability of climate models to respond to climate variability. Here we simulate summer SMB and surface melting from 1979 to 2017 using the Regional Atmosphere Model (MAR) at 10 km resolution over the drainage basins of the Amundsen Sea glaciers in West Antarctica. Our simulations reproduce the mean present-day climate in terms of near-surface temperature (mean overestimation of 0.10 • C), near-surface wind speed (mean underestimation of 0.42 m s −1 ), and SMB (relative bias < 20 % over Thwaites glacier). The simulated interannual variability of SMB and melting is also close to observation-based estimates.
For all the Amundsen glacial drainage basins, the interannual variability of summer SMB and surface melting is driven by two distinct mechanisms: high summer SMB tends to occur when the Amundsen Sea Low (ASL) is shifted southward and westward, while high summer melt rates tend to occur when ASL is shallower (i.e. anticyclonic anomaly). Both mechanisms create a northerly flow anomaly that increases moisture convergence and cloud cover over the Amundsen Sea and therefore favors snowfall and downward longwave radiation over the ice sheet. The part of interannual summer SMB variance explained by the ASL longitudinal migrations increases westward and reaches 40 % for Getz. Interannual variation in the ASL relative central pressure is the largest driver of melt rate variability, with 11 % to 21 % of explained variance (increasing westward). While high summer SMB and melt rates are both favored by positive phases of El Niño-Southern Oscillation (ENSO), the Southern Oscillation Index (SOI) only explains 5 % to 16 % of SMB or melt rate interannual variance in our simulations, with moderate statistical significance. However, the part explained by SOI in the previous austral winter is greater, suggesting that at least a part of the ENSO-SMB and ENSO-melt relationships in summer is inherited from the previous austral winter. Possible mechanisms involve sea ice advection from the Ross Sea and intrusions of circumpolar deep water combined with melt-induced ocean overturning circulation in ice shelf cavities. Finally, we do not find any correlation with the Southern Annular Mode (SAM) in summer.

Introduction
From 1992 to 2017, the Antarctic continent has contributed 7.6 ± 3.9 mm to the global mean sea level (Shepherd et al., 2018), and this contribution may increase over the next century (Ritz et al., 2015;DeConto and Pollard, 2016;Edwards et al., 2019). The recent mass loss from the Antarctic ice sheet is dominated by increased ice discharge into the ocean (Shepherd et al., 2018), but both surface mass balance (SMB) and ice discharge may significantly affect the Antarctic contribution to future sea level rise (Asay-Davis et al., 2017;Favier et al., 2017;Pattyn et al., 2018). Despite recent improvements of ice sheet models motivated by newly available Published by Copernicus Publications on behalf of the European Geosciences Union.
satellite products over the last 10-20 years, large uncertainties remain in both the SMB and ice dynamics projections, hampering our ability to accurately predict future sea level rise (Favier et al., 2017;Shepherd and Nowicki, 2017).
The largest ice discharge changes in Antarctica are observed in the Amundsen sector with an increase of 77 % over the last decades (Mouginot et al., 2014). Current changes in the dynamics of glaciers flowing into the Amundsen Sea are dominated by ocean warming rather than changes in surface conditions over the ice sheet (Thoma et al., 2008;Pritchard et al., 2012;Turner et al., 2017;Jenkins et al., 2016Jenkins et al., , 2018. Increased oceanic melting can trigger marine ice sheet instability, leading to increased ice discharge, thinning ice, and retreating grounding lines (Weertman, 1974;Schoof, 2007;Favier et al., 2014;Joughin et al., 2014). In parallel, increased surface air temperature can lead to surface melting, subsequent hydrofracturing and possibly major thinning and retreat of outlet glaciers after the collapse of ice shelves (De-Conto and Pollard, 2016). Surface melting, leading to meltwater ponding, drainage into crevasses and hydrofracturing, is thought to be the main cause of the Larsen ice shelf collapse over the last decades in the Antarctic Peninsula (van den Broeke, 2005;Scambos et al., 2009;Vaughan et al., 2003). While surface melting is currently limited to relatively rare events over the Amundsen Sea ice shelves (Nicolas and Bromwich, 2010;Trusel et al., 2012) and underlying reasons for melt pond formation versus active surface drainage network remain unclear (Bell et al., 2018), the rapid surface air warming observed (Steig et al., 2009;Bromwich et al., 2013) and projected (Bracegirdle et al., 2008) in this region suggests that surface melting could increase in the future. Our study focuses on the two atmospheric-related aspects that can significantly affect the contribution of the Amundsen Sea sector to sea level rise, i.e., snowfall accumulation that is expected to increase in a warmer climate and therefore to reduce the mean sea level and surface melting that could potentially induce more ice discharge and therefore increase the mean sea level.
Understanding the interannual variability of SMB and surface melting is key to (i) quantify the signal-to-noise ratio in climate trends, (ii) identify opportunities for seasonal predictions and (iii) assess the capacity of climate models to respond to global climate variability. Furthermore, years with particularly strong surface (or oceanic) melting could trigger irreversible grounding line retreat without the need for a long-term climate trend. Interannual variability in the Amundsen Sea region is usually described in terms of connections with the El Niño-Southern Oscillation (ENSO), the Southern Annular Mode (SAM) and the Amundsen Sea Low (ASL). Our study revisits these connections through dedicated regional simulations based on the MAR model (Fettweis et al., 2017;Agosta et al., 2019). Hereafter, we start by reviewing recent literature on these climate connections.
The El Niño-Southern Oscillation (ENSO; Philander et al., 1989) is the leading mode of ocean and atmosphere vari-ability in the tropical Pacific. It is the strongest climate fluctuation at the interannual timescale and can bring seasonal to multi-year climate predictability (e.g. Izumo et al., 2010). Global climate models predict an increasing number of extreme El Niño events in the future, with large global impacts (Cai et al., 2014(Cai et al., , 2017. Interannual and decadal variability in the tropical Pacific affects air temperature (Ding et al., 2011), snowfall (Bromwich et al., 2000;Cullather et al., 1996;Genthon and Cosme, 2003), sea ice extent (Pope et al., 2017;Raphael and Hobbs, 2014) and upwelling of circumpolar deep water favoring ice shelf basal melting (Dutrieux et al., 2014;Steig et al., 2012;Thoma et al., 2008) in West Antarctica. Recent studies found concurrences between El Niño events and summer surface melting over West Antarctic ice shelves (Deb et al., 2018;Nicolas et al., 2017;. These connections are generally explained in terms of Rossby wave trains excited by tropical convection during El Niño events and inducing an anticyclonic anomaly over the Amundsen Sea (Ding et al., 2011). Paolo et al. (2018) reported a positive correlation between ENSO and the satellitebased ice shelf surface height in the Amundsen Sea over 1994-2017. Based on a detailed study of the extreme El Niño-La Niña sequence from 1997 to 1999, these authors suggested that El Niño events could increase snow accumulation but also increase ocean melting even more, thus leading to an overall ice shelf mass loss. The impact of ENSO was found to be stronger for the Dotson ice shelf and eastward and weaker for Pine Island and Thwaites (Paolo et al., 2018). However, the aforementioned studies were based on the analysis of a few recent ENSO events and did not account for the highly variable properties of ENSO over multi-decadal periods (e.g. Deser et al., 2012;Newman et al., 2011).
The Southern Annular Mode (SAM; Hartmann and Lo, 1998;Limpasuvan and Hartmann, 1999;Thompson and Wallace, 2000) is the dominant mode of atmospheric variability in the Southern Hemisphere and corresponds to a variation in the strength and position of the circumpolar westerlies. Over the last 3 to 5 decades, the SAM has exhibited a positive trend; i.e., westerly winds have been strengthening and shifting poleward (Chen and Held, 2007;Jones et al., 2016;Marshall, 2003). Medley and Thomas (2019) found similar patterns for the SAM trends and the reconstructed snow accumulation trend over 1801-2000. By contrast, the temperatures above the melting point over the Amundsen ice shelves were found to be largely insensitive to the polarity of the SAM (Deb et al., 2018). The SAM phase has also been suggested to influence the ENSO teleconnection to the south Pacific: in-phase ENSO and SAM events (i.e. El Niño-SAM− or La Niña-SAM+) favor anomalous transient eddy momentum fluxes in the Pacific that make the ENSO teleconnection to the South Pacific stronger than average (Fogt et al., 2011).
The Amundsen Sea Low (ASL; Raphael et al., 2016;Turner et al., 2013a) is a dynamic low-pressure system located in the Pacific sector of the Southern Ocean and moving across the Ross, Amundsen and Bellingshausen seas. The ASL is important regionally and variations in its central pressure and position respectively reflect the second and third leading modes of the Southern Hemisphere climate (Scott et al., 2019, their Fig. 3). A westward shift of the ASL induces northerly flow anomalies over the Amundsen Sea, leading to warmer conditions and increased moisture transport over the ice sheet (Hosking et al., 2013Thomas et al., 2015;Raphael et al., 2016;Fyke et al., 2017). Variations in the ASL central pressure also largely impact the West Antarctic climate: anticyclonic anomalies near 120 • W lead to marine air intrusion over the ice sheet, thereby increasing cloud cover, longwave downward radiations and surface air temperature over the West Antarctic Ice Sheet (WAIS; . While a deepening of the ASL is predicted for the twenty-first century in response to greenhouse gas emissions, its high regional variability makes future changes of the ASL difficult to predict Turner et al., 2009).
Importantly, ENSO and SAM are not independent of each other, and both modes of climate variability impact the ASL (Fogt and Wovrosh, 2015). SAM influences the ASL central pressure since it affects the mean sea level pressure over Antarctica (Turner et al., 2013a). The second and third leading modes of variability in the South Pacific have been suggested to be affected by Rossby wave trains induced by tropical convection anomalies (Mo and Higgins, 1998). In terms of ASL, it corresponds to a migration further west (east) during the La Niña (El Niño), but the difference has a low statistical significance (Turner et al., 2013b).  recently reported that El Niño conditions favored blocking in the Amundsen Sea as well as a negative SAM phase, both leading to warm surface air anomalies in West Antarctica.
In this study we revisit the influence of ENSO, SAM and ASL on summer SMB and melting over the drainage basins of the Amundsen sector in West Antarctica for the 1979-2017 period. While the summer focus on melt rates is obvious, SMB in DJF (i.e. December-January-February) only represents 15 % of the annual SMB. It is nonetheless interesting to analyze the similarities and differences in what drives SMB and melting, and the modes of variability and their teleconnections to the Amundsen Sea region both have strong seasonal characteristics, so that each season needs to be considered separately. To do so, we simulate the surface conditions of the Amundsen Sea region over 1979-2017 using the polar-adapted Regional Atmosphere Model (MAR) forced by the ERA-Interim reanalysis. Section 2 describes the methodology followed in the study and presents the model and observations used for comparison. The model results are analyzed and evaluated against observations in Sect. 3. After evaluating the model skills (Sect. 3.1), we analyze and discuss our results on the potential impact of large-scale climate variabilities on the SMB and melting in Sects. 3.2 and 4. The conclusions are provided in Sect. 5.

Model
To estimate SMB and surface melt over the Amundsen sector we use the Regional Atmosphere Model (MAR; Gallée and Schayes, 1994) and specifically version 3.9.3 (http://mar. cnrs.fr, last access: 25 September 2019). The model solves the primitive equations under the hydrostatic approximation. It solves conservation equations for specific humidity, cloud droplets, raindrops, cloud ice crystals and snow particles (Gallée, 1995;Gallée and Gorodetskaya, 2010). MAR represents coupled interactions between the atmospheric surface boundary layer and the snowpack using the Soil Ice Snow Vegetation Atmosphere Transfer (SISVAT) originally developed by De Ridder and Gallée (1998). The snow-ice part of SISVAT includes submodules for surface albedo, meltwater percolation, and refreezing and snow metamorphism based on an early version of the CROCUS model (Brun et al., 1992). MAR has been largely evaluated in polar regions (e.g. Amory et al., 2015;Gallée et al., 2015;Lang et al., 2015;Fettweis et al., 2017;Kittel et al., 2018;Agosta et al., 2019;Datta et al., 2019).
Our domain includes the drainage basins of the Amundsen Sea Embayment glaciers and a large part of the Amundsen Sea until 65 • S using oblique stereographic projection (EPSG: 3031). It covers an area of 2800 km × 2400 km at 10 km horizontal resolution ( Fig. 1) and 24 vertical sigma levels located from approximately 1 to 15500 m above the ground. We use 30 snow layers, resolving the first 20 m of the snowpack, with a fine vertical resolution at the surface (1 mm) increasing with depth; snow layer thickness varies dynamically depending on the physical properties of overlying snow layer properties. If neighboring layers have similar properties, then layers are associated together. The radiative scheme and cloud properties are the same as in Datta et al. (2019) and the surface scheme including snow density and roughness is the same as in Agosta et al. (2019). The model is forced, over the period 1979-2017, by ERA-Interim reanalysis (Dee et al., 2011), which performs well over Antarctica Huai et al., 2019), at 6-hourly temporal resolution and relaxed over ∼ 50 km laterally (pressure, wind, temperature, specific humidity; the relaxation zone is shown in Fig. 1), at the top (i.e. above 10 km) of the troposphere (temperature, wind) and at the surface (sea ice concentration, sea surface temperature). The Bedmap2 surface elevation dataset is used for the ice sheet topography (Fretwell et al., 2013). The snowpack density and temperature are initialized from the pan-Antarctic simulation from Agosta et al. (2019). Drifting snow is relatively infrequent in the Amundsen region (Lenaerts et al., 2012) Table S1 in the Supplement for station names) located no more than 15 km from the closest continental MAR grid point (even if the domain resolution is 10 km, stations over islands or capes that are not resolved can be located farther than 15 km from the closest continental MAR grid point). For each location, modeled values (surface pressure, near-surface temperature and near-surface wind speed) are computed as the average-distance-weighted value of the four nearest continental grid points. A second selection criterion is also applied in order to reduce comparison errors due to the difference between the model surface elevation and the actual AWS elevation: we only retain observations with an elevation difference lower than 250 m. This two-stage selection leaves 41 suitable AWSs in our domain (Fig. 1).
To evaluate the simulated SMB, we use airborne-radar data from Medley et al. (2013Medley et al. ( , 2014 covering the period 1980-2011. These data were collected through NASA's Operation IceBridge campaign over the Thwaites and Pine Island basins. They are based on the CReSIS radar (Center for Remote Sensing of Ice Sheets), which is an ultra-wideband radar system able to measure the stratigraphy of the upper 20-30 m of the snowpack with a few centimeters in vertical resolution. Airborne-radar data were verified with 190 firn core accumulation records. To evaluate the SMB regional pattern at a broader scale, we also compared the simulated SMB with the observations gathered in the GLACIOCLIM-SAMBA dataset thoroughly described by Favier et al. (2013) and updated by Wang et al. (2016) that are covered by our domain. Similar to Kittel et al. (2018) and Agosta et al. (2019), we selected the observations for which the measurement period extends from 1950 to 2018. Observations before 1979 (i.e., the beginning of our study period) were compared to the average SMB simulated by MAR provided they cover a period of at least 5 years, while observations after 1979 were compared to the SMB modeled by MAR for the observation period. We then compared the modeled SMB computed by using a four-nearest inverse-distance-weighted method for each of the 124 selected SMB observations.
To evaluate simulated surface melt, we use satellitederived estimates of surface meltwater production over 1999-2009 from Trusel et al. (2013), provided at 4.45 km resolution, and based on the QuickSCAT backscatter and calibrated with in situ observations. We also use data from Nicolas et al. (2017), who provide the number of melt days at 25 km resolution over Antarctica. This product is based on passive microwave observations from the Scanning Microwave Multichannel Radiometer (SMMR), the Special Sensor Microwave/Imager (SSM/I) and the Special Sensor Microwave Imager/Sounder (SSMIS) spaceborne sensors and covers the 1978-2017 period. For a given grid cell and a given day, melt is assumed to occur as soon as one of the two daily observations of brightness temperature exceeds a threshold value. As the identification of melt days may be sensitive to the algorithm, we also use the dataset from Picard et al. (2007), extended to 2018 (http:// pp.ige-grenoble.fr/pageperso/picardgh/melting/, last access: 25 September 2019). This dataset is also based on SMMR and SSM/I but uses the algorithms from Torinesi et al. (2003) and Picard and Fily (2006) to retrieve melt days. It is provided as daily melt status at 25 km resolution over the Antarctic continent from 1979 to 2018.

Climate indices
To describe the ENSO, we use the Southern Oscillation Index (SOI) from the Global Climate Observing System (GCOS) Working Group on Surface Pressure (Ropelewski and Jones, 1987; https://www.esrl.noaa.gov/psd/gcos_wgsp/ Timeseries/SOI/, last access: 25 September 2019). The SOI corresponds to the normalized pressure difference between Tahiti and Darwin based on observations. The Rossby wave trains connecting the equatorial Pacific to Antarctica are expected to develop within a few weeks in response to ENSO anomalies (e.g. Hoskins and Karoly, 1981;Mo and Higgins, 1998;Peters and Vargin, 2015), so we first use the synchronous (DJF) SOI in Sect. 3. The lagged relationship to ENSO is discussed in Sect. 4, where we use other 3-month averages of SOI such as JJA (June-July-August). SOI is preferred to NINO3.4 because it gives slightly stronger correlations with the variability in the Amundsen Sea region (as also found by Holland et al., 2019), but very similar results were obtained using NINO3.4 (not shown).
We use the SAM index from NOAA/CPC (https: //stateoftheocean.osmc.noaa.gov/atm/sam.php, last access: 25 September 2019) to describe the primary mode of atmospheric variability in the Southern Ocean (e.g., Marshall, 2003). The SAM index is calculated as the difference of mean zonal pressure between the latitudes of 40 and 65 • S based on NCEP/NCAR reanalysis which produces a SAM that is consistent with other reanalyses after 1979 (Gerber and Martineau, 2018). In the negative (positive) phase, the mean sea level pressure anomaly between the Antarctic and midlatitudes is positive (negative) and leads to a weaker (stronger) polar jet. Thus, positive (negative) values of the SAM index correspond to westerlies that are stronger (weaker) than average over the middle to high latitudes (50-70 • S) and weaker (stronger) westerlies in the midlatitudes (30-50 • S).
The Cryosphere, 14, 229-249, 2020 www.the-cryosphere.net/14/229/2020/ We use two other indices to describe the evolution of the migration and intensity variations in the Amundsen Sea Low (ASL). The datasets are provided by the British Antarctic Survey (https://legacy.bas.ac.uk/ data/absl/ASL-index-Version2-Seasonal-ERA-Interim_ Hosking2016.txt, last access: 25 September 2019) and calculated from the ERA-Interim reanalysis. To describe the migration, we use the longitudinal position of the ASL defined as the position of the minimum pressure within the box 80-60 • S, 170-298 • E , defined in degrees east. A decrease in the longitudinal position index hence corresponds to a westward shift of the ASL. To describe the intensity of the ASL, we use the relative central pressure of the ASL calculated as the minimum pressure in the aforementioned box minus the average pressure over that box . A more intense ASL (deeper depression) is therefore represented by a lower index.
The SAM and ASL indices are defined regionally, and we do not expect any lag with summer SMB, so these indices are therefore calculated as DJF averages. All the correlations are calculated using detrended time series.
The correlations between these four indices are indicated in Table 1. A significant anticorrelation is obtained between the SAM index and ENSO (i.e. −SOI) as previously reported by Fogt et al. (2011). There is no significant relationship between the ASL longitudinal position and ENSO or SAM, as previously reported by Turner et al. (2013a). The relative central pressure also varies independently from SAM, ENSO and the ASL longitudinal position. Numerous previous studies used the absolute rather than relative central pressure to characterize the ASL, but this index is strongly correlated to the SAM index and cannot be considered independently ( Table 1). As proposed by Hosking et al. (2013), the ASL relative central pressure (i.e. actual central pressure minus pressure over the AS sector) allows for a better understanding of West Antarctic climate as it removes the influence of large-scale variability such as ENSO and SAM.

Results
We first evaluate the simulations with regard to observations (Sect. 3.1). Then we analyze the interannual variations in SMB and melting (Sect. 3.2).

Model evaluation
We first evaluate the near-surface temperature and nearsurface wind speed in comparison to AWS data (Fig. 2).
Our MAR configuration reproduces the daily near-surface temperatures, with a mean bias of 0.10 • C and a mean correlation of 0.93 for the whole year and 0.86 for summer months (Fig. 2a). The statistics per station show a root-mean-square error (RMSE) varying from 2.66 (10th percentile) to 4.15 • C (90th percentile) and a mean bias varying from −1.97 to 1.31 • C for the whole year (see Supplement for more details).
The model tends to overestimate the lowest observed wind and underestimate the highest observed wind speeds (regressions in Fig. 2b). The model agreement with observations is nonetheless good on average, with a mean underestimation of 0.42 m s −1 . The statistics per station show a RMSE varying from 1.73 to 3.69 m s −1 and a mean bias varying from −3.08 to 0.85 m s −1 for the whole year. The variance of the wind speed simulated by MAR is lower than observed. Less satisfactory results are generally found for the stations located on an island. This can be explained by the resolution of 10 km, which is still too coarse to resolve small topographic features. For both near-surface temperature and wind speed, the statistics for the summer period (DJF) are very similar to the statistics for the whole year. Our results show very similar model skills compared to other simulations in the same region (Deb et al., 2018;Lenaerts et al., 2017) or at coarser resolution over the whole ice sheet .
www.the-cryosphere.net/14/229/2020/ The Cryosphere, 14, 229-249, 2020 . Scatter plots of observed vs. simulated daily near-surface temperature (a) and daily near-surface wind speed (b) for the selected AWSs (see corresponding locations and names in Fig. 1). The statistics, including RMSE, correlation (R), bias, and standard deviations (σ ), are calculated for individual stations and provided as multi-station mean over the whole year and over the summer months (DJF). The range of RMSE and biases across individual stations is also indicated with the 10th percentile and the 90th percentile of all RMSE values. The lines represent least-mean-square linear fit between simulated data and observations. The complete statistical analyses for individual AWSs are provided in the Supplement (Tables S1-S2). We now assess the simulated SMB compared to the SMB from Medley et al. (2013Medley et al. ( , 2014) derived from airborne radar over the period 1980-2011. The simulated SMB is well captured by MAR with a mean relative overestimation of approximately 10 % over the Thwaites basin and local errors smaller than 20 % at all locations (Fig. 3). The interannual variability is also well simulated by MAR with a correlation of 0.90 (Fig. 4). In order to have a broad overview of the SMB evaluation, we also compared the simulated SMB with the GLACIOCLIM-SAMBA dataset  over the Ross and Siple Coast sector (See Fig. S1 in the Supplement). The bias of simulated SMB compared to observation SMB is less than 10 mm w.e. a −1 and local bias can reach 30 mm w.e. a −1 . However, the relative bias between the GLACIOCLIM-SAMBA dataset and simulated SMB is more pronounced with only 44 % of GLACIOCLIM-SAMBA sites showing a relative error with simulated SMB lower than 20 %. All SMB components are shown in Table 2.
The areas of highest surface melt (> 100 mm w.e. a −1 ) are located near the coast and particularly over Abbot, Cos- The Cryosphere, 14, 229-249, 2020 www.the-cryosphere.net/14/229/2020/ grove and the eastern part of Pine Island ice shelf, while more extreme values (> 200 mm w.e. a −1 ) are found near the peninsula in both simulated and observed datasets (Fig. 5).
Even if the simulated and observed patterns are similar, the simulated surface melt is a factor of 2 lower than observations locally (e.g. over Abbot ice shelf and the peninsula). While the interannual melt rate variability is well reproduced with a correlation of 0.80, the surface melt rate simulated by MAR is underestimated by 18 % on average compared to QuickSCAT estimates (Fig. 6a). Surface melt rate over Pine Island basins is well simulated by MAR (Fig. 6b) with R equal to 0.80 compared to drainage basins with low surface melt (i.e. Crosson, Dotson) where R is equal to 0.14 and 0.24, respectively. This melt underestimation, particularly pronounced over drainage basins with low surface melt rate, could be explained by the slight overestimation of the snowfall accumulation (10 %-20 %), as the presence of a fresh snow layer of high albedo overlying snow or ice layers of lower albedo likely reduces melt. MAR surface melt presents a slight overestimation over Getz ice shelf (Fig. 5) possibly explained by wind advection, foehn effect or even snow metamorphism simulated by MAR. Further work is needed to understand such local biases. MAR is fully driven by low-resolution ERA-Interim sea ice cover and temperature; therefore possible underestimation of the presence of polynyas can also play a role in the melt biases.
We also compare the number of melt days to the satellite products from Nicolas et al. (2017) and Picard et al. (2007). To avoid no-melt-day areas in the time series computation, we use the area where the annual number of melt days for each dataset is more than 3 melt days per year, which corresponds approximately to the ice shelf zone. As with the amount of surface melt, the number of melt days over the domain is underestimated by MAR (Fig. 7). The amplitude of the underestimation is not very sensitive to the melt rate threshold used to define a melt day in MAR. A threshold of 1 mm w.e. d −1 (as in Datta et al., 2019) gives a mean underestimation of 4.8 d per year compared to observation from Nicolas et al. (2017), while a threshold of 3 mm w.e. d −1 (as in Deb et al., 2018;Lenaerts et al., 2017) gives a mean underestimation of 4.9 d per year. This underestimation is less pronounced (0.8 to 0.9 d per year depending on the threshold) when using Picard et al. (2007) as a reference. The interannual variability in the number of melt days is reproduced with correlations of 0.69 and 0.43 to the two satellite products (Fig. 7). Previous study on the Antarctic peninsula also found that MAR melt occurrence is comparable to satellite products, but slightly underestimated over the western coast of the Peninsula (Datta et al., 2019).
Overall, MAR simulates the interannual variability of the Amundsen sector well, and we are now going to use these simulations to investigate the drivers of interannual variability of SMB and surface melting.

Drivers of summer interannual variability
In this subsection, we first investigate the large-scale conditions leading to interannual anomalies in summer SMB or surface melting. For sake of clarity, we only consider the Pine Island and Thwaites basins (together) as a first approach. To identify large-scale conditions leading to high (low) SMB, we calculate composites defined as the average of summers presenting a SMB greater than the 85th (lower than the 15th) interannual percentile, and we proceed similarly for surface melt composites. We choose the 85th and 15th percentiles to optimize the signal-to-noise ratio.
Sea surface pressure composites show that distinct mechanisms affect the interannual variability of summer SMB and surface melting (Fig. 8). Summers with high SMB are on average characterized by a far westward (by ∼ 30 • ) and southward (by 3-4 • ) migration of the ASL center, while the reverse migration is found for summers with low SMB, although with a smaller displacement (∼ 15 • eastward). In contrast, years with high surface melt rates are characterized by a much smaller ASL migration, and no migration is found for years with low surface melt rates, but the pressure gradients differ between the high and low composites. Therefore, we hereafter consider the variability of SMB and surface melting separately.
To further characterize the tropospheric circulation associated with years of low or high summer SMB, we plot composites of both the 500 hPa geopotential height (Fig. 9a, b) and the 500 hPa geopotential height divided by the domainaveraged value for each season (Fig. 9c, d). The latter has the advantage of highlighting changes in regional gradients (related to the regional circulation) rather than largerscale changes in geopotential height. Both provide similar composites, but the statistical significance is higher in Fig. 9c, d. On average, low-SMB summers are characterized by a northward and eastward ASL migration (shown through a dipole in the 500 hPa normalized geopotential composite www.the-cryosphere.net/14/229/2020/ The Cryosphere, 14, 229-249, 2020  in Fig. 9a, c), which is associated with an offshore surface wind anomaly over the glaciers of the Amundsen Sea sector (Fig. 9e). Conversely, high-SMB summers are characterized by a southward and westward ASL migration (Fig. 9b, d), which is associated with an onshore surface wind anomaly over the glaciers of the Amundsen sector (Fig. 9f). The circulation anomalies typical of high-SMB summers favor the southward transport of precipitable water as indicated by the composites of integrated vapor transport (Fig. 10a, b). Increased moisture transport towards the Amundsen Sea Embayment leads to denser cloud cover (Fig. 10c, d) and increased SMB. On average, high-melt summers are also associated with increased moisture transport towards the Amundsen Sea Embayment and conversely for low-melt summers (Fig. 11a, b), but the mechanism is somewhat different from the case of SMB. The ASL migration during high-melt summers is much smaller than for the high-SMB summers (Fig. 8b). As previously done for SMB, we plot composites of both the 500 hPa geopotential height (Fig. 12a, b) and the 500 hPa geopotential height divided by the domain-averaged value for each season (Fig. 12c, d), the latter better highlighting regional circulation changes (geopotential gradients). Summers with high surface melt rates show a significant increase in the 500 hPa geopotential height over the Bellingshausen Sea (Fig. 12b), i.e. an anticyclonic anomaly, and small westward ASL migration as shown in the 500 hPa normalized geopotential composite (Fig. 12d). This anomaly is against the ASL mean circulation and creates a northerly flow anomaly over the ice sheet in the Amundsen sector (Fig. 12e, f). This anticyclonic anomaly was described by  in terms of enhanced blocking activity. As in , we find that high-melt summers are associated with denser cloud cover (Fig. 11c, d) and increased downward longwave radiation (Fig. 11e, f), and therefore surface air warming, while the opposite occurs for low-melt The Cryosphere, 14, 229-249, 2020 www.the-cryosphere.net/14/229/2020/ summers. Composites of sensible heat flux indicate that heat is lost by the snow surface to the atmosphere for high-melt summers, i.e. high melt summers are not caused by foehn events on average (Fig. S2). Now that we have described the mechanisms in play for summers with high and low SMB or surface melt rates, we investigate the connections between the leading modes of climate variability (ENSO, SAM and ASL variability) and summer SMB and surface melting over the individual Amundsen drainage basins (shown in Fig. 1).
In line with the previous composite analysis for highand low-SMB composites, the SMB in all the drainage basins is anticorrelated to the ASL longitudinal position (Table 3, fourth column). This anticorrelation has little statistical significance for Abbot and Cosgrove, but for Dotson and Thwaites the ASL longitudinal position explains nearly 40 % of the SMB interannual variance (explained variance given by square correlations). The ENSO-SMB relationship has moderate levels of statistical significance, with positive SMB correlations to −SOI for all basins but a part of SMB variance explained by ENSO that remains below 16 % (Table 3, second column). −SOI and the ASL longitudinal location are not significantly connected together (Table 1); therefore their connection to SMB can be considered independent from each other. Finally, the SMB is significantly correlated to neither the ASL relative central pressure (Table 3, fifth row) nor the SAM index (Table 3, third column) for all the basins. To better describe interplays, we also calculate a multi-linear regression of SMB on the four indices (last column of Table 3). Accounting for several indices increases the explained SMB variance compared to a single index, indicating an interplay of the ASL and ENSO. Overall, 16 % to 49 % of the summer SMB variance (increasing westward) can be explained by a linear combination of the climate indices.
We now investigate similar relationships, but with surface melt rates instead of SMB. By contrast to SMB, the surface melt connection to the ASL relative central pressure is stronger than its connection to the ASL longitudinal position (Table 4, fourth and fifth columns), which again highlights the two distinct mechanisms explaining high or low melt rates vs. high or low SMB. The part of the melt rate variance explained by the ASL relative central pressure increases westward, from 12 % for Abbot to 21 % for Getz. Even though the effect of the ASL central pressure dominates, there is still a moderate anticorrelation between melt rates and the ASL longitudinal position, suggesting that the mechanism explaining high and low SMB can explain a small part of the melt rate variance (less than 10 %). In a way similar to SMB, SOI explains less than 9 % of the melt rate's variance, with moderate statistical significance (Table 4, second column), and as for summer SMB there is no significant relationship to the SAM. We have repeated the calculations considering the number of melt days, and we find very similar results in terms of correlations (Table 4, second line in each row). Relatively similar conclusions can be drawn from observational estimates of the number of melt days (values in italic in Table 4), except that satellite estimates indicate a stronger correlation to −SOI, even exceeding the correlation to the ASL central pressure in the case for most drainage basins (the variance explained by −SOI reaching 25 %). As the SAM index is significantly anticorrelated to ENSO (Table 1), the stronger melt-SOI correlation in the observational products goes together with a stronger melt-SAM anticorrelation than in our simulations. To better describe interplays, we also calculate a multi-linear regression of melt rates on the four indices (last column of Table 4). Accounting for several indices increases the explained melt rate variance compared to a single index, which indicates an interplay of the fours modes of variability. Overall, 21 % to 30 % of the summer melt rate variance can be explained by a linear combination of the climate indices.
The part of explained variance never exceeds 50 % of the summer melt and SMB variance. Possible reasons for this are as follows. (i) The modes of variability do not explain all the variance locally; for example, the leading EOF of sea surface temperature (SST) in the equatorial Pacific (representing ENSO) only accounts for 50 % to 70 % of the SST variance (e.g. Roundy, 2014), meaning that the tropical convection thought to influence Antarctica is not completely described by SOI or NINO3.4. (ii) Assuming that a large part of the tropospheric circulation variability is explained by ENSO, SAM and ASL indices, there are reasons why the connection may be weaker for SMB and surface melting because of their nonlinear dependence on sea ice and evaporation in coastal regions, the evolution of snow properties, etc. (iii) Strong modulation of the southeast Pacific extratropical circulation by Rossby wave trains is not only due to the existence of El www.the-cryosphere.net/14/229/2020/ The Cryosphere, 14, 229-249, 2020  Niño events but also depends on the exact spatial distribution of deep convection in the tropical central Pacific and the strength of the polar jet (Harangozo, 2004). (iv) A part of the variability of SMB and melting may be stochastic, i.e. not necessarily driven by variability with spatiotemporal coherence at large scales.

Discussion
The composite analysis and the correlation of SMB and melt rates to the ASL indices give a consistent picture. Summers tend to be associated with high SMB when the ASL migrates westward and southward because this places the northerly flow (ASL eastern flank) over the Amundsen Sea, thereby increasing the southward humidity transport and snowfall. This corresponds to the large-scale features described by Hosking et al. (2013) but is here described for the SMB of individual drainage basins. By contrast, longitudinal migrations of the ASL are not the main driver of surface melting variability, as previously noted by Deb et al. (2018). Summers tend to be associated with high surface melt rates when the Amundsen-Bellingshausen region experiences blocking, i.e. anticyclonic conditions, which tends to decrease the climatological southerly flow (western flank of the ASL) and to favor marine air intrusions that make cloud cover denser with increasing downward longwave radiation, as described by . While the role of the ASL now appears to be quite clear, the exact impact of ENSO on SMB and surface melt rates remains elusive. Earlier studies analyzing the impact of ENSO on precipitation in West Antarctica had difficulties understanding the mechanisms and the robustness of the signal, because they had to rely on relatively short observation and reanalysis periods (Bromwich et al., 2000;Cullather et al., 1996;Genthon and Cosme, 2003). Using a dedicated SMB model over a longer time period, we have shown here that the ENSO-SMB relationship in austral summer exists, but it is relatively weak as SOI alone cannot explain more than 16 % of the interannual variance in summer SMB. The relationship between ENSO and the number of melt days was identified by Deb et al. (2018) using both regional simulations and a satellite product. It was then thoroughly described The Cryosphere, 14, 229-249, 2020 www.the-cryosphere.net/14/229/2020/   Table 4. Correlation R between −SOI, SAM, and ASL indices and MAR surface melt rates (bold), MAR number of melt days (regular), number of melt days from satellite products (italic, first value for Nicolas et al. (2017) and second for Picard et al. (2007)), over individual ice shelves in summer. The statistical significance (Welch's t test) is written within brackets. The last column shows the correlation of a multi-linear regression to the four indices using a least absolute shrinkage and selection operator (LASSO, Tibshirani 1996).  Table 3). While we obtain results similar to those of  when using the number of melt days derived from satellite products, both the number of melt days and the melt rates simulated by MAR indicate less variance explained by SOI, that is, between 5 % and 9 % for the individual drainage basins. Our MAR simulations certainly contain biases in the representation of the melting process and the way it affects surface properties such as albedo and roughness, but it is also possible that the number of melt days derived from microwave satellite data is biased due to variability in surface conditions, percolation within fresh snow, meltwater ponding (observed on Pine Island; Kingslake et al., 2017) and satellite overpass time (Tedesco, 2009;. More work will be needed to understand these differences. Numerous publications have explained the remote effects of ENSO on the West Antarctic climate through Rossby wave trains that connect the convective anomalies associated with ENSO in the equatorial Pacific to Antarctica (e.g., Yuan and Martinson, 2001). However, austral winter and spring conditions are more favorable for Rossby wave trains to be formed and to propagate to high southern latitudes than summer conditions (Harangozo, 2004;Lachlan-Cope and Connolley, 2006;Ding et al., 2011, and references therein). The poleward propagation of tropically sourced Rossby waves in summer is indeed inhibited by the strong polar front jet in the South Pacific sector at that time of the year, which leads to Rossby wave reflection away from the Amundsen Sea region (Scott Yiu and Maycock, 2019). This lack of direct
connection in summer was supported by Steig et al. (2012), who found the weakest correlations between NINO3.4 and wind stress anomalies in the Amundsen Sea in DJF compared to other seasons. Therefore, we investigated possible lags in the relationships to ENSO. While ENSO peaks in DJF, it starts to develop in MAM (March-April-May), as indicated by the growing SOI autocorrelation from a 9-to 6-month lag (Fig. 13a). The first implication of this is that any signal correlated to SOI in DJF will be correlated to SOI in the previous JJA without the need for a lagged physical mechanism. Nevertheless, the correlation between SMB or melt rates in DJF and SOI in the preceding JJA is higher than the synchronous correlation for all the drainage basins (solid curves in Fig. 13b-h), which suggests that the lagged relationship is not only a simple statistical artifact. The results of Ding et al. (2011) andSteig et al. (2012) suggest that there could be a lagged mechanism whereby ENSO would influence West Antarctica in austral spring or winter, with a delayed response of SMB and melting in the following austral summer. The number of melt days derived from satellite data also gives 6-month-lagged correlations to SOI that are as high or higher than synchronous correlations for most ice shelves (dashed curves in Fig. 13b-h).
We now discuss possible explanations for this lag. As mentioned previously, the Rossby wave trains connecting the equatorial Pacific to Antarctica are expected to develop Figure 11. (a, b) Vertical integrated vapor transport (IVT) along the y axis (negative toward the continent (kg m −2 , same formula as for Fig. 10), (c, d) cloud cover (no units, from 0 to 1) and (e, f) downward longwave radiation (W m −2 ) anomalies during lowmelt summers (left) and high-melt summers (right). Anomalies are calculated as high or low composites minus the climatology over 1979-2017. The hatched area represents significance > 90 % calculated with Welch's t test.
within a few weeks in response to ENSO convective anomalies (e.g. Hoskins and Karoly, 1981;Mo and Higgins, 1998;Peters and Vargin, 2015). Therefore, the lag has to come from anomalies stored in a slower medium, such as snowpack, ocean or sea ice. Snow surface melting in DJF is correlated neither to the temperature of snow layers within the first 2 m in the previous months (not shown) nor to the snow accumulated over the previous months (not shown). This indicates that heat diffusion in snow or preconditioned porosity or albedo of snow is not responsible for the 6-month lag. By contrast, we find that El Niño events in JJA significantly reduce the sea ice cover in the following DJF (Fig. 14). This is reminiscent of Clem et al. (2017), who found stronger lagged correlation between SON ENSO and DJF sea ice cover than synchronous correlation in DJF, with consequences on summer air temperatures. We suggest two possible explanations for this lagged ENSO-sea ice relationship. First, it could be slowly advected from the Ross Sea. Pope et al. (2017)  ing in MAM created a dipole of sea ice anomalies, with decreased (increased) concentration in the Ross Sea (Amundsen and Bellingshausen seas). Using a novel sea ice budget analysis, they showed that the decreased concentration in the Ross Sea was then advected eastward, reaching the Amundsen Sea in SON and DJF. There is also another possible pathway for a lagged ENSO-sea ice relationship. The zonal wind stress over the Amundsen Sea continental shelf break is a good proxy for the transport of Circumpolar Deep Water (CDW) onto the continental shelf (Thoma et al., 2008;Holland et al., 2019). Steig et al. (2012) noted significant correlations between wind stress and ENSO in JJA and SON but not in DJF. All these studies as well as Paolo et al. (2018) pointed out scales of a few months for the buildup and advection of CDW on the continental shelf and then into the ice shelf cavities where they produce basal melting, and Paolo et al. (2018) reported correlations between ENSO and ice shelf thinning 6 months later. As stronger ice shelf melt rates tend to decrease sea ice in this region due to the entrainment of warm CDW towards the surface Merino et al., 2018), the connection through CDW intrusions may also explain a part of the lag between ENSO and DJF sea ice in the Amundsen Sea. We suggest that both mechanisms (eastward advection of sea ice anomalies and anomalous intrusions of CDW) may explain the 6-month lag between DJF SMB or melting and ENSO, and we leave the details of the ocean-sea ice processes for future research.
Beyond the ASL and ENSO, we also find that the SAM is not significantly related to summer SMB and surface melt over individual drainage basins at interannual timescales, which agrees with Deb et al. (2018). This may appear contradictory to the results obtained by Medley and Thomas (2019), showing that the positive SAM trend from 1957 to 2000 largely explains the pattern of annual SMB trends over the Antarctic ice sheet. First of all, their residual SMB trend (i.e. not related to SAM) is particularly strong in the Amundsen Sea Embayment (their Fig. 1e), highlighting that only a part of the SMB trend in that region may be related to the SAM trend. The multi-decadal SAM trend is also related to ozone depletion and emissions of greenhouse gases, and the interannual SAM variability may have different characteristics and impacts on SMB. Furthermore, the absence of a SMB-SAM relationship in our MAR simulations is specific to the austral summer, which represents 15 % of the annual SMB, and correlations are more significant for the other seasons (Table S3). Therefore, the significant SAM-SMB relationship suggested by Medley and Thomas (2019) for annual SMB is not necessarily contradictory to our results. Lastly, previous studies have suggested that the SAM-ENSO anticorrelation may diminish the impact of ENSO on surface melting and SMB. Partial correlations used to disentangle the SAM and ENSO influences on SMB do indicate a slightly stronger SMB-ENSO correlation when the effect of SAM is removed (in particular for Abbot and Cosgrove; see second and third columns of Table 5), but the effect is relatively small. For melt rates, the SAM modulation is very weak for all the basins (Table 5, fourth and fifth columns).
Lastly, we discuss the relationship between surface melt and snowfall over the ice shelves of the Amundsen sector (last rows of Table 2). According to Table 2, runoff is null over all the ice shelves, which means that the firn is never saturated. In other words, all surface meltwater and rainfall refreeze within the firn. This is consistent with Pfeffer et al. (1991), who estimated that the melt rate needed to saturate the firn with water and lead to hydrofracturing can be estimated as 0.7 times the snowfall rate (both melt and snowfall rates expressed in kilograms per square meter per second or millimeter water equivalent). This indicates that meltwater ponding and complex surface hydrological flows are unlikely to develop over West Antarctic ice shelves under the current climate. To reach saturation at the scale of the entire ice shelf in the future (and therefore to initiate hydrofracturing), the 0.7 ratio of Pfeffer et al. (1991) suggests that melt The Cryosphere, 14, 229-249, 2020 www.the-cryosphere.net/14/229/2020/  rates would need to be multiplied by 2.5 (Cosgrove) to 40 (Crosson) compared to present conditions.

Conclusions
In this paper we have analyzed possible drivers for summer surface melt and SMB interannual variability over the last decades in the Amundsen sector, West Antarctica. For this, we have simulated the 1979 to 2017 period with the Re- gional Atmosphere Model, MAR. We have first evaluated our model configuration in comparison to observational products (i.e. AWS, airborne-radar and firn-core SMB, melt days from satellite microwave, and melt rates from satellite scatterometer). MAR gives good results for near-surface temperatures (mean overestimation of 0.10 • C), near-surface wind speeds (mean underestimation of 0.42 m s −1 ) and SMB (local relative bias < 20 % over the Thwaites basin). The mean surface melt rate over the Amundsen Sea region is underestimated by 18 % compared to the estimates derived from QuickSCAT (Trusel et al., 2013), and the interannual variability of surface melting is relatively well reproduced in terms of melt rate (R = 0.80) or number of melt days (R = 0.43 to 0.69 depending on the satellite product) as also found by previous studies using the same MAR version (i.e. Datta et al., 2019). Similar underestimation was also found in another regional atmospheric model of the Amundsen region (underestimation of 30 %-50 % found by Lenaerts et al., 2017). Overall, www.the-cryosphere.net/14/229/2020/ The Cryosphere, 14, 229-249, 2020 our results indicate that MAR is a suitable tool to study interannual variability in the Amundsen sector. Then, we have analyzed the interannual variability of summer SMB. The strongest summer SMB occurs over Thwaites and Pine Island glaciers when the ASL migrates far westward (by typically 30 • ) and southward (by typically 3-4 • ). This promotes a southward flow on the eastern flank of the ASL, towards the glaciers, with resulting increased moisture convergence, precipitation and therefore SMB. Our study hence provides further support for the connection between Antarctic precipitation and the ASL longitudinal position that was previously described by Hosking et al. (2013) based on the ERA-Interim reanalysis. In terms of climate indices, this corresponds to an anticorrelation between SMB and the ASL longitudinal position. This anticorrelation is found for all the drainage basins of the Amundsen Sea Embayment, and the part of the SMB variance explained by the ASL longitudinal migrations ranges from 2 % to 41 % (increasing westward). A small part of the SMB variance is also related to ENSO, with higher SMB during El Niño events and lower SMB during La Niña, but less than 8 % of the SMB variance is explained by ENSO variability. This SMB connection to ENSO is independent from its connection with the ASL longitudinal position.
We have also analyzed the interannual variability of summer surface melt rates. The strongest surface melting occurs over Thwaites and Pine Island glaciers when the ASL undergoes an anticyclonic anomaly (likely the signature of blocking activity), which is visible through anomalies of the ASL relative central pressure. Such an anomaly promotes a southward anomaly of near-surface winds and moisture convergence over the Amundsen Sea Embayment. As recently described by , this leads to increased cloud cover and downward longwave radiation, which in turn increases surface melting. As for SMB, we do not find that surface melt rate variability in our simulations is strongly connected to ENSO as it does not explain more than 9 % of the total variance in simulated summer surface melt rate (or 12 % of the number of melt days). By contrast and for unknown reasons, the variance in number of melt days derived from satellite products indicates that as much as 25 % of the variance in these products could be explained by −SOI.
We also suggest that at least a part of the ENSO-SMB and ENSO-melt relationships in summer is inherited from the previous austral winter (JJA). Rossby wave trains generated by convective anomalies related to developing El Niño events in austral winter significantly affect the Antarctic region, and we suggest that this has some impact on SMB and surface melting in the Amundsen sector 6 months later. Such a delay could be related either to sea ice anomalies generated by ENSO in the Ross Sea in austral winter and taking 6 months to be advected to the Amundsen Sea (Pope et al., 2017) or to marine intrusions of Circumpolar Deep Water that are favored by El Niño events in austral winter (Steig et al., 2012). Circumpolar Deep Water may take 6 months to reach ice shelf cavities where increased basal melting favors the entrainment of this water towards the ocean surface . It should nonetheless be noted that even accounting for this 6-month lag, the influence of ENSO on summer SMB and melt rates remains weak, not explaining more than 15 % variance.
Lastly, we propose that the rate of surface water needed to saturate the firn and lead to hydrofracturing has to increase by a factor of 2.5 to 40 depending on the ice shelf. Such an increase could be reached under strong warming scenarios given the exponential temperature dependence described by Trusel et al. (2015), although snowfall is also expected to increase (Krinner et al., 2008;Agosta et al., 2013;Ligtenberg et al., 2013;Lenaerts et al., 2016;Palerme et al., 2017), requiring even more meltwater to reach saturation. In their projections, Kuipers Munneke et al. (2014) found that the western part of Abbot as well as Cosgrove could become watersaturated before the end of the twenty-second century, but the other ice shelves of the Amundsen sector remained nonsaturated. Further work will be needed to assess the robustness of these projections, with other firn models and global projections.
Code and data availability. The MAR code (version 3.9.1) is available on the MAR website (http://mar.cnrs.fr/, last access: 17 January 2020); outputs from the Amundsen simulation presented in this study are available on https://doi.org/10.5281/zenodo.2815907.
Author contributions. The study was designed by MD-M and NCJ. Setup of the MAR domain configuration was made by MD-M, CA, AD and NCJ. CA, XF, HG, CK and CA developed and tuned the MAR model for Antarctica, and they contributed to improving and interpreting our simulations. CK developed the scripts used to compare MAR to AWS data. JDW and VF contributed to the interpretation of our results related to interannual variability. All the authors significantly contributed to this paper.
Review statement. This paper was edited by Kenichi Matsuoka and reviewed by two anonymous referees.