Melting over the Northeast Antarctic Peninsula ( 1999-2009 ) : 1 evaluation of a high-resolution regional climate model 2 3

Surface melting over the Antarctic Peninsula (AP) may impact the stability of ice shelves and therefore the 14 rate at which grounded ice is discharged into the ocean. Energy and mass balance models are needed to understand 15 how climatic change and atmospheric circulation variability drive current and future melting. In this study, we evaluate 16 the regional climate model MAR over the AP at a 10 km horizontal spatial resolution between 1999 and 2009, a period 17 when active microwave data from the QuikSCAT mission is available. This is the first time that this model, which has 18 been validated extensively over Greenland, has been applied to the AP at a high resolution and for a relatively long 19 time period (full outputs are available to 2014). We find that melting in the northeastern AP, the focus area of this 20 study, can be initiated both by sporadic westerly föhn flow over the AP mountains and by northerly winds advecting 21 warm air from lower latitudes. A comparison of MAR with satellite and automatic weather station (AWS) data reveals 22 that satellite estimates show greater melt frequency, a larger melt extent, and a quicker expansion to peak melt extent 23 than MAR in the center and east of the Larsen C ice shelf. These differences are reduced in the north and west of the 24 ice shelf, where the comparison with satellite data suggests that MAR is accurately capturing melt produced by warm 25 westerly winds. MAR shows an overall warm bias and a cool bias at temperatures above 0°C as well as fewer warm, 26 strong westerly winds than reported by AWS stations located on the eastern edge of the Larsen C ice shelf, suggesting 27 that the underestimation of melt in this region may be the product of limited eastward flow. At higher resolutions 28 (5km), MAR shows a further increase in wind biases and a decrease in meltwater production. We conclude that non29 hydrostatic models at spatial resolutions better than 5km are needed to better-resolve the effects of föhn winds on the 30 eastern edges of the Larsen C ice shelf. 31

that satellite estimates show greater melt frequency, a larger melt extent, and a quicker expansion to peak melt extent 23 than MAR in the center and east of the Larsen C ice shelf. These differences are reduced in the north and west of the 24 ice shelf, where the comparison with satellite data suggests that MAR is accurately capturing melt produced by warm 25 westerly winds. MAR shows an overall warm bias and a cool bias at temperatures above 0°C as well as fewer warm, 26 strong westerly winds than reported by AWS stations located on the eastern edge of the Larsen C ice shelf, suggesting 27 that the underestimation of melt in this region may be the product of limited eastward flow. At higher resolutions 28 (5km), MAR shows a further increase in wind biases and a decrease in meltwater production. We conclude that non-29 hydrostatic models at spatial resolutions better than 5km are needed to better-resolve the effects of föhn winds on the 30 eastern edges of the Larsen C ice shelf. version of the same model was still unable to resolve high temperatures associated with the initiation of föhn flow 10 during a short period. We note that because these modelling studies use a non-hydrostatic model, they are limited to 11 short periods due to the prohibitive computational cost . 12 Models are limited by the parameterization of physics and our incomplete understanding of the physical 13 processes driving the observed changes. Regional climate models (RCMs) such as the Modèle Atmosphérique 14 Régionale (MAR), evaluated here, can be used for simulating the coupled atmosphere/surface system at a continental 15 and decadal scale (Gallée and Schayes, 1994). The trade-off, in this case, is that RCMs might not be able to capture 16 physical processes with the required accuracy and must be thoroughly evaluated with in-situ and remotely sensed 17 observations. Several studies have used passive microwave estimates for melt occurrence alongside in-situ 18 temperature data (Liu et al., 2003;Ridley, 1993 Melt occurrence over the AP has also been investigated using the QuikSCAT satellite product at a ~2.225km resolution 23 (Long and Hicks, 2010) in combination with model outputs from the RCM RACMO2 (Regional Atmospheric Climate 24 Model) and in-situ temperature estimates ( Barrand et al., 2013). Raw backscatter values from QuikSCAT have also 25 been used to estimates melt flux over the AP (Trusel et al., 2013;Trusel et al., 2012). A recent study using 5.5km 26 horizontal resolution run of RACMO 2.3 over the AP suggested that a further increase in resolution would be required 27 to properly resolve föhn wind propagation, which would imply the removal of the hydrostatic assumption (Van timescales. Additionally, these high-resolution runs can easily be compared to, and potentially nested into, continental-35 scale runs of the same model. 36 We consider two conditions for identifying melting based on previous work comparing MAR outputs 1 (version 3.2) and satellite microwave melt estimates that found that passive microwave estimates were sensitive to a 2 meltwater content of 0.4% (or mm w.e.) in the first meter of the snowpack (Tedesco et al., 2007). The first condition 3 (LWC0.4) determines melt occurrence in MAR when the daily-averaged integrated liquid water content (LWC) in the 4 first meter of the snowpack exceeds 0.4% for at least three consecutive days. The second condition (MF0.4) determines 5 melting when total meltwater production over the day exceeds 0.4 mm w.e., and is intended to capture both sporadic 6 melt (which may refreeze) and melt which has percolated into the snowpack column below 1m, i.e. equivalent satellite-7 based estimates could have potentially shown melt occurrence during some portion of the day. A sensitivity test was 8 conducted with multiple thresholds, finding that the differences between a threshold of 0.1 and 1 mm w.e. (suggested 9 by Franco et al., 2013 as a melt threshold for Greenland) was negligible overall, but more substantial on the northern 10 Larsen C Ice Shelf, where the 4 mm w.e. threshold proved insufficient to capture melt occurrence (Fig. S2)

Microwave satellite estimates of melt extent, duration 24
Spaceborne microwave sensors can detect the presence of liquid water in snow over those regions where poor or no 25 observations and unlike sensors in the visible range, microwave sensors are only weakly affected by the presence of 26 clouds. In the case of active measurements (e.g., radar, scatterometer), the presence of wet snow is associated with a 27 sharp decline in backscatter (s 0 ) (Ashcraft and Long, 2000), whereas in the case of passive microwave data the 28 detection is associated with an increase in brightness temperature (Tb) (Mote et al., 1993;Tedesco et al., 2007). In 29 either passive or active microwave estimates, even the presence of a relatively small amount of liquid water (i.e. a few 30 percent) triggers a substantial increase in the imaginary part of the dielectric constant (Ashcraft and Long, 2006;Ulaby 31 and Stiles, 1980). Threshold-based methods for melt detection from passive microwave data range from a combination of multiple 23 frequencies and polarizations (Abdalati and Steffen, 1995) to using a single frequency, single polarization (e.g., Mote 24 et al. 1993, Tedesco 2009), as is used in this study. Three algorithms are used here which are described in detail in 25 (Tedesco, 2009). These include the 240-algorithm where the threshold was determined as the value above which an 26 increase in liquid water content above 1% no longer produces an increase in Tb, based on output of an electromagnetic 27 model. The original threshold of 245K was found to be insufficiently sensitive and reduced to 240K for this study 28 (Tedesco, 2007) (M. Tedesco, personal communication). The second algorithm uses the winter mean threshold-based 29 method ALA: 30 where snowmelt is assumed to occur when the brightness temperature (Tb) exceeds a threshold brightness temperature 32 (Tc) based on the mean winter (JJA) Tb, the wet snow Tb (Twet_snow, equal to 273K) and a mixing coefficient (a, equal 33 to 0.47). For the ALA algorithm, Ashcraft and Long (2006) presume a wet layer of 4.7 cm and a Liquid Water Content based on the on the winter mean threshold (Twinter) and a threshold value (DT), in this case 30K (Zwally and Fiegles, 1 1994) 2

AWS measurements 4
We evaluate the MAR simulation of the near-surface atmosphere using pressure, temperature and wind speed data 5 collected by three automatic weather stations (AWS) on the AP (Fig 1). The comparison between MAR outputs and 6 AWS data for surface pressure are provided in supplementary data. Data  Metrics are computed for December-January-February (DJF, summer). We did not compute a seasonal average when 14 more than 5 consecutive days of data were missing. The five-day period was chosen as an upper limit for the length 15

Statistical Methods 30
To evaluate and quantify the differences between MAR outputs and AWS data for temperature and wind speed we 31 use a mean bias. Additional statistical measures shown in supplemental data include the coefficient of determination 32 (R 2 ), root mean squared error (RMSE) and mean error (ME) (Wilks, 1995). We assess the extent to which each station located with AWS stations vs all other gridpoints in the full MAR domain (Fig. S7). We ignore all R 2 statistics where 1 the p-value exceeds 0.05. 2 To capture wind speed frequency distributions, we fit available data for each season for MAR (for the full 3 2000-2009 period), AWS (when AWS data are available) and MAR-R (MAR values collected only when AWS data 4 is available) with a Weibull distribution (Wilks, 1995). The shape (b) parameter roughly captures the degree of skew, 5 with higher values being closer to a normal distribution. The scale (l) parameter approximates the peak frequency 6 (we note that this is not equivalent to the arithmetic mean). We report expected values (i.e. first moment or mean) for 7 each windspeed distribution using the best Weibull fit. 8

Results: Melt Occurrence and Meltwater Production 9
In this section, we show results concerning total meltwater production in the AP and compare melt occurrence 10 estimated by MAR with estimates from three passive microwave algorithms as well as QuikSCAT ft3. The relative 11 sensitivity of each melt occurrence criteria, as well as their associated temperature biases, are first compared at the 12 location of the Larsen Ice Shelf AWS. We then identify spatial biases for melt occurrence at the domain scale, finding 13 substantial differences in the center of the Larsen C Ice Shelf as well as to the north and west of the NE basin, a region 14 which includes the former Larsen A and B ice shelves as well as the northernmost portions of the Larsen C ice shelf 15 (Section. 3.2). These differences could result from either weaknesses in the MAR representation of wind dynamics 16 (discussed in Section 4) or from limitations of the satellite sensor or algorithm. Finally, we compare the climatology 17 and inter-annual variability of melt extent (calculated by multiple algorithms) over the CL and NL region (Section. 18

Meltwater production over the AP 20
We show MAR meltwater production over the 1999-2009 period (Fig. 2). The total annual meltwater production 21 estimated by MAR shows substantial inter-annual variation with the NE basin accounting for the highest meltwater 22 production, closely followed by the SW basin (in green). The NE basin is divided into three regions: the NL and CL 23 masks (discussed in Section 3.2) and the remainder of the basin. We note that the SW basin does not covary with the 24 NE basin and the subregions of the NE basin do not consistently covary with one another. The meltwater production 25 shown here does not account for refreezing and we note that the effects of refrozen melt on the snowpack will vary 26 regionally depending on local properties. The NL region dominates meltwater production in the NE basin in most melt production during the study period (only the preceding year is lower). Declining aggregate meltwater production 29 across the AP does not necessarily correspond to declining meltwater production in the most vulnerable regions of the 30 northeastern AP (including the Larsen C Ice Shelf). Because melt in the NL region is particularly sensitive to föhn-31 induced melt, we note that changes in circulation patterns may affect the northwest regions differently than the 32 southern regions. The strong relationship between wind direction and temperature bias points to the need for isolating dominant inter-annual patterns of melt in the Northern Larsen C Ice Shelf and associating them with large-scale 1 atmospheric drivers. 2 A comparison between mean annual meltwater production from 2000-2009 calculated using RACMO2.3p2 3 (5.5 km) vs MAR (10km) is shown in Fig. 3. MAR shows higher meltwater production overall ( Fig. 3b vs 3a), with a 4 difference of over 150 mm w.e. on the Larsen C ice shelf north of 67°S latitude. Over the NE basin, MAR meltwater 5 shows enhanced meltwater production near the AP mountains, including towards the southern edges, and declines 6 eastward and southward. By comparison, meltwater production from RACMO2.3p2 melt declines southward, but no 7 similar west-to-east gradient is apparent. Although inter-annual standard deviations over the northern Larsen C ice 8 shelf are generally above 100 mm w.e. in both models, there are major differences in other regions, with MAR 9 meltwater production exceeding RACMO2.3p2 values by 30 mm w.e.on the southern Larsen C ice shelf as well as 10 the George VI ice shelf ( Fig. 3d vs 3c). Van Wessem et al. (2015a) suggest that even at 5.5 km resolution, the 11 underestimation of the height and slope of the orographic barrier may result in an underestimation of föhn winds as 12 well as precipitation in RACMO2.3p2. We note that in addition to the difference in horizontal model resolution, 13 RACMO2.3p2 contains 40 atmospheric layers while MAR implements 23 layers. While the differences in total 14 meltwater production from RACMO2.3p2 and MAR could be a product of dissimilar physics, the potential effect of 15 model resolution on meltwater production in MAR is specifically discussed in Section 5. While melt occurrence and 16 meltwater production are not related in any linear fashion, we note that the spatial pattern produced by MAR, i.e. the 17 eastward gradient from the edge of the AP, is also shown in observed melt occurrence estimates, most notably from 18 the PMW zwa and QS algorithms ( Fig. 5f,g), as discussed in greater detail in the next section. to PMW ALA while the T metric (AvgT2m > 0°C) compares poorly to satellite-based measures (Fig. 4a). We find 26 that at colder temperatures (when MAXT2m < 0°C), AvgT2m values reported by MAR are substantially higher than 27 those reported by the AWS when only MAR reports melt (Fig. 4b). However, at higher temperatures (where MaxT2m 28 >= 0°C), the AWS reports higher MaxT2m temperatures than MAR and biases are even stronger when only 29 observation-based metrics report melt (Fig. 4e). We note that the Larsen Ice Shelf AWS is located on the eastern edge 30 of the Larsen C ice shelf and the major discrepancies in melt occurrence at this location will be explored further in 31 Section 4, where we further expand the analysis of melt occurrence and temperature biases to include wind direction 32 biases as well. 33 In Fig RACMO2.3p2 reports substantially higher melt occurrence than MAR at the center of the Larsen C ice shelf as well 18 as a comparatively limited west to east gradient. Because overall average annual meltwater production in MAR was 19 shown to be substantially higher, with a stronger west-to-east gradient away from the AP (Fig. 3), we conclude that 20 in comparison to RACMO2.3p2, MAR produces melt less frequently, but with greater intensity. 21 In summary, a comparison between observed and modeled data sources show two distinct spatial patterns for 22 maximum melt occurrence. QuikSCAT ft3 as well as both MAR melt metrics show the highest range of melt days in 23 the northern and western edges of the Larsen C Ice Shelf (including both high and low elevation regions) while PMW 24 algorithms show the highest number of melt days in the center of the Larsen C Ice Shelf, where elevations are lower 25 and topography is less complex. We hypothesize that the major difference in spatial patterns between algorithms/melt 26 metrics is related to the different resolutions of the data sources (~2.2225 km for QuikSCAT, 10km for MAR and 27 25km for PMW), such that QuikSCAT is better able to resolve melt where topography is complex , such as near the 28 spine of the AP. Secondarily, the differences are a product of the depths presumed for the calculation of meltwater 29 content. This is true for both the MAR metrics and for the three PMW algorithms; the "ALA" algorithm, for example, 30 presumes a 4.7cm depth and a 1% liquid water content. (see Section 2). To confirm this, we find the maximum depth 31 to which meltwater percolates (according to MAR) associated with the number of days when melt occurs (according 32 to PMW algorithms). Histograms for total PMW melt days in Fig. S4 show three peaks (two major inflection points) Spatial regions defined as having "low" melt occurrence are highly heterogeneous with regard to elevation, 1 meltwater percolation and the relative sensitivity of PMW algorithms. Low melt occurrence regions largely include 2 the spine of the AP and regions just east of it. Bedmap2 (Fretwell et al., 2013) reports a large range of elevations while 3 MAR reports low coincident meltwater production and a relatively shallow meltwater depth. Both the ALA and ZWA 4 algorithms report melt at higher elevations (above approximately 1300m and 1900m, respectively) than the 240 5 algorithm, which neither reports any melt occurrence above 1100m in the NE basin nor at lower elevations to the north 6 and south. (Table S1, rows 1,4,7 and Fig. S6). Where melt occurrence is low, the 240 and ALA algorithms generally 7 detect melt only where MAR reports a maximum meltwater percolation depth below 0.4 m, (Fig S6a,b), whereas the 8 ZWA algorithm can detect melt at a substantially shallower depth of 0.1 m (Fig S6c). Although generally meltwater 9 in MAR rarely percolates below 3m, in low melt-occurrence regions, modeled meltwater occasionally percolates 10 below 10m in the beginning of the melt season (Fig. S6 a,b,c, column "N", indicating November). We remind the 11 reader that melt occurrence within the firn layer (as calculated by MAR MF0.4) will capture melt that can refreeze 12 immediately, so this does not necessarily correspond to melt which is retained in the snowpack. Rather, the snowpack 13 layer depth represents the deepest layer which is affected by the melt process according to MAR. 14 By contrast, where PMW reports high melt occurrence in the NE basin, MAR consistently reports high 15 coincident meltwater production, low elevations and the deepest average meltwater percolation in the region. In the 16 month of January, we find that where PMW algorithms report melt, coincident MAR meltwater percolates to 2 m into 17 the snowpack for 35-47% of the total day-pixels in the NE basin which report any melt, and as deep as 3 meters for 18 more than 30% of total day-pixels (Table S1, (Table  29 S1, row 11,12). 30 The "NL" (Northern Larsen) mask is defined by finding the mean latitude of the CL region and including all 31 portions of the NE basin above this latitude, but excluding the CL region (Fig. 2, inset). In the NL region, elevation is 32 highly-variable, with a mean value ~600m and MAR and QS detect melt both earlier and more often than for PMW 33 algorithms. The NL region includes the eastern spine of the AP and most inlets (including Cabinet Inlet and SCAR in NDJF), for each year as well as the climatological average. The PMWAll algorithm is typically treated as the most 4 restrictive condition while the PMW zwa and QuikSCAT ft3 are the most sensitive. Melt extent is defined as the total 5 area reporting melt daily between Nov 1 st and February 28 th (austral summer, including November to show early 6 melt) (Fig. 6). 7 The melt extent climatology for PMWAll in the CL region shows the initial increase in sustained melt 8 occurring around December 15 th with melt extent peaking in January, followed by a series of increasingly smaller melt 9 pulses ending with refreezing at the end of February. While MAR shows peak melt extent at the same point in the 10 season, the progression from melt onset is more gradual, average peak melt extent is generally smaller and interannual 11 variability (indicated by the grey envelope) during peak melt extent is larger (Fig. 6c vs Fig. 6a). In the CL region, the 12 PMWAll metric is generally restricted by the low sensitivity of the 240 algorithm. Interannual variability for melt 13 extent is substantial, with PMWAll reporting a larger melt extent than MAR towards the middle of the melt season in 14 most years (Fig. 6b,d) (Fig. 6 g,h). We note that for several years, both QuikSCAT ft3 and PMW ZWA report substantial melt 22 occurrence early in the season (~Nov 15 th ) and that the QuikSCAT ft3 climatology frequently reports melt occurrence 23 in the CL region well after February (Fig. 6g). 24 The NL region includes areas which reported low melt occurrence in all PMW algorithms, variable meltwater 25 percolation depth in MAR was variable , and a large range of elevations was observed (Section 3.2), implying that the 26 mask defined by the combined PMWAll algorithm is less clearly linked to consistent modeled physical properties in 27 this region. Here, the MAR melt extent climatology (Fig. 7a,b) is consistently larger than PMWAll throughout the 28 season (Fig. 7c,d). In comparison to the ZWA (Fig. 6c) and QuikSCAT ft3 (Fig. 7g) (Fig. 7d), but less than the PMW ZWA (Fig. 7f) or QuikSCAT ft3 (Fig. 7h)  algorithms, which are more sensitive. Notably, MAR melt occurrence is comparatively low during the peak melt 3 period. By contrast, in the NL region, MAR reports greater melt occurrence than the most restrictive measure 4 (PMWAll) during peak melt, but far less than the highly-sensitive QuikSCAT ft3 algorithm. The eastern AP is generally substantially colder than the western AP, and temperature-driven melt primarily results 9 from either large-scale advection from lower latitudes or from westerly föhn flow over the spine of the AP (Marshall 10 et al., 2006). Here, we assess the bias in temperature and melt occurrence associated with wind direction at three AWS 11 locations on the Larsen C Ice Shelf (shown in Fig. 1). We first discuss wind direction and wind speed biases during 12 the summer season at all three locations (without regard to melt occurrence) (Section 4.1). For prominent wind 13 direction biases, we quantify the associated temperature and melt occurrence biases in order to capture atmospheric 14 conditions where MAR reports less melt occurrence than observations (Section 4.2). All MAR and satellite data used 15 are co-located to the grid cell associated with the AWS (Fig. 1), and we remind the reader that all three stations, at the 16 eastern edge of the CL region (Fig. 2 inset), are located where MAR reported substantially less melt occurrence than 17 PMW algorithms, QuikSCAT ft3 or AWS temperature-based criteria. 18 Fig. 8 shows wind frequency distributions during the summer season, color-coded for wind direction as represented 20 by the pie graph at the right. We note that AWS data are 3-hourly averages and ERA-Interim are 6-hourly averages 21 for wind speed and direction, while MAR produces daily-averaged outputs. For this reason, a direct comparison 22 between Weibull parameters derived from MAR vs AWS data is not fully justified. The Larsen Ice Shelf AWS has 23 full temporal coverage during the QuikSCAT period while AWS14 and AWS15 were installed after termination of 24 the QuikSCAT mission. These last two stations are used in this study to demonstrate that (a) similar wind biases 25 persisted after the QuikSCAT period at multiple locations, as AWS 14 the Larsen Ice Shelf AWSs are co-located to 26 the same MAR grid cell and that (b) wind biases vary slightly by latitude, AWS15 being located slightly to the south. 27

Aggregate wind direction biases 19
Both MAR and AWSs at all stations show a larger proportion of northerly winds at lower windspeeds (Fig. 8, in  28 yellow and blue), although AWSs report a greater frequency of southwesterly and northwesterly flow ( Table 2 col. 29 4,5 rows 4-9). At the Larsen Ice Shelf AWS location, both AWS and MAR report dominant northeasterly flow (Table  30 2, rows 4,8, col2). However, the Larsen Ice Shelf AWS reports slightly more flow which is either southwesterly 31 (28.9% for AWS vs. 23.2% in MAR) or northwesterly (19.3% for AWS vs. 14.1% in MAR) while MAR reports more 32 southeasterly flow overall (23.5% in MAR vs. 17.4% in AWS). These biases are more pronounced at the southern 33 AWS15, where modelled temperature correlates with a larger portion of the southern Larsen C Ice Shelf than for a smaller proportion of southwesterly flow in the 180°-225° range (especially at the southernmost AWS15 location), 1 although easterly flow is equivalent to AWS-reported estimates. We note that although ERA-Interim has been shown 2 to reproduce the basic structure of föhn flow (Grosvenor et al., 2014), the horizontal spatial resolution may be too 3 coarse to adequately capture southwesterly gap flow here. As discussed further in Section 5, westerly flow towards 4 the stations used in this study may be strongly affected by the fine-scale representation of topography (which is coarse 5 in ERA-Interim) and the lowered orographic barrier due to the smoothing of topography in the northwest in ERA-6 Interim may contribute to the enhanced northwesterly flow reported by ERA-Interim. 7

Wind and temperature biases concurrent with observed melt occurrence 8
When daily-averaged temperature (AvgT2m) values are high, it is more likely that melt is sustained, while high 9 maximum daily temperatures (MaxT2m) can also occur during sporadic melt. Melt occurrence is strongly influenced 10 by the temperature of the snow column as well as at the surface; internal melting can occur even when the surface is 11 frozen due to net outgoing longwave radiation (Holmgren, 1971;Hock, 2005). It is therefore possible for melt to occur 12 despite a cold bias. In general, we find a small, but consistent warm MAR bias for AvgT2m, and a consistent cold 13 MaxT2m bias ( where melt is likely, although melt is still possible due to other components of the energy balance. 17 The cold MaxT2m temperature bias is strongest during northerly flow in general (Table 2,  proportion of northwesterly flow declines (but increases at the AWS). We find that the major flow biases account for 23 a relatively small proportion of melt which is captured by observations but not by MAR. The easterly flow bias 24 accounts for 8%(9%) of days where PMWAll(QS) melt occurrence is not also captured by MAR (Table S9) while the 25 southerly flow bias accounts for 6%(6%) of days when PMW(QS) melt occurrence is not also reported by MAR (Table  26 S8). For these wind direction biases, Fig. 9 presents temperature values when observed sources, either PMW All or 27 QuikSCAT ft3, report melt, but MAR does not. We refer to the condition where PMWAll reports melt (but MAR does 28 not) as "PMWEx" (i.e. PMW exclusive-or), with the equivalent condition for QuikSCAT ft3 called "QSEx". We limit 29 the melt days shown in each figure panel to a specific wind bias, thus showing how the wind bias directly influences 30 temperature-driven melt in both satellite-based observations as well as MAR. Tables S8-S12 contain relative 31 proportions of each case (flow bias) divided for each restriction (i.e. MAR, QSEx or PMWEx), as well as the timeseries 32 mean and biases for AvgT2m, AvgT2m>0°C (excluding days when AvgT2m values from AWS are below 0°C), 33 especially when only satellite-observed melt occurs, are higher. When MAR reports melt, MAR AvgT2m values 1 cluster near 0°C, with a small overall warm bias (Tables S8,S9,

row 4, col 8). Under omission conditions (PMWEx 2
and QSEx), AvgT2m values are lower, and the MAR bias is slightly negative, although the standard deviation is high 3 (Tables S8, S9, row 5,6, col 7). With all flow cases, only QuikSCAT ft3 shows melt at very low observed AvgT2m 4 values. By contrast, AWS MaxT2m values are substantially higher than MAR values (the latter clustering around 0°C) 5 (Fig. 9b,d). Temperature biases associated with southwesterly flow are similar to those shown by the overall bias 6 towards easterly flow in MAR, and are shown in Table S10,S11. 7 Northwesterly winds are most likely to produce föhn-induced melt and we find that on days when MAR 8 reports melt, only 13.2% of winds are northwesterly while AWS reports 25.2% of flow as northwesterly ( Table 2,  9 rows 9,10, col 5). Northwesterly winds show the highest expected windspeeds as well as the highest standard deviation 10 for both MAR and AWS (Table 2, rows 19,20, col 5). While the temperature bias when wind directions are in 11 agreement is relatively minimal, the temperature bias when northwesterly winds are misrepresented is substantial. 12 When MAR reports melt but misrepresents northwesterly winds (this condition accounts for 3% of all MAR melt 13 days), the cool bias for MaxT2m > 0°C is above 4°C (Table S12,  AvgT2m > 0°C (Fig. 9e). The PMWEx and QSEx conditions still report melt at lower temperature values, and the 19 MAR bias remains positive. Although a cold MAR bias persists, MaxT2m values are generally in better agreement at 20 the Larsen IS AWS location during this condition (Fig 9f, Table S12). 21

Discussion and Conclusions 22
We conclude that MAR captures melt which occurs just east of the AP (which is normally the product of westerly 23 föhn flow) with acceptable accuracy according to satellite estimates, but that that melt is underestimated with respect at temperatures below 0°C, when melt is less likely to occur, but which may still impact the refreezing process. 2 However, when maximum daily temperatures (MaxT2m) and average daily temperatures (AvgT2m) exceed 0°C, 3 MAR shows a substantial cold bias. This is particularly evident when MAR misrepresents westerly winds or northerly 4 winds, and the temperature bias is most extreme when northwesterly flow is misrepresented, i.e. the condition when 5 the most intense föhn flow would be likely. However, this represents only a small proportion of the melt occurrence 6 bias, i.e. melt occurrence reported by satellite estimates, but not by MAR. 7 We demonstrate the impact of westerly winds on melt during a single season, specifically during both mid-8 December and the beginning of January of the 2001-2002 season. During both of these periods, satellite-based melt 9 extent in the CL region increases substantially, while MAR melt extent declines after an initial pulse (Fig. S9a). In 10 December, MAR shows an increase in northwesterly flow, both at the station and throughout the region while AWS 11 reports northwesterly winds at slightly higher speeds. Beginning approximately on January 1 st , the NL region reports 12 substantial northwesterly flow, followed by southwesterly flow, although neither is reported at the Larsen Ice Shelf 13 AWS station east of the NL region. Over January, while both AWS and MAR report northeasterly flow, the AWS 14 station also reports substantial high-speed southwesterly flow not captured by MAR. After this period (beginning on 15 approximately Jan. 1 st ), AWS AvgT2m temperatures consistently exceed MAR AvgT2m values until the end of the 16 season (Fig. S11), suggesting that because MAR did not accurately model the initial intrusion of westerly winds, 17 subsequent temperature-induced melt was limited over the eastern Larsen C ice shelf, where this AWS is located. 18 Presuming that the flow characteristics are largely similar in this relatively flat region, we conclude that the 19 underestimation of melt in the CL region is partially due to the absence of westerly flow, but that this flow is adequately 20 captured directly east of the AP (comprising the NL region). 21 Previous work has suggested that southwesterly föhn winds can result from gap flow (Elvidge et al. 2015), 22 although we note that the southwesterly jets studied in this single campaign were typically cooler and moister than 23 surrounding air, i.e. föhn flow produced from isentropic drawdown. While a version with a higher spatial resolution 24 may potentially resolve topography sufficiently to include the initial intrusion of southwesterly gap flow, as well as 25 northwesterly föhn flow, it may also further inhibit subsequent eastward flow when the hydrostatic assumption is 26 retained. While a higher resolution of MAR v3.5.2 (used throughout this study) was not run due to computational 27 constraints, the enhanced computational efficiency of a newer version of the MAR model (MAR v3.9, Section 2.1) 28 could enable higher resolution runs over extended periods in the future. 29 To assess both the potential future application of MAR v3.9 over the AP as well as the effects of both vertical 30 and horizontal resolution on modelled melt estimates, we compare melt occurrence and flow characteristics from Nov 31 1,2004 to March 31, 2005 between multiple versions of the MAR model. This included three versions of v3.9 (Section 32 2.1), with two 5km and 10km resolution versions run with 24 vertical layers as well as an additional 10km resolution 33 version with 32 vertical layers (10km V). The effect of the enhanced horizontal resolution on topography is substantial; 34 the maximum height of the AP in the 5km version of the model is 2567m, but only 2340m in the 10km version. We 35 find that the effect of increasing horizontal resolution to 5km is to limit the consistent strong melt production just 36 leeward of the AP and that an increase in either horizontal resolution or vertical discretization limits eastward flow (Fig. S12). As compared to AWS data at the Larsen IS AWS, all MAR configurations largely replicated the dominant 1 southwesterly and northeasterly flow, although we found an enhanced bias for southeasterly flow with the enhanced-2 resolution versions of the model (Fig. S13). The effects of local topography on wind speed should be relatively limited 3 as the region surrounding the Larsen ice shelf AWS station is relatively flat. Bedmap2 (Fretwell et al., 2013) reports 4 mean (standard deviation) elevation values of 37.38m (0.53m) in the 5km surrounding the station and 37.37m (0.78m) 5 in the 10km surrounding the station. The mean (standard deviation) values for slope are 0.015°(0.018°) at both 6 resolutions. We conclude that a further increase in vertical discretization or horizontal resolution may potentially 7 reduce flow towards the eastern edge of the Larsen C ice shelf, although the effect of better-resolved topography may 8 allow more westerly flow in MAR to cross the AP. 9 As has been suggested by previous studies (Van Wessem et al., 2015a), the implementation of a non-hydrostatic 10 model may improve the representation of westerly föhn flow over the eastern Larsen Ice Shelf (Hubert Gallée, personal 11 communication). We note that previous work has suggested that a 5km non-hydrostatic model was still unable to 12 capture föhn flow on the eastern portion of the Larsen C ice shelf (according to the AWS records), partially due to the 13 inability to simulate southwesterly föhn jets, and that resolutions as high as 1.5km are required to simulate föhn flow 14 accurately (Turton et al., 2017). However, recent work found that spatial resolutions as high as 2km in the non-15 hydrostatic WRF model were still unable to fully-resolve the steep surface temperature increases associated with the 16 beginning of föhn flow (Bozkurt et al., 2018), suggesting that neither increased spatial resolution nor a non-hydrostatic 17 model may be sufficient to fully capture the effects of föhn flow. We conclude from the main analysis that reduced 18 eastward propagation of westerly winds may contribute to a lack of MAR melt in the CL region as compared to 19 satellite estimates but that melt just east of the AP (the NL) region is represented with relative accuracy. This is further 20 confirmed by the similarity between the spatial trends for melt occurrence as compared to QuikSCAT estimates. We intrinsic to the satellite data itself or a product of the algorithm selected for melt detection (Ashcraft and Long, 2006). 31 Products derived from QuikSCAT are limited in temporal resolution because the satellite passes daily, and may 32 therefore ignore sporadic melt occurring at other times of the day. However, previous studies have compared total 33 melt days from the QuikSCAT ft3 algorithm with a measure derived from surface temperature at seven automatic 34 weather stations and shown a positive QuikSCAT ft3 bias compared to AWS (Steiner and Tedesco, 2014). Similarly, In future work, we will extend this model run to the 1982-2017 period as well as explore a higher-resolution run 1 of a newer version of MAR, producing hourly outputs for the near-surface atmosphere. These runs will allow us to 2 examine the frequency of föhn winds, the concurrent meltwater production and the effects of föhn-induced melt on 3 the snowpack. We will use this multi-decadal record to examine interannual trends of föhn winds in all seasons as 4 well as the cumulative effect of a changing regional climate on the snowpack of the NE basin.

Observation-based regions of high melt occurrence (Section 3.2)
CL region high melt at the center-east of the Larsen C ice shelf, melt days exceeding 1 std dev of PMWAll mean melt occurrence NL region high melt in the north and west of the NE basin, consisting of the NE basin above the mean latitude of CL region which excludes the CL region Conditions for melt occurrence (Section 4.2)

PMWEx
PMWAll reports melt occurrence but MAR does not QSEx QuikSCAT ft3 reports melt occurrence but MAR does not MAR-R criteria when MAR data is used only when AWS data is available