Brief communication: Evaluation of the near-surface climate in ERA5 over the Greenland Ice Sheet

, Abstract. The ERA5 reanalysis, recently made available by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a new reanalysis product at a high resolution replacing ERA-Interim, and is considered to provide the best climate reanalysis over Greenland to date. However, so far little is known about the performance of ERA5 over the Greenland Ice Sheet (GrIS). This studyshows ( (cid:58) In (cid:58)(cid:58)(cid:58)(cid:58) this (cid:58)(cid:58)(cid:58)(cid:58)(cid:58) study, (cid:58)(cid:58)(cid:58) we (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) compare (cid:58)(cid:58)(cid:58) the (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) near-surface (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) climate (cid:58)(cid:58)(cid:58)(cid:58)(cid:58) from (cid:58)(cid:58)(cid:58) the (cid:58)(cid:58)(cid:58)(cid:58) new (cid:58)(cid:58)(cid:58)(cid:58)(cid:58) Abstract. The ERA5 reanalysis, recently made available by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a new reanalysis product at a high resolution replacing ERA-Interim, and is considered to provide the best climate reanalysis over Greenland to date. However, so far little is known about the performance of ERA5 over the Greenland Ice Sheet (GrIS). In this study, we compare the near-surface climate from the new ERA5 reanalysis to ERA-Interim, to the Arctic System Reanalysis (ASR) as well as to a state-of-the-art polar regional climate model (MAR). The results show (1) that 5 ERA5 outperforms insigniﬁcantly ERA-Interim in a comparison with near-surface climate observations over GrIS, even if it has to be mentioned that ASR better models the near-surface temperature than both ERA renalyses. (2) Polar regional climate models (e.g. MAR) are still a useful tool to downscale the GrIS climate compared to ERA5, in particular the near-surface temperature in summer that has a key role for representing snow and ice processes such as the surface melt. However, assimilating satellite data and using a more recent radiative scheme enable both ERA and ASR reanalyses to represent more satisfactorily 10 than MAR the downward solar and infrared ﬂuxes. (3) MAR near-surface climate is not affected when forced at its lateral boundaries by either ERA5 or ERA-Interim. Therefore, forcing polar regional climate models with ERA5 starting from 1950 will enable to perform long and homogeneous surface mass balance reconstructions.

and spatial resolution (0.3°) than ERA-Interim (0.75°). This new generation of reanalysis products has been already evaluated over North America as forcing field for a land surface model (Albergel et al., 2018), over Europe (Urraca et al., 2018) and over the Arctic Ocean (Wang et al., 2018) but not yet over the Greenland Ice sheet (GrIS).
Because of the finer resolution of ERA5 (∼31 km over the equator and ∼15 km over Greenland), the question of the relevance of using regional reanalyses (e.g. Arctic System Reanalysis, ASR, Bromwich et al., 2016Bromwich et al., , 2018 or polar-oriented 5 regional climate models (RCMs) (e.g., Fettweis et al., 2017; to study the near-surface climate of the GrIS can be raised. The spatial resolutions are now more similar while the time and spatial evolution of snow pack properties, and the surface energy balance (Rae et al., 2012), remain poorly represented in global reanalyses (e.g., Bougamont et al., 2007;Reijmer et al., 2012;Goelzer et al., 2013;Vaughan et al., 2013;Vernon et al., 2013;van Kampenhout et al., 2018). Moreover, in the context of the substitution of the ERA-Interim reanalysis, it is relevant to assess the new product, ERA5, as a forcing dataset 10 for (regional) climate models or positive degree day models simulating the surface mass balance (SMB), not yet represented in global reanalyses.
The main goals of this study are (1) to evaluate ERA5 against ERA-Interim and ASR reanalyses by comparison with a set of near-surface climate observations covering the GrIS not assimilated in the reanalyses (Ahlstrom et al., 2008), (2) to highlight the added value of using the state-of-the-art RCM MAR (Modèle Atmosphérique Régional, Fettweis et al., 2017) forced by 15 both ERA-Interim and ERA5 to simulate the near-surface climate of the GrIS, and (3) to evaluate the sensitivity of MAR based near-surface climate to the forcing used (ERA-Interim and ERA5 reanalyses) at its lateral boundaries.

The ERA5 reanalysis
The latest generation of ECMWF reanalyses, ERA5 (E5 hereafter, Hersbach and Dick, 2016), has a higher spatial ( :::::::: resolution :: of ∼ 31 km and ::: 0.25 • : (about 15 km at :::: over Greenland) and ::::: higher : temporal (hourly analysis fields and 3-hourly for the ensemble of data assimilation) output-resolution than EI. E5 has replaced EI. E5 is now available from 1979 to near-real time, but is planned to start in 1950. Beside the higher time and spatial resolution, the main improvements compared to EI consist in a 30 higher number of vertical levels (137 versus 60 in EI), an improved 4D-VAR assimilation system, more consistent sea surface condition input products, a globally better balance between precipitation and evaporation, and more (new) data assimilated (ECMWF, 2018).

5
The outputs have a 3-hourly time resolution covering the 2000s (2010 -2016) using version 3.6.0 of the Polar Weather Research Forecast model (Polar WRF,  and the community WRF data assimilation system based on a 3D-Var technique. ASRv2 improves the comparison of near-surface climate variables with observations compared to ASRv1 and EI over the Arctic (Bromwich et al., 2018).
2.2 The model MAR 10 The model MAR is a RCM specifically designed for polar areas (Amory et al., 2015;Lang et al., 2015;Kittel et al., 2018;Agosta et al., 2019) and abundantly evaluated over Greenland (e.g., Fettweis et al., 2011Fettweis et al., , 2017. In this study, we use the last version of MAR (3.9.6). The main improvements compared to the previous MAR version used in Delhasse et al. (2018)   (2) data records clearly subject to instrument malfunction. The AWSs excluded and the reasons for their exclusion are listed in Table S2 of the supplementary material. The robustness of the observed temperature time series has been improved with a selection criterion excluding measurements when the ventilation of the station is not active. An unventilated temperature can be significantly warm biased by solar radiation and thus cannot be considered reliable.

Comparison method 5
When we stated our study, only the 2010 -2017 period was available for E5, while ASRv2 is available until 2016. We have therefore limited the comparison to the time period spanning 2010 -2016.
Here we assess the near-surface climate of the GrIS simulated by E5 against PROMICE observations at a daily time scale.
We also compare it to the previous reanalysis generation, EI, the regional reanalysis ASR and two MAR simulations. Four variables are evaluated here as proxy of the near-surface climate: 2-m temperature (T2M), 10-m wind speed (W10M), short-10 wave downward radiative flux (SWD) and long-wave downward radiative flux (LWD).
Modelled values of these essential climate parameters are computed for each AWS location following an average-distanceweighted values of the four nearest grid points. To evaluate modelled values, we compare the correlation, the root mean square error (RMSE), the centred RMSE (RMSEc, Eq. S1 in supplementary materials) and the mean bias (MB) between daily observations and each modelled datasets. These statistics are calculated for each day of AWS observations, averaged over 2010 -2016 and for all AWSs, by applying a weighted average according to the number of available observations for each station.

5
For T2M statistics, we initially tried to correct modelled temperature values from the altitude difference between the station and the model interpolated elevation with a time variable vertical temperature gradient. As the comparisons did not improve, we concluded that applying such a correction would add more uncertainties than using the raw modelled fields without elevation correction.

10
The comparison of daily observations and model values for the four main variables is summarized in The added value of MAR is recognisable in the yearly absolute values of MB that are smaller. In summer the temperature biases in both MAR simulations are the highest, but the same simulations show the lowest RMSE(c) and highest correlation with observations (0.87). Both ERA reanalyses perform worse than MAR, while ASR shows similar statistics as MAR.
MAR E5 is colder in summer than MAR EI , but the two simulations produce similar temporal variability.
Two explanations can be given for the statistical differences in T2M between the models. First, a difference in altitude may 25 exist between the AWS and the corresponding interpolated model elevation, which mainly influences the annual MB. For example, the interpolated elevation of the EI grid is 770 m higher at AWS QAS_L (see Table S1 in supplementary materials) while the difference in altitude is lower for the other models (151 m for E5, 6 m for ASR and 119 m for MAR). This difference leads to a negative MB of EI (-4.89 • C, Table S15) and erroneously suggests that this model is colder at this location. The second difference in the modelling of T2M is the better representation by the two regional models (MAR and ASR) of the physical 30 processes at the surface of the GrIS. This consequently results in a better representation of surface-atmosphere interactions, which are influenced by the melt of the snow pack when the excess energy is used to melt snow or ice and not to warm the surrounding air, and by the density of the snow pack, which is better modelled in the polar RCMs. Resolving surface processes (i.e. melt-albedo feedback) that are driving the near-surface temperature and melt variability is particularly relevant in summer when the statistics of both ERA datasets are worse than those of RCMs.
The finer resolution of the RCMs and the inherently better representation of the topography could also play an important To conclude, MAR shows the best accuracy when modelling T2M which might also lead to a better representation of the 15 surface energy balance and melt (not evaluated here).

Wind speed
W10M in each model is well correlated with observations (Table 1, annually > 0.79 and in summer > 0.73) and has an insignificant RMSE representing 70 % of the daily variability (taken as the standard deviation), except for ASR in summer where the RMSE is higher than the standard deviation. 20 Wind speed depends on synoptic atmospheric features, but also on interactions with the surface and local topographic conditions. These generate persistent and widespread katabatic winds, and winds being channelled through valley in mountainous coastal areas of Greenland. It is difficult for models to correctly represent the surface wind regime in mountainous areas due to their resolution exceeding the topographical length scales.
E5 correlate better with in-situ observations than EI, MAR and ASR at the annual and summer time scales, and also has a 25 smaller RMSE and RMSEc.
Despite the improved representation of W10m in E5, both EI and E5 underestimate W10M (negative bias between -1.06 and -1.04 ms −1 ) as also shown by Moore et al. (2016) over Greenland and Jones et al. (2016) over Antarctica. Nevertheless not all PROMICE AWSs are located near the relatively steep ice sheet margin where mountains may or may not be present, disallowing the models to well capture the katabatic winds in the shallow atmospheric boundary layer. The models can also be 30 not able to reproduce the near-surface temperature deficit, and/or they have too few levels near the ice sheet surface. It should be noted that the underestimation of wind speed would be even larger at the height (∼3 m) at which wind is measured by the PROMICE AWSs.
6 W10M in ASR and in both MAR simulations (UV1 in Table 1) is overestimated with a positive bias higher than 1.3 ms −1 .
The biases are reduced for MAR wind speed at ∼2 m (UV2 in Table 1), which is more similar to the height of the AWS measurements. However, the correlation of the wind speed is neither sensitive to the vertical level used in MAR (2 m vs 10 m), nor to EI versus E5 forcing.

Longwave downward radiative flux 5
Contrary to the near-surface wind speed and temperature observations that are usually assimilated in reanalyses, observed downward radiative fluxes are usually not. Forecasted radiative fluxes simulated by the three reanalysis models have been compared to in-situ observation of radiative fluxes. Table 1 shows that each model has a satisfactory representation of LWD. Differences with PROMICE observations are small, with all the models underestimating LWD by 10 -16 Wm −2 .

10
E5 performs the best for LWD producing the highest correlation coefficients (0.94 annually, 0.89 in summer) and the smallest RMSE.
The two MAR simulations are similar, but the reanalyses show more favourable comparisons. The temporal variability of LWD is better represented by the reanalyses, yet the yearly MB are smaller for MAR EI (-11.35 Wm −2 ) and MAR E5 (-10.58 Wm −2 ). 15 The better LWD statistics of the three reanalyses compared to MAR EI and MAR E5 is likely related to their assimilation of radiance from satellite data, as well as the assimilation of sparse coastal temperature and humidity profiles from radiosondes (Dee et al., 2011;Bromwich et al., 2016). This enables a better representation of incident radiative fluxes that depends on clouds and their microphysical characteristics, including the thickness, water phase and temperature. MAR does not assimilate such observations, but is only forced at its lateral boundaries every 6-hours (specific humidity and temperature). Clouds in 20 MAR are the outcome of the own climate and microphysics of the model. Table 1 shows that each model performs well at representing SWD (yearly correlation >= 0.97 and summer correlation >= 0.88) and differences with PROMICE observations are not significant.

Shortwave downward radiative flux
Similar to the LWD statistics, the reanalyses represent SWD better than the RCMs, with E5 providing the best statistics.

25
The ASR reanalysis overestimates SWD (yearly MB = 7 Wm −2 and summer MB = 23 Wm −2 ) whereas other models underestimate SWD (MB = -4 Wm −2 on average), as also highlighted by Bromwich et al. (2018). Large LWD and SWD biases in ASR indicate that additional model improvements in Polar WRF are necessary to better capture the radiative cloud effects despite improved model cloud physics between ASRv1 and ASRv2 (Bromwich et al., 2018).
The assessment of SWD as represented by both MAR experiments reveals no significant difference, but a less accurate SWD 30 temporal variability than in the ERA reanalyses.
In general, the accurate model representation of incident radiative fluxes (LWD and SWD) depends on radiative scheme of model. The scheme used by MAR is the one from ERA-40 (the previous ECMWF reanalysis before EI) which has been updated for the EI and E5 reanalyses. This, combined with the fact that the reanalyses assimilate observations within the RCM domains, enables them to simulate clouds better, explaining the higher accuracy of radiative fluxes simulated by the ERA reanalyses compared to MAR forced by these same reanalyses.

Additional analysis
The same statistical comparison with GC-Net  observations was performed to better cover the Green-5 land ice sheet, as GC-Net stations are mainly located in the accumulation area. However, it is important to note that GC-Net observations are assimilated into reanalyses (EI, E5 and ASR) but not into MAR. Therefore, the comparison of models with GC-Net observations was carried out separately from PROMICE observations in order to keep the independence of the PROMICE comparison with data assimilation. The conclusions of this comparison are identical to the results presented above, except that the assimilation of this data set into reanalyses favours the reanalyses for the representation of T2M with respect to 10 MAR. A more detailed analysis of the results can be found in the supplementary materials (see Table S5).

Discussion and conclusions
We have evaluated essential near-surface climate variables (2-m temperature, 10-m wind speed and energy downward fluxes) simulated by the new ERA5 reanalysis against EI, ASR, and MAR forced by EI and by E5 for the period 2010 -2016.
EI is usually used as a reference over Greenland while ASR is a regional reanalysis specifically developed for the Arctic 15 region. E5 outperforms EI for most variables, but not significantly. ASR is able to model temperature more accurately than the other global reanalyses. Near-surface wind speed is underestimated by both ERA reanalyses.
MAR performs less satisfactorily than the reanalyses in terms of downward solar and infrared fluxes likely because of its older radiative scheme, and because it does not assimilate satellite data within its domain. Still, near-surface temperature, especially in summer, is calculated with more accuracy by MAR, suggesting that MAR does well in resolving processes in 20 the shallow atmospheric boundary layer over the Greenland ice sheet. A good representation of T2M is important because of its importance to snow and ice processes such as the surface melt. In order to better simulate SMB, there is still an interest of using polar RCMs like MAR, not constrained by observations, to represent the near-surface climate over Greenland in the ablation zone compared to E5.
We also evaluated the sensitivity of MAR to its lateral forcing, using both E5 and EI. For each analysed variable, results 25 from both MAR simulations are highly similar, except that MAR E5 generates slightly lower near-surface temperatures than MAR EI , illustrating the ability of regional climate models to simulate climate in detail when forced by reanalyses.
Since September 2019, E5 has replaced EI, and it covers a long and homogeneous period (planned from 1950 to present).
This represents a significant advantage compared to the discontinuity between ERA-40 and ERA-Interim in 1979, which can be of consequence to SMB reconstructions (e.g. Fettweis et al., 2017). In this study we showed that E5 is superior in simulating 30 the near-surface climate of the GrIS over EI, while the advantage is not large. However, when reconstructing SMB back in time to 1950, using E5 as forcing has clear advantages in terms of continuity.
contributed to the editing of the manuscript.
Competing interests. The authors declare that they have no conflict of interest.

Introduction
Reanalyses are global datasets describing the recent history and current state of the atmosphere, land surface, and oceans. They 15 merge sparse observations into a space-and time-continuous product over the whole Earth. These datasets, commonly used in geophysical sciences, enable for instance to evaluate recent climate trends (e.g., Belleflamme et al., 2015;Hanna et al., 2018) and to constrain numerical climate models at their boundaries (e.g., Fettweis et al., 2017;. The ERA5 reanalysis (Hersbach and Dick, 2016), recently made available by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a new reanalysis product that has replaced ERA-Interim since the 1st September 2019, con-20 sidered until now as the best reanalysis over Greenland (Chen et al., 2011;Jakobson et al., 2012;Lindsay et al., 2014;Fettweis et al., 2017). In addition to the model improvements listed in Hersbach and Dick (2016), ERA5 is available at higher vertical and spatial resolution (0.3°) than ERA-Interim (0.75°). This new generation of reanalysis products has been already evaluated 1 over North America as forcing field for a land surface model (Albergel et al., 2018), over Europe (Urraca et al., 2018) and over the Arctic Ocean (Wang et al., 2018) but not yet over the Greenland Ice sheet (GrIS).
Because of the finer resolution of ERA5 (∼31 km over the equator and ∼15 km over Greenland), the question of the relevance of using regional reanalyses (e.g. Arctic System Reanalysis, ASR, Bromwich et al., 2016Bromwich et al., , 2018 or polar-oriented regional climate models (RCMs) (e.g., Fettweis et al., 2017; to study the near-surface climate of the GrIS can 5 be raised. The spatial resolutions are now more similar while the time and spatial evolution of snow pack properties, and the surface energy balance (Rae et al., 2012), remain poorly represented in global reanalyses (e.g., Bougamont et al., 2007;Reijmer et al., 2012;Goelzer et al., 2013;Vaughan et al., 2013;Vernon et al., 2013;van Kampenhout et al., 2018). Moreover, in the context of the substitution of the ERA-Interim reanalysis, it is relevant to assess the new product, ERA5, as a forcing dataset for (regional) climate models or positive degree day models simulating the surface mass balance (SMB) , not yet represented in 10 global reanalyses.
The main goals of this study are (1) to evaluate ERA5 against ERA-Interim and ASR reanalyses by comparison with a set of near-surface climate observations covering the GrIS not assimilated in the reanalyses (Ahlstrom et al., 2008), (2)  The fourth-generation reanalysis from the ECMWF (ERA-Interim, Dee et al., 2011), available at a spatial resolution of ∼ 20 0.75 • (about 41 km over Greenland) and a time resolution of 6-hourly for analysis fields, has been widely used over the Arctic (e.g., Kapsch et al., 2014;Bieniek et al., 2016) and especially over Greenland (e.g., Lucas-Picher et al., 2012;Bennartz et al., 2013;Merz et al., 2013;Cox et al., 2014). The ERA-Interim reanalysis (EI hereafter) is considered as the reference in this study.

The ERA5 reanalysis 25
The latest generation of ECMWF reanalyses, ERA5 (E5 hereafter, Hersbach and Dick, 2016), has a higher spatial resolution of ∼ 0.25 • (about 15 km over Greenland) and higher temporal (hourly analysis fields and 3-hourly for the ensemble of data assimilation) output-resolution than EI. E5 has replaced EI. E5 is now available from 1979 to near-real time, but is planned to start in 1950. Beside the higher time and spatial resolution, the main improvements compared to EI consist in a higher number of vertical levels (137 versus 60 in EI), an improved 4D-VAR assimilation system, more consistent sea surface condition input 30 products, a globally better balance between precipitation and evaporation, and more (new) data assimilated (ECMWF, 2018). 2

The Arctic System Reanalysis
ASR is a regional reanalysis product for the Arctic region . ASR version 2 (called ASR hereafter, Bromwich et al., 2018) has a horizontal resolution (15 km)    and the community WRF data assimilation system based on a 3D-Var technique.

5
ASRv2 improves the comparison of near-surface climate variables with observations compared to ASRv1 and EI over the Arctic (Bromwich et al., 2018).

The model MAR
The model MAR is a RCM specifically designed for polar areas (Amory et al., 2015;Lang et al., 2015;Kittel et al., 2018;Agosta et al., 2019) and abundantly evaluated over Greenland (e.g., Fettweis et al., 2011Fettweis et al., , 2017. In this study, we use the last neither assimilated in the reanalyses nor in MAR so that model output is truly independent from the observations.

Automatic weather stations
Among the time series from the 25 AWSs available in the PROMICE dataset, we dismissed the ones established after the end of our study period (2010 -2016). The remaining 21 AWSs (Figure 1) are mentioned in supplementary materials (Table S1), 25 also listing differences in elevation at the AWS sites between model and reality. For each of the model variables of interest (pressure, 2-m temperature, 10-m wind speed, short-wave and long-wave downward radiative fluxes), we excluded the AWS when: (1) a too large difference in elevation between the station and the corresponding grid cells of all models (> ∼ 250 m), and (2) data records clearly subject to instrument malfunction. The AWSs excluded and the reasons for their exclusion are listed in Table S2 of the supplementary material. The robustness of the observed temperature time series has been improved with a selection criterion excluding measurements when the ventilation of the station is not active. An unventilated temperature can be significantly warm biased by solar radiation and thus cannot be considered reliable.

Comparison method
When we stated our study, only the 2010 -2017 period was available for E5, while ASRv2 is available until 2016. We have therefore limited the comparison to the time period spanning 2010 -2016.

5
Here we assess the near-surface climate of the GrIS simulated by E5 against PROMICE observations at a daily time scale.
We also compare it to the previous reanalysis generation, EI, the regional reanalysis ASR and two MAR simulations. Four variables are evaluated here as proxy of the near-surface climate: 2-m temperature (T2M), 10-m wind speed (W10M), shortwave downward radiative flux (SWD) and long-wave downward radiative flux (LWD).
Modelled values of these essential climate parameters are computed for each AWS location following an average-distance- 10 weighted values of the four nearest grid points. To evaluate modelled values, we compare the correlation, the root mean 4 square error (RMSE), the centred RMSE (RMSEc, Eq. S1 in supplementary materials) and the mean bias (MB) between daily observations and each modelled datasets. These statistics are calculated for each day of AWS observations, averaged over 2010 -2016 and for all AWSs, by applying a weighted average according to the number of available observations for each station.
For T2M statistics, we initially tried to correct modelled temperature values from the altitude difference between the station and the model interpolated elevation with a time variable vertical temperature gradient. As the comparisons did not improve, we 5 concluded that applying such a correction would add more uncertainties than using the raw modelled fields without elevation correction.

Results
The comparison of daily observations and model values for the four main variables is summarized in Table 1. Before analysing each variable in the next sections, it should be noted that all models succeed in representing the daily variability of surface 10 pressure and then synoptic circulation with correlation values in the range 0.97 -0.99 (listed in supplementary materials, Table   S6 -S10).

Temperature
All model-generated temperatures correlate with PROMICE measurements with values exceeding 0.96 at the annual scale and higher than 0.82 in summer. With RMSE values representing about 30% of the daily variability (taken as the standard deviation) 15 the biases can be considered as statistically insignificant (i.e. lower than the daily variability of the PROMICE observations).
ASR outperforms the ERA reanalyses and E5 does not outperform EI: despite E5 correlating better in summer (0.85 VS 0.83), EI has a smaller RMSE in summer than E5 (2.50 • C vs 1.93 • C ).
The added value of MAR is recognisable in the yearly absolute values of MB that are smaller. In summer the temperature biases in both MAR simulations are the highest, but the same simulations show the lowest RMSE(c) and highest correlation 20 with observations (0.87). Both ERA reanalyses perform worse than MAR, while ASR shows similar statistics as MAR.
MAR E5 is colder in summer than MAR EI , but the two simulations produce similar temporal variability.
Two explanations can be given for the statistical differences in T2M between the models. First, a difference in altitude may exist between the AWS and the corresponding interpolated model elevation, which mainly influences the annual MB. For example, the interpolated elevation of the EI grid is 770 m higher at AWS QAS_L (see Table S1 in supplementary materials) 25 while the difference in altitude is lower for the other models (151 m for E5, 6 m for ASR and 119 m for MAR). This difference leads to a negative MB of EI (-4.89 • C, Table S15) and erroneously suggests that this model is colder at this location. The second difference in the modelling of T2M is the better representation by the two regional models (MAR and ASR) of the physical processes at the surface of the GrIS. This consequently results in a better representation of surface-atmosphere interactions, which are influenced by the melt of the snow pack when the excess energy is used to melt snow or ice and not to warm the 30 surrounding air, and by the density of the snow pack, which is better modelled in the polar RCMs. Resolving surface processes when the statistics of both ERA datasets are worse than those of RCMs.
The finer resolution of the RCMs and the inherently better representation of the topography could also play an important role in the better representation of climate variables. However, it appears to be irrelevant here, since the new reanalysis E5 has a resolution similar to MAR and ASR, and E5 does not outperform EI in terms of daily near-surface temperature. For example,

5
AWSs where elevation differences are less than 100 m (NUK_U, KPC_U, KAN_U, UPE_U, TAS_A, NUK_N), T2M from EI and E5 are better represented annually than in summer. By contrast, T2M from both RCMs for the same AWS have significantly better statistics in summer than both ERA reanalyses.
Finally, annual T2M simulation by ASR (correlation 0.98) is slightly better than for MAR (correlation 0.97), while the two MAR simulations have smaller RMSE values. The difference between ASR and MAR might be due to the assimilation of To conclude, MAR shows the best accuracy when modelling T2M which might also lead to a better representation of the surface energy balance and melt (not evaluated here).

Wind speed
W10M in each model is well correlated with observations (Table 1, annually > 0.79 and in summer > 0.73) and has an insignificant RMSE representing 70 % of the daily variability (taken as the standard deviation), except for ASR in summer where the RMSE is higher than the standard deviation.
Wind speed depends on synoptic atmospheric features, but also on interactions with the surface and local topographic con-20 ditions. These generate persistent and widespread katabatic winds, and winds being channelled through valley in mountainous coastal areas of Greenland. It is difficult for models to correctly represent the surface wind regime in mountainous areas due to their resolution exceeding the topographical length scales.
E5 correlate better with in-situ observations than EI, MAR and ASR at the annual and summer time scales, and also has a smaller RMSE and RMSEc.

25
Despite the improved representation of W10m in E5, both EI and E5 underestimate W10M (negative bias between -1.06 and -1.04 ms −1 ) as also shown by Moore et al. (2016) over Greenland and Jones et al. (2016) over Antarctica. Nevertheless not all PROMICE AWSs are located near the relatively steep ice sheet margin where mountains may or may not be present, disallowing the models to well capture the katabatic winds in the shallow atmospheric boundary layer. The models can also be not able to reproduce the near-surface temperature deficit, and/or they have too few levels near the ice sheet surface. It should 30 be noted that the underestimation of wind speed would be even larger at the height (∼3 m) at which wind is measured by the PROMICE AWSs.
W10M in ASR and in both MAR simulations (UV1 in Table 1) is overestimated with a positive bias higher than 1.3 ms −1 .
The biases are reduced for MAR wind speed at ∼2 m (UV2 in Table 1), which is more similar to the height of the AWS 6 measurements. However, the correlation of the wind speed is neither sensitive to the vertical level used in MAR (2 m vs 10 m), nor to EI versus E5 forcing.

Longwave downward radiative flux
Contrary to the near-surface wind speed and temperature observations that are usually assimilated in reanalyses, observed downward radiative fluxes are usually not. Forecasted radiative fluxes simulated by the three reanalysis models have been 5 compared to in-situ observation of radiative fluxes. Table 1 shows that each model has a satisfactory representation of LWD. Differences with PROMICE observations are small, with all the models underestimating LWD by 10 -16 Wm −2 .
E5 performs the best for LWD producing the highest correlation coefficients (0.94 annually, 0.89 in summer) and the smallest RMSE.

10
The two MAR simulations are similar, but the reanalyses show more favourable comparisons. The temporal variability of LWD is better represented by the reanalyses, yet the yearly MB are smaller for MAR EI (-11.35 Wm −2 ) and MAR E5 (-10.58 Wm −2 ).
The better LWD statistics of the three reanalyses compared to MAR EI and MAR E5 is likely related to their assimilation of radiance from satellite data, as well as the assimilation of sparse coastal temperature and humidity profiles from radiosondes 15 (Dee et al., 2011;Bromwich et al., 2016). This enables a better representation of incident radiative fluxes that depends on clouds and their microphysical characteristics, including the thickness, water phase and temperature. MAR does not assimilate such observations, but is only forced at its lateral boundaries every 6-hours (specific humidity and temperature). Clouds in MAR are the outcome of the own climate and microphysics of the model.

Shortwave downward radiative flux
20 Table 1 shows that each model performs well at representing SWD (yearly correlation >= 0.97 and summer correlation >= 0.88) and differences with PROMICE observations are not significant.
Similar to the LWD statistics, the reanalyses represent SWD better than the RCMs, with E5 providing the best statistics.
The ASR reanalysis overestimates SWD (yearly MB = 7 Wm −2 and summer MB = 23 Wm −2 ) whereas other models underestimate SWD (MB = -4 Wm −2 on average), as also highlighted by Bromwich et al. (2018). Large LWD and SWD 25 biases in ASR indicate that additional model improvements in Polar WRF are necessary to better capture the radiative cloud effects despite improved model cloud physics between ASRv1 and ASRv2 (Bromwich et al., 2018).
The assessment of SWD as represented by both MAR experiments reveals no significant difference, but a less accurate SWD temporal variability than in the ERA reanalyses.
In general, the accurate model representation of incident radiative fluxes (LWD and SWD) depends on radiative scheme 30 of model. The scheme used by MAR is the one from ERA-40 (the previous ECMWF reanalysis before EI) which has been updated for the EI and E5 reanalyses. This, combined with the fact that the reanalyses assimilate observations within the RCM

Additional analysis
The same statistical comparison with GC-Net  observations was performed to better cover the Greenland ice sheet, as GC-Net stations are mainly located in the accumulation area. However, it is important to note that GC-5 Net observations are assimilated into reanalyses (EI, E5 and ASR) but not into MAR. Therefore, the comparison of models with GC-Net observations was carried out separately from PROMICE observations in order to keep the independence of the PROMICE comparison with data assimilation. The conclusions of this comparison are identical to the results presented above, except that the assimilation of this data set into reanalyses favours the reanalyses for the representation of T2M with respect to MAR. A more detailed analysis of the results can be found in the supplementary materials (see Table S5).

Discussion and conclusions
We have evaluated essential near-surface climate variables (2-m temperature, 10-m wind speed and energy downward fluxes) simulated by the new ERA5 reanalysis against EI, ASR, and MAR forced by EI and by E5 for the period 2010 -2016.
EI is usually used as a reference over Greenland while ASR is a regional reanalysis specifically developed for the Arctic region. E5 outperforms EI for most variables, but not significantly. ASR is able to model temperature more accurately than the 15 other global reanalyses. Near-surface wind speed is underestimated by both ERA reanalyses.
MAR performs less satisfactorily than the reanalyses in terms of downward solar and infrared fluxes likely because of its older radiative scheme, and because it does not assimilate satellite data within its domain. Still, near-surface temperature, especially in summer, is calculated with more accuracy by MAR, suggesting that MAR does well in resolving processes in the shallow atmospheric boundary layer over the Greenland ice sheet. A good representation of T2M is important because of 20 its importance to snow and ice processes such as the surface melt. In order to better simulate SMB, there is still an interest of using polar RCMs like MAR, not constrained by observations, to represent the near-surface climate over Greenland in the ablation zone compared to E5.
We also evaluated the sensitivity of MAR to its lateral forcing, using both E5 and EI. For each analysed variable, results from both MAR simulations are highly similar, except that MAR E5 generates slightly lower near-surface temperatures than 25 MAR EI , illustrating the ability of regional climate models to simulate climate in detail when forced by reanalyses.
Since September 2019, E5 has replaced EI, and it covers a long and homogeneous period (planned from 1950 to present).
This represents a significant advantage compared to the discontinuity between ERA-40 and ERA-Interim in 1979, which can be of consequence to SMB reconstructions (e.g. Fettweis et al., 2017). In this study we showed that E5 is superior in simulating the near-surface climate of the GrIS over EI, while the advantage is not large. However, when reconstructing SMB back in time 30 to 1950, using E5 as forcing has clear advantages in terms of continuity.