Impact of radiation penetration on Antarctic surface melt and subsurface snow temperatures in RACMO2.3p3

This study investigates the sensitivity of modeled surface melt and subsurface heating on the Antarctic ice sheet to a new spectral snow albedo and radiative transfer scheme in the Regional Atmospheric Climate Model (RACMO2), version 2.3p3 (Rp3). We tune Rp3 to observations by performing several sensitivity experiments and assess the impact on temperature and melt by changing one parameter at a time. When fully tuned, Rp3 compares well with in situ and remote sensing observations of surface mass and energy balance, melt, temperature, albedo and snow grain specific surface area. Furthermore, the introduction 5 of subsurface heating in Rp3 significantly improves the snow temperature profile. Near surface snow temperature is especially sensitive to the prescribed fresh snow specific surface area and fresh dry snow metamorphism. These processes, together with the refreezing grain size and subsurface heating, are important for melt around the margins of the Antarctic ice sheet. Moreover, small changes in the albedo and the aforementioned processes can lead to an order of magnitude overestimation of melt, locally leading to runoff and a reduced surface mass balance. 10


Introduction
The contemporary climate of the Antarctic ice sheet (AIS) has been relatively stable, but recently the ice sheet has started losing mass at an accelerated pace (Shepherd et al., 2018). As the AIS contains enough water to raise global mean sea level by 58 m (Fretwell et al., 2013), it is imperative to understand the driving mechanisms behind recent mass loss. Present-day AIS mass loss has been ascribed to the thinning and breakup of ice shelves, the floating extensions of the ice sheet, due to warming of both and ice processes in a dedicated glaciated tile Van Dalum et al., 2020).
Dry snow metamorphism in RACMO2 is calculated using the parameterization of the SNICAR snow model (Gelman Constantin et al., 2020), based on the scheme of Flanner and Zender (2006), which considers the impact of temperature, temperature gradient with depth, layer density and initial grain size distribution on grain growth. Based on Eq. (16) of Flanner and Zender (2006), RACMO2 uses the following expression for dry snow metamorphism in meters per time step: 70 dr dt = dr dt 0 τ τ + 10 6 · (r − αr 0 ) 1/κ · ∆t · 10 −6 3600 . (1) With r the grain radius, r 0 the initial grain radius, dr dt 0 the initial grain growth rate, ∆t the time step, τ and κ empirical parameters and α a newly introduced tuning parameter that will be changed as an experiment. The grain radius is then converted to SSA using SSA = 3 rρice (Grenfell and Warren, 1999), with ρ ice the density of ice, which is set to 917 kg m −3 (Bader, 1964). This parameterization uses three regimes based on the initial SSA following observations of Legagneux et al. (2004): 1) for an 75 SSA of 60 m 2 kg −1 or lower, 2) 60-80 m 2 kg −1 and 3) 80-100 m 2 kg −1 . Snow metamorphism is fastest for the first regime and slowest for the last. In RACMO2, we assume by default a fresh snow SSA of 60 m 2 kg −1 , hence using the first regime, but this will be changed as a sensitivity experiment.
The latest model version, RACMO2.3p3 (Rp3), includes several updates. The spectrally-integrated (broadband) snow albedo scheme of Gardner and Sharp (2010) is replaced by the Two-streAm Radiative TransfEr in Snow model (TARTES, Libois et al. To properly account for subsurface heating, it has to be considered that heat can still reach the surface within a model time step up to the maximum skin layer equilibration depth (SLED). Between the surface and the SLED, the fraction of shortwave radiation absorbed that attributes to the surface energy balance (SEB) decreases linearly from 1 to 0. The remaining energy contributes to subsurface heating. In other words, the SLED is the maximum depth at which some energy can still equilibrate with the surface within a model time step. Beyond the SLED, all absorbed energy leads to subsurface heating (Van Dalum 95 et al., 2021b).
The multilayer firn module of RACMO2 has also been updated. Numerical diffusion is reduced by a new merging routine that limits the mixing of layers with distinct characteristics. Furthermore, the vertical resolution in snow is increased. Rp3 now typically has 50 to 60 layers, with a maximum of 100. Model output, however, is limited to the upper 20 layers. The impact of the aforementioned model updates for the Greenland ice sheet has been investigated extensively by comparing with in situ 100 and remote sensing measurements (Van Dalum et al., 2020, 2021b, which shows improvements compared to the previous RACMO2 version, Rp2.

Surface mass balance and energy budget
The specific surface mass balance (SMB) represents the net mass gain or loss over a glaciated surface. Some surface processes contribute to mass gain, i.e., snowfall (SN) or rain (RA), and others contribute to mass loss, i.e., sublimation (SU), drifting 105 snow erosion (ER) and runoff (RU). RU includes all liquid water not retained or refrozen in the snowpack. In RACMO2, we adopt the following definition, in kg m −2 or mm w.e. yr −1 : Formally this definition of the SMB represents the climatic mass balance (Cogley et al., 2011), as internal accumulation, or refreezing, is included.

110
Melt energy (M ) is modeled in RACMO2 as the residual energy flux of the SEB of a melting snow or ice surface, with all fluxes in W m −2 and defined positive when directed to the surface: with SW d , SW u , LW d and LW u the downward and upward shortwave and longwave radiative fluxes, LHF and SHF the turbulent latent and sensible heat fluxes and G s the subsurface conductive heat flux. Net shortwave and longwave radiative 115 fluxes (SW n and LW n ) are defined as SW d +SW u and LW d +LW u , respectively. In Rp3, some shortwave radiation is allowed to penetrate through the surface, heating layers below. When snow layer temperature is at melting point, the excess energy is modeled as melt. Percolation of meltwater is modeled using the tipping-bucket method (Coléou and Lesaffre, 1998), where layers are filled with water until the irreducible water content is reached. Any excessive water then percolates to the next unsaturated layer where it can refreeze, run off or be retained by capillary forces, all in a single time step.  (2018)). IMAU-FDM provides the snow grain size, water concentration, temperature, layer thickness and snow and ice density for all initial active layers. No impurities are prescribed in the snowpack, as the impurity concentration of the AIS is typically very low (Warren and Clarke, 1990;Doherty et al., 2010;Dang et al., 2015). snow SSA of 60 m 2 kg −1 , no snow metamorphism tuning, i.e., α in Eq. (1) set to 1, and a refreezing grain size of 1 mm. In GRL, we set the SLED to 5 mm after scale analysis (Van Dalum et al., 2021b). In Rp2, the SLED is not defined, as no radiation penetration occurs and all absorbed shortwave radiation contributes to the SEB.
Four more experiments are performed using Rp3, changing one parameter at a time. In the fresh snow grain size (FSG) experiment, the fresh snow SSA is increased from 60 to 100 m 2 kg −1 , reducing r from 55 µm to 37 µm. An SSA of 100 135 m 2 kg −1 better matches observations of fresh snow at Dome C (Libois et al., 2015). Furthermore, this changes the dry snow metamorphism rate from the fastest to the slowest regime, reducing snow growth by an order of magnitude ( Fig. 1). This current parameterization, however, is not optimized for Antarctic conditions. Therefore, in the next experiment we reduce fresh dry snow metamorphism (FSM) even more by setting the tuning parameter α in Eq.
(1) to 0.25. This reduces fresh snow metamorphism considerably, but its impact diminishes with increasing SSA (Fig. 1). As grain size significantly impacts the 140 albedo (Gardner and Sharp, 2010;He and Flanner, 2020), slower snow metamorphism reduces shortwave radiation absorption in the snowpack, hence snow temperatures are expected to decrease. We also reduce the grain size of refrozen snow from 1 to 0.25 mm (RFG), fitting better with Antarctic observations (Domine et al., 2007), which is expected to further reduce melt. The  final experiment is the control run (CON), where the SLED is increased to 10 mm following the scale analysis of Van Dalum et al. (2021b) to better conform to a model time step of 6 minutes. A larger SLED reduces energy available for subsurface 145 heating, further lowering the snowpack temperature.
Running these experiments is computationally demanding, hence only Rp2, GRL and CON are run for the full time period: 1979. FSG, FSM and RFG are run for 1979-1990. For all experiments, 1979-1984 is considered as spin-up. Statistical significance between model versions or observations is determined by using statistical bootstrapping with 2 standard deviation significance.

Observational data
In this study, we use several observational data sets to evaluate the SMB and SEB components, snow and 2-m air temperature, 10-m wind speed and SSA. Here, we provide a brief overview of the observational data sets.

SMB
Modeled SMB is compared with 1870 SMB measurements including isolated observations and traverses on the EAIS (Fig. 2b).
155 Favier et al. (2013) describe this data set in more detail. In addition, melt fluxes are compared with the output of the surface energy balance model (EBM) of Jakobs et al. (2019). This model is forced with high-quality meteorological and radiation observations to specifically produce a melt rate estimate for Neumayer station (Fig. 2b).

Automatic weather stations
The SEB components, 10-m wind speed and 2-m temperature are evaluated using automatic weather station (AWS) data of 160 nine stations, most of them located in DML (Fig. 2b). Some are located on an ice shelf (4, 11) or close to the ice-sheet margin (5, 16) and others are more in-land, hence covering several climatic regimes. All data are monthly averaged. Van Wessem et al.
(2018) provide a more detailed overview of the AWS specifications.

QuikSCAT melt fluxes
The time series of the satellite radar backscatter from the SeaWinds scatterometer aboard QuikSCAT (QSCAT) is used to 165 produce a seasonal meltwater product covering the entire AIS (Trusel et al., 2013). This melt product uses an empirical relation between the satellite product and in situ observations. The QSCAT melt product is provided on a 4.45 km resolution grid, but is resampled to the RACMO2 grid with the nearest neighbor method. Here, we use QSCAT to evaluate the modeled ice sheet wide surface meltwater fluxes between 2000-2009. Table 2. Statistics of the monthly-averaged downward, upward and net longwave and shortwave fluxes during summer (LW d , LWu, LWn, SW d , SWu, SWn, respectively), albedo, sensible heat flux (SHF), latent heat flux (LHF), 2-m temperature (T2m), skin temperature (Tskin) and 10-m wind speed (V10m) using AWS data of DML between 1997 and 2012 (locations shown in Fig. 2b). We use the ratio of the monthly sum of SWu and SW d to determine the albedo. For all variables, 202 observations are available. The correlation coefficient (R 2 ), bias and root-mean-square error (RMSE) are shown for Rp2, GRL and CON. In all following figures, Rp2 is in black, GRL in red and CON in blue.

Rp2
GRL CON  (Brucker et al., 2011). Probes are positioned down to 21 m depth, but we limit the evaluation to the upper 2 m. Temperatures are measured every 10 cm starting between 10 and 60 cm depth, and every 20 cm between 80 and 200 cm.

175
The SSA of the upper snow layers at Dome C are retrieved by Picard et al. (2016) between 2013 and 2015 by using an algorithm applied to observed spectral albedos. This SSA product is representative for the upper two centimeters, as the albedo for such a vertically homogeneous snow layer, with an SSA of 50 m 2 kg −1 or larger, is representative for more than 95% of the observed surface albedo (Fig 1. of Picard et al., 2016). Measurements are available between September and March. 3 Results: Temperature 180 Figure 2 shows the yearly-averaged T 2m difference for GRL and CON with Rp2. Considerably higher temperatures are simulated in GRL, with some areas more than 2.0 • C warmer with respect to Rp2. The temperature in CON (Fig. 2b) is on average only 0.1 to 0.3 • C lower than Rp2. In summer (not shown), the signal of Fig. 2 is amplified. A comparison with observations in DML during summer (Table 2), which is the season where any changes in the albedo have the strongest impact on the SEB, shows that the temperature of Rp2 is modeled well, with a small bias of -0.3 • C and a root-mean-square error (RMSE) of 1.4 • C.

185
The bias of GRL and CON are larger: 2.0 • C and -0.8 • C respectively. This illustrate the high sensitivity to the implemented changes on the T 2m for the AIS in RACMO2. The new radiative transfer scheme results in a lower albedo, which is especially important during summer and will be discussed in more detail in Sect. 5. Including radiation penetration leads to higher subsur-  face snow temperatures, enhancing snow metamorphism and subsequently enhancing radiation absorption. Due to this positive feedback, inaccuracies in the modeled (sub)surface snow metamorphism (Flanner and Zender, 2006) are amplified in Rp3.

190
To investigate the exact cause of deviating temperatures, we show the yearly-averaged T 2m difference with Rp2 for all sensitivity experiments for 1985-1990 (Fig. 3). As Rp2 models T 2m fairly well, differences with this model version cannot be too large and it is used as a benchmark. All implemented changes lower the temperature, although some changes impact it more than others. A significant lowering of the temperature is induced by the increase of the fresh snow SSA to 100 m 2 kg −1 in the FSG experiment (Fig. 3b). FSG also uses a different fresh snow regime in the grain growth parameterization (Sect. 2.1, 195 Fig. 1) and grains with a high SSA consequently remain at the surface for longer.
The strongest temperature lowering occurs when we further reduce the fresh dry snow metamorphism (Fig. 3c) by implementing a tuning parameter (Eq. (1)). As Fig. 1 illustrates, this tuning reduces in particular the snow metamorphism for small grains, i.e., up to 100 times slower metamorphism in FSM than FSG. This tuning makes that surface layers with a high SSA (>50 m 2 kg −1 ) are more persistent between snow deposition events, consequently lowering the surface temperature and 200 hence, through turbulent and longwave exchange between the surface and near-surface atmosphere, reducing T 2m . The significant temperature differences between Fig. 3a and Fig. 3c shows how sensitive RACMO2 is to grain size and underlines the importance of an accurate snow metamorphism scheme. Higher temperatures are relatively persistent on some of the ice shelves (Fig. 3c), especially in DML. These regions are characterized by melt in summer that refreezes in the snowpack. As meltwater refreezes, it increases snow grain size, resulting 205 in more solar radiation absorption and therefore higher temperatures. Reducing the refreezing snow grain size consequently reduces the temperature difference on relatively dry locations with melt (Fig. 3d). Increasing the SLED further lowers the temperature as subsurface heating is reduced (Fig. 3e). The temperature in CON is now somewhat too low during summer (Tab 2). This bias can be further reduced by slightly changing α in Eq. (1).  (Brucker et al., 2011) in CON than in Rp2 and GRL. During summer ( Fig. 4a and b), we observe 215 that Rp2 is somewhat too cold compared to measurements. Results improve for CON, showing the significance of subsurface heating, although the skin temperature in DML is somewhat underestimated (Table 2). Figure 4a and b also show that in GRL, i.e., without tuning, the snow temperatures are significantly overestimated by up to 10 • C. During autumn (Fig. 4c), temperature profiles of Rp2 and CON, and to a lesser extent GRL, are more similar, as the impact of radiation penetration diminishes towards winter. Compared to observations, however, temperatures in autumn are too high. In the previous section, we illustrated the importance of grain size on the temperature of the AIS. Compared to in situ observations at Dome C (Picard et al., 2016), the SSA of the upper two centimeters in the CON simulation follows the yearly cycle well (Fig. 5). The SSA drops gradually over time during spring and summer to values around 40 m 2 kg −1 , which is somewhat higher than observed. In GRL, the SSA is too low as it drops below 20 m 2 kg −1 . The SSA decline during spring 225 is delayed by a few weeks, but the rate of change is similar to observations. After summer, the SSA gradually increases with deposition of fresh snow, but only reaches 40 to 50 m 2 kg −1 for GRL, significantly below observations. For CON, the SSA gradually increases to 80 to 90 m 2 kg −1 , which is in agreement with observations. Note that the average SSA of the upper two centimeters never reaches the prescribed fresh snow SSA of 100 m 2 kg −1 , as large snowfall events at this polar desert site are rare . To summarize, the GRL settings lead to unrealistically low SSAs. The CON settings somewhat 230 underestimate snow metamorphism, leading to higher SSA during summer, but this can be fine tuned using α in Eq. (1). Table 2 shows the statistics of SEB components compared to AWS observations in summer from DML in Rp2, GRL and CON.

Surface energy balance and albedo analysis
All fluxes toward the surface are defined positive.
The longwave radiation of Rp2 and CON correlate well with observations, but some biases are observed. The underestima-235 tion of LW d illustrates that the atmosphere in RACMO2 is too cold. This could be due to too few clouds, too low atmospheric humidity or biases in the radiation scheme for these cold conditions. This is partly compensated by underestimated LW u , resulting in a relatively small LW n bias. In GRL, the bias of LW n is larger, as higher surface temperatures lead to an overestimation of LW u , while only partly compensated by increased LW d . changed, it illustrates that the atmosphere is too transparent, likely due to similar reasons as causing the LW d differences. For Rp2 and CON, this bias is compensated by SW u , as the albedo is somewhat too high during summer. Table 2 also shows that the albedo of GRL is on average too low, which is discussed in more detail in Sect. 5.1, resulting in a lower compensating SW u and consequently a larger SW n bias.
On average, the SHF is overestimated in RACMO2 during summer, despite an underestimation of the wind speed (V 10m , 245 Table 2). This can be either due to an incorrect representation of the roughness length or an incorrect temperature gradient between surface and atmosphere. In GRL, turbulent heat exchange is smaller while T 2m is overestimated. For a stable surface layer, this therefore suggests that the temperature of lower atmospheric layers is too high in RACMO2. Similarly, GRL also shows a better LHF representation than Rp2 and CON. Hence, turbulent fluxes can still be further improved. the slope, the intercept, R 2 the correlation coefficient, the bias and root-mean-square error (RMSE).

250
Year-round monthly-averaged albedo in DML compared to observations is shown in Fig. 6. Figure 6a illustrates that the spread in data points in GRL is similar to CON but with a lower average. Moreover, an albedo lower than 0.8 is sometimes modeled in GRL and is shown by observations, while absent in Rp2 and CON (Fig. 6b).
Yearly averaged, the albedo of CON is relatively homogeneous over the AIS (Fig. 7a) with a high albedo (> 0.8) almost everywhere due to the abundance of fine-grained snow. Compared to Rp2, the differences are generally small, with somewhat 255 higher albedos in West Antarctica (Fig. 7c). The albedo of GRL is significantly smaller than Rp2 (Fig. 7b), showing the impact of snow properties on the radiative transfer scheme in Rp3. The largest differences in both GRL and CON are observed for the Amery ice shelf, where bare ice can be found at the surface during summer. The transition from snow to bare ice is faster due to higher snow temperatures, leading to more snow-free days and consequently a lower mean albedo. Note that the albedo in Rp2 is fixed for bare ice, while TARTES and SNOWBAL are called in Rp3, allowing a variable ice albedo depending on 260 atmospheric conditions (Van Dalum et al., 2020). Figure 8 shows a case study at Neumayer for the one-year period July 2012 to July 2013 at local noon, illustrating the various processes that impact the albedo on seasonal and sub-seasonal time scales. In general the albedo is high (close to 0.9, Fig.   8b) but fluctuating, mostly depending on cloud cover (Fig. 8f). The albedo is on average lower than the broadband albedo February, on the other hand, induces a strong albedo increase, resulting in a large albedo difference of more than 0.1 with PKM ( Fig. 8d). Such differences reduce over time when snow metamorphism occurs or if more fresh snow is deposited. This illustrates that a simple exponential decay function is not enough to properly capture radiation penetration.

Neumayer case study
The impact of cloud cover on irradiance is shown in Fig. 8a. It shows that infrared (IR) radiation is filtered out by clouds, but that cloud content (Fig. 8f) is too small to considerably impact UV and visible irradiance. As the spectral albedo of IR 275 radiation is low (Dang et al., 2015;Warren, 2019), the broadband albedo in Rp3 consequently increases with increasing cloud content. Compared to G&S and PKM, cloud cover induces stronger albedo variations in CON, as this effect is now explicitly taken into account.
Solar zenith angle (SZA) also impacts the albedo. The albedo increases with SZA, as it increases the angle of incidence of radiation, leading to a higher likelihood for light to scatter out of the snowpack (Solomon et al., 1987;Gardner and Sharp, 280 2010). The spectral distribution of light also changes with increasing SZA. For high SZA (>80 • ), a relatively larger part of the irradiance is IR (Fig. 8a), for which the spectral albedo is low, partly compensating the albedo increase. This effect, however, is not captured in G&S and PKM, but is included in Rp3. Consequently, the albedo is lower for CON for high SZA, as can be seen at the beginning of May during clear-sky conditions (Fig. 8c, d).
Compared to observations, the daily mean albedo product of CON is often too high (Fig. 8h), especially during spring and layers to better fit with SSA observations (Fig. 5) and temperature (Fig. 3) does not necessarily lead to a smaller bias in the SEB components or albedo. The analysis of the SEB shows that RACMO2 has some compensating biases, i.e., clouds and turbulence. Nonetheless, despite a slightly overestimated albedo, CON provides a better representation of the near-surface climate, albedo and near-surface snow state than the GRL experiment.

290
6 Surface mass balance Figure 9 shows the mean yearly-accumulated SMB, melt, precipitation and sublimation difference with Rp2 for GRL (upper row) and CON (lower row). In CON, the SMB differences are generally small (lower than 10 mm w.e. yr −1 ), with somewhat and spatial melt variability. This demonstrates the high sensitivity of the implemented changes for this region, as the snowpack is close to the melting point in summer and additional energy absorption therefore leads to a stronger meltwater flux.
In GRL (upper row of Fig. 9), a strong SMB decrease is modeled for ice shelves in the AP, DML and Amery ice shelf.
More inland, the SMB increases somewhat, which is mainly caused by an ice sheet-wide precipitation increase. It is, however, partially compensated by more sublimation. As the precipitation parameterization has not been changed, the moisture of this 305 excess precipitation has been taken up locally. Further analysis showed that it relates to unrealistic features during summer in GRL. Due to the higher surface temperature, sublimation increases and a cloud-topped shallow convective layer is modeled for the interior of the ice sheet. These clouds subsequently provide the additional precipitation. This synoptic weather pattern is, however, not backed by observations. Furthermore, melt has increased strongly around the margins of the entire AIS and all ice shelves. This melt changes the snow structure and leads to extensive runoff on several smaller ice shelves in DML, where the 310 snowpack is close to saturation in summer, and on the Amery, Larsen C, Wilkins and George VI ice shelves. Finally, compared to 1870 SMB observations in the EAIS (locations shown in Fig. 2b), the difference between CON and GRL is small and both agree well with measurements (Fig. A1), with a bias of 15.0 and 15.6 mm w.e. yr −1 , RMSE of 80.64 and 82.27 mm w.e. yr −1 and correlation coefficient of 0.45 and 0.44, respectively.

315
To investigate what causes the strongly overestimated melt in GRL, Fig. 10 shows the melt difference with Rp2 for all sensitivity experiments. By increasing the fresh snow SSA (Fig 10b) and reducing fresh dry snow metamorphism (Fig 10c), less radiation  increased by as much as 490% in GRL with respect to Rp2. Each sensitivity experiment lowers the melt flux, resulting in only a 7.0% increase in CON (Table A1). As a result, the domain-integrated yearly-averaged SMB modeled in GRL is lower (2370 Gt yr −1 ) than CON (2407 Gt yr −1 ).

Comparison with QuikSCAT
In this section we compare modeled melt with QuikSCAT (QSCAT) data (Sect. 2.4.3). QSCAT shows that virtually no melt occurs on the majority of the AIS (Fig. 11) and that there are only small melt fluxes (< 100 mm w.e. yr −1 ) around most of the Integrating melt over the AIS shows a similar pattern (Fig. 12), with melt in GRL almost an order of magnitude larger than QSCAT in every year. This shows the impact of internal heating and melt-albedo feedback, as these processes significantly 6.1.2 Comparison with an EBM Figure 13 shows the cumulative melt at Neumayer station as calculated by the EBM of Jakobs et al. (2019), which is forced by meteorological data, and compares it with Rp2, GRL and CON. Also for this location, GRL predicts excessively high melt.

345
This figure confirms that GRL significantly overestimates melt and that tuning is necessary. CON initially underestimates melt, which is compensated by increased meltwater production in the warm years of 2004, 2010 and 2014, ending closer to the cumulative melt of the EBM than Rp2. In conclusion, melt of the AIS is somewhat sensitive to fresh snow SSA and fresh dry snow metamorphism and is highly sensitive to the refreezing grain size and SLED. Hence, subsurface heating can warm the snowpack considerably, enhancing 350 melt. Despite the low average melt flux in Antarctica, the impact of subsurface heating should not be neglected for a physical description of (sub)surface melt.

Summary and conclusions
In this study, we investigated the impact of a new snow albedo and radiative transfer scheme in the latest version of RACMO2 Analysis of the SEB shows that RACMO2 exhibit some small (max. 10 W m −2 ) persistent biases in the net radiative fluxes, 370 which is caused by too transparent clouds and overestimated turbulent surface fluxes. These biases demonstrate that Rp2 and CON provide a better representation of the surface climate than GRL, and, moreover, that there is room for model development of RACMO2.
The higher (subsurface) temperatures in GRL lead to excessive melt around the margins and on the ice shelves, locally leading to runoff and a reduced SMB. Integrated over the AIS, melt in GRL is one order of magnitude larger than observed by 375 QSCAT and also considerably larger than measured at Neumayer station. In contrast, CON and Rp2 compare well with these observations. Melt is progressively reduced by all sensitivity experiments, especially in RFG and CON, showing the sensitivity of the AIS to the refreezing grain size and SLED. It is clear that GRL does not produce realistic meltwater fluxes and that the standard Greenland settings of Rp3 should not be used for the AIS. This is undesirable, as model settings should preferably not depend on location and/or tuning to local conditions, and shows that more research into this problem is needed.

380
In conclusion, introducing a more physically based albedo scheme in RACMO2 that allows for radiation penetration and subsurface heating improves, after tuning, subsurface snow temperatures in Antarctica. Incorrectly modeling snow conditions can lead to an order of magnitude melt overestimation and can significantly impact the climate and lower the SMB of the AIS.
Furthermore, as is shown in the GRL experiment, only a small lowering of summer albedo by, for example, global warming induced melting would lead to a very different near-surface summer climate in Antarctica.

385
Author contributions. CTvD, WJvdB and MRvdB initiated this study and analyzed the results. CTvD lead the writing of the manuscript, performed the simulations and implemented model changes. All authors contributed to the discussion on the manuscript.
Data availability. Data are available at 27 km resolution for Antarctica for CON and GRL (1979) and FSG, FSM and RFG (1979-1990 Figure A1. Yearly accumulated SMB in the EAIS in CON and GRL compared to observations. The gray line is the 1-on-1 line and the red and blue lines are linear regression of the data, with N the number of observations, the slope, the intercept, R 2 the correlation coefficient, the bias and root-mean-square error (RMSE). The intercept, bias and RMSE are in mm w.e. yr −1 .