WRF v.3.9 sensitivity to land surface model and horizontal resolution changes over North America

. Understanding the differences between regional simulations of land-atmosphere interactions and near-surface conditions is crucial for a more reliable representation of past and future climate. Here, we explore the effect of changes in the model’s horizontal resolution on the simulated energy balance at the surface and near-surface conditions using the Weather Research and Forecasting (WRF) model. To this aim, an ensemble of twelve simulations using three different horizontal 5 resolutions (25 km, 50 km and 100 km) and four different Land Surface Model (LSM) conﬁgurations over North America from 1980 to 2013 is developed. Our results show that ﬁner resolutions lead to higher surface net shortwave radiation and maximum temperatures at mid- and high latitudes. At low latitudes over coastal areas, an increase in resolution leads to lower values of sensible heat ﬂux and higher values of latent heat ﬂux, as well as lower values of surface temperatures and 10 higher values of precipitation and soil moisture in summer. The use of ﬁner resolutions leads then to an increase in summer values of latent heat ﬂux, convective and non-convective precipitation and soil moisture at low latitudes. The effect of the LSM choice is larger than the effect of horizontal resolution on the near-surface temperature conditions. By contrast, the effect of the LSM choice on the simulation of precipitation is weaker than the effect of horizontal resolution, showing larger 15 differences among LSM simulations in summer and over regions with high latent heat ﬂux. energy shortwave (SNET, W/m 2 ), longwave (LNET, W/m 2 ), net radiation absorbed by the soil (RNET, W/m 2 ), latent heat ﬂux (LH, W/m 2 ), sensible heat ﬂux (HFX, W/m 2 ) and ground heat ﬂux (GHF, W/m 2 ). The temporal aver- ages of near-surface conditions are estimated using outputs of 2 m air temperatures (SAT, o C ), daily maximum SAT (TASMAX, o C ), daily minimum SAT (TASMIN, o C ), soil temperature at 1m depth (GST 1m, o C ), accumulated convective and non-convective precipitation at the surface (PRECIP C and PRECIP NC, mm/day ), soil moisture contained in the ﬁrst soil meter (SM 1m, m 3 /m 3 ) and total cloud fraction (TCLDFR, % ). All values are computed using the annual and seasonal (boreal winter, DJF; summer, JJA) averages over the 34-year period (1980-2013) after discarding the ﬁrst year of the simulation (1979) as spin up, which is enough to avoid the effect of initial conditions (García-García et al., 2020). Thus, we estimated the anomalies of all outputs for each LSM simulation relative to the multi-model mean (the mean of CLM4, NOAH, NOAH-MP, and NOAH-MP-DV outputs) for each set of simulations with different resolution (25 km, 50 km and 100km). Similarly, we estimated the change in the simulation of all variables between the 100 km and 50 km simulations and between the 50 km and 25 km simulations for all LSM conﬁgurations. When required, outputs of all WRF experiments were mapped to the grid of the observational reference employed in this study by selecting the nearest model grid point. A Student’s t-test considering autocorrelation was used to identify signiﬁcant differences between simulations with different LSMs

estimated the temporal averages of surface energy fluxes for the analysis period  focusing on the following energy components: net shortwave radiation (SNET, W/m 2 ), net longwave radiation (LNET, W/m 2 ), net radiation absorbed by the soil (RNET, W/m 2 ), latent heat flux (LH, W/m 2 ), sensible heat flux (HFX, W/m 2 ) and ground heat flux (GHF, W/m 2 ). The temporal aver-160 ages of near-surface conditions are estimated using outputs of 2 m air temperatures (SAT, o C), daily maximum SAT (TASMAX, o C), daily minimum SAT (TASMIN, o C), soil temperature at 1m depth (GST 1m, o C), accumulated convective and non-convective precipitation at the surface (PRECIP C and PRECIP NC, mm/day), soil moisture contained in the first soil meter (SM 1m, m 3 /m 3 ) and total cloud fraction (TCLDFR, %). All values are computed using the annual and seasonal (boreal 165 winter, DJF; summer, JJA) averages over the 34-year period (1980-2013) after discarding the first year of the simulation (1979) as spin up, which is enough to avoid the effect of initial conditions (García-García et al., 2020). Thus, we estimated the anomalies of all outputs for each LSM simula- configuration. We calculated annual and seasonal WRF and DAYMET biases in these variables relative to the CRU data for the analysis period averaging over six subregions, due to the large climate differences over North America. These six subregions cover Central and North America (NA) and are adapted from Giorgi and Francisco (2000): Central America, CAM; Western North America, WNA; Central North America, CNA; Eastern North America, ENA; Alaska, ALA; and simulation of extreme events than on surface climatologies (Prein et al., 2013;Di Luca et al., 2015;Rummukainen, 2016). We examined this by calculating the bias in the 95th percentile of maximum and minimum temperature and accumulated precipitation within all experiments using the DAYMET product as reference. The net radiation absorbed by the ground surface enhances turbulent (latent and sensible) fluxes at the surface and/or warms the soil surface, which leads to an increase in the emitted longwave radiation (Bonan, 2002). The relationship between these variables is shown by their corresponding 200 ensemble mean of LSM simulations, indicating similar latitudinal patterns in net radiation, turbulent fluxes and near-surface temperatures with higher fluxes and temperatures at lower latitudes (see for example Figure 2 for the LSM ensemble mean of the 50 km experiments). The net radiation results from adding net shortwave and longwave radiations, whose mean values have similar spatial distributions but with opposite sign (Figure 2a). This indicates that more shortwave radiation reaches 205 the land surface than is reflected due to surface albedo, while the radiation emitted from the soil due to its surface temperature, is higher than the longwave radiation reaching the soil surface ( Figure 2a).
The energy proportion of net radiation that is propagated through the soil is much smaller than the rest of surface energy fluxes (Figure 2b). Areas with high latent heat flux coincide with areas with high convective precipitation rates at low and middle latitudes (Figure 2b  s , where T s is surface temperature and σ is the Stefan-Boltzmann constant. Thus, the CLM4 simulation produces the largest outgoing longwave radiation (see Figure   S7), shown in Figure 3b by the largest negative anomalies and therefore the highest air and soil tem-235 peratures ( Figure 4). The opposite behaviour is performed by the NOAH LSM, yielding the lowest upward longwave radiation (Figures 3b and S7), and one of the coldest temperature climatologies relative to the multi-model mean ( Figure 4). The relationship between the radiation and temperature anomalies is also supported by high spatial correlation coefficients (Table S1). These correlation coefficients show the link of maximum temperatures to both shortwave and longwave net radiation, 240 particularly in summer, while minimum temperatures show higher correlation with longwave net radiation than with shortwave net radiations in most of LSM simulations (Table S1).
The simulation of sensible heat flux reaches the highest values using the CLM4 configuration over the boreal forest and the lowest values in western US. Meanwhile, the NOAH simulation reaches the lowest sensible heat fluxes over the boreal forest and the highest values in western US (Figure 3e).

245
The spatial patterns of LSM anomalies in sensible heat flux are similar to the LSM anomalies in net shortwave radiation and daily maximum temperatures (Figures 3a and 4a), which is also shown by high spatial correlation coefficients, particularly in summer (Table S1). LSM differences in ground heat flux are smaller than for the rest of the energy fluxes due to the small magnitude of the GHF in comparison with the rest of energy components ( Figure 2g). The NOAH LSM reaches the lowest  (Figures 3f and S1). The spatial pattern of LSM differences in ground heat flux differ from the soil temperature results, whose spatial correlation coefficients are higher with the longwave net radiation mainly in summer. LSM differences are larger for the simulation of soil temperatures than for the simulation of air temperatures particularly 255 at high latitudes in summer where LSM differs largely in the simulation of shortwave net radiation, probably due to different estimates of surface albedo under different land cover and soil moisture values (Figures 2a,5c, and S3).
The CLM4 simulation produces the highest latent heat flux values and convective precipitation rates, particularly over southwestern NA, while the NOAH simulation provides the lowest latent 260 heat flux and convective precipitation values over the same areas (Figures 3d and 5a). LSM differences in latent heat flux, convective and non-convective precipitation are larger in summer (Figures S1 and S5). LSM differences in JJA convective precipitation rates are particularly large at low latitudes, where the CLM4 LSM produces high latent heat flux over the western part of the domain and low latent heat flux over the eastern NA relative to the multi-model mean ( Figure S1 and S5).

265
This differences between the east and west in the CLM4 simulation of latent heat flux are not reflected in the values of convective precipitation rates, hence the low spatial correlation coefficients between both variables (Table S2). There are also LSM differences in the non-convective term of precipitation, with larger anomalies in summer than in winter (Figures 5b and S5). Thus, the NOAH LSM produces the highest precipitation anomaly at mid-latitudes, where the same LSM produces 270 high values of total cloud fraction relative to the multi-model mean (Figure 5b and d). This relationship between non-convective precipitation and cloud cover is also shown by high spatial correlation coefficients (Table S2)

Resolution impact on surface energy fluxes and near-surface conditions
The response of surface energy fluxes and near-surface conditions to changes in spatial resolution varies considerably with the season, while they behave similarly using different LSM components Consistently with the effect of LSM differences on near-surface conditions (Section 4.1), the spatial pattern of the resolution impact on net shortwave radiation is similar to the resolution-induced 315 changes in daily maximum temperatures (Figures 6a and 7a). The response of minimum temperature to changing resolution is, however, smaller than for maximum temperatures at mid-high latitudes ( Figure 7b). Over eastern North America, JJA minimum temperature increases with the use of coarser resolutions, while it decreases over western North America (Figure 7b). The response of mean temperature to resolution is mainly driven by the resolution impact on maximum tempera- Non convective precipitation increases with the use of coarser horizontal resolutions over the Rocky Mountains, particularly in winter, where the model also represents higher percentage of cloud cover with coarser resolutions (Figure 8b and d). Over the east coast of the US, the use of coarser 330 resolutions leads to lower non-convective precipitation rates (Figure 8b). This behaviour is also present in JJA over the arctic areas of our domain (Figure 8b). Although the response of convective precipitation to resolution in winter is not significant, in JJA the use of coarser horizontal resolutions yields a decrease in convective precipitation over coastal areas and the Rocky Mountains, where the simulation also reaches lower latent heat flux values (Figures 6d and 8a). Although the spatial pattern 335 of soil moisture is very patchy in DJF and JJA, soil moisture tends to decrease at low latitudes in JJA with the use of coarser resolutions (Figure 8c). At mid-latitudes, however, the use of coarser resolution leads to an increase in soil moisture during the year at most locations (Figure 8c).
In summary, at low NA latitudes the JJA values of the three variables associated with the surface water balance (LH, PRECIP and SM 1m) decrease with the use of coarser horizontal resolutions, 340 although showing large spatial variability (Figures 6d and 8a and c). At mid-and high NA latitudes, there are differences in the response of the water balance variables to the use of coarser resolutions.
For example, soil moisture increases with coarser resolutions over a large area at mid-latitudes, while convective precipitation increases just over a few grid cells in central NA, decreasing over most coastal areas (Figure 8a and c). Latent heat flux decreases with the use of coarser resolutions 345 over most regions at high and mid-latitudes, particularly over coastal areas (Figure 6d).  Figure 9a). Over these areas, finer horizontal resolutions are associated with warmer JJA maximum temperatures, reducing the bias relative to the CRU dataset at middle latitudes (Figures 7a and 9a). In the ALA and GRL regions, the WRF model with the CLM4 and the NOAH-MP LSM components overestimates JJA maximum temperatures, increasing the bias in these simulations with the use of finer resolution (Figure 9a). The WRF bias in maximum 365 temperatures in winter is greatly improved over the boreal forest and the Rocky areas by using the CLM4 as LSM ( Figure S15). Over the same areas the CLM4 simulated very high values of shortwave net radiation and sensible heat flux in comparison with the rest of land surface models ( Figure   3a and e), which may be related to the CLM4 albedo estimate. Despite the WRF underestimation of mean maximum temperature, extreme maximum temperatures are overestimated by the WRF model, 370 particularly at high latitudes and using the CLM4 LSM ( Figure 10a). As expected based on the literature, the resolution effect on the bias in extreme maximum temperatures is larger than on the bias in mean maximum temperatures, however LSM differences are still larger than resolution-induced changes.

Comparison of temperature and precipitation against observations
The performance of the WRF model in reproducing daily minimum temperatures from the CRU 375 observations is slightly better than reproducing maximum temperatures at mid-and low latitudes, but it is worse at high latitudes particularly in DJF (Figure 9b). Experiments using the CLM4 LSM yield a warmer climatology over most areas and for all seasons than the experiments with the other LSM components, implying smaller biases in the CLM4 simulations for most of regions ( Figure   9b). The WRF bias in minimum temperature is large in winter over the central and eastern areas of 380 North America and at mid-and high latitudes (subdomains ALA, GRL, CNA and ENA in Figure   9b). The resolution impact on these results is again weaker than the effect of the LSM component.  (Figure 9a and b).
The WRF model simulates large positive biases in daily accumulated precipitation at the surface over most of North America during all seasons, with larger biases in summer (Figures 9c and S17).
of North America in summer and winter ( Figure S17). Dry biases are reduced when using finer horizontal resolutions, while wet biases are larger when using smaller scales (Figure 9c). This is due to the intensification of the water cycle with the use of finer horizontal resolutions discussed in section 4.2 and presented in Figures 6d and 8. For example in winter, the dry bias shown in all 400 experiments over the southeastern NA is associated with an increase in non-convective precipitation using finer resolutions (Figures S17 and 8b). In summer, the bias in precipitation is larger using finer resolutions over most of coastal areas where an increase in convective precipitation and latent heat flux were shown with the use of finer resolutions (Figures 6d, 8a and S17). The impact of resolution on the accumulated precipitation is stronger than the effect of the LSM component, which affects 405 precipitation mainly in summer (Figure 9). The results show larger bias in extreme precipitation than for the mean accumulated precipitation, but yielding similar conclusions (Figure 10).
In summary, the LSM impact on temperatures is larger than the resolution effect, while the opposite is true for precipitation climatologies, i.e. differences in precipitation arising from changes in resolution are larger than LSM differences (Figure 9). The influence of both the LSM choice 410 and resolution intensifies in summer comparing with the rest of seasons, probably because of the larger energy exchanges and the consequent intensification of land-atmosphere coupling in summer (Zhang et al., 2008;Mei and Wang, 2012). The CLM4 LSM generates the smallest biases relative to the CRU database in the WRF simulation of mean maximum and minimum temperature, however it also yields the larger biases in extreme maximum and minimum temperatures. The use of finer The large WRF sensibility of precipitation rates to resolution is also supported by the literature (e.g. Pieri et al., 2015). Our results also shown large seasonal differences, mainly caused by the different contribution of convective precipitation in summer and winter. In summer at middle and low 440 latitudes, the use of finer grid cells leads to a change in the energy partition into sensible and latent heat flux, increasing latent heat and decreasing sensible heat ( Figure 6). This increase in latent heat flux over these areas is probably the caused of the higher values of convective precipitation ( Figure   8a). This resolution-induced increase in precipitation through changes in convective processes has also been suggested in the literature (Prein et al., 2016). At high latitudes both turbulent heat fluxes 445 increase in summer with the use of finer resolutions, mainly due to the increase in shortwave net radiation probably related to the decrease in cloud cover shown in Figure 8d. These changes in cloud cover with resolution may be caused by the performance of the microphysical parameterizations at different resolutions and the improvement in the representation of orography (Pieri et al., 2015;Prein et al., 2016).

450
Previous evaluations of RCMs using several soil models over different domains reached the conclusion that the most complex LSM components, that is, the LSM components representing more physical phenomena, outperform others (Chen et al., 2014;Van Den Broeke et al., 2018;Liu et al., 2019). Over North America, our results indicate that the WRF simulation of temperature conditions using the CLM4 LSM outperforms the simulation of mean maximum and minimum temperatures 455 generated by the NOAH and NOAH-MP LSMs, but it yields larger biases in extreme maximum and minimum temperatures (Figures 9 and 10). The simulation of precipitation in summer is, however, slightly better represented by the NOAH LSM than by the other LSM components (Figure 9). Nonetheless, the comparison of all WRF experiments with observations shows overestimated values of precipitation over most of North America in agreement with other studies using WRF over the 460 western US (Jin et al., 2010;Chen et al., 2014) and over Europe (Pieri et al., 2015). Atmospheric parameterizations were not tested in our study; however, other WRF sensitivity experiments using several microphysics schemes over Europe found a positive bias in precipitation for all simulations, which was considerably reduced in summer within a convective permitting simulation (Pieri et al., domains using several LSM components, horizontal resolutions, microphysics parameterizations, and reanalysis products as initial and boundary conditions (Figures 9 in this manuscript, Pieri et al., 2015;Chen et al., 2014;Jin et al., 2010). Therefore, the results included here together with the results reported in the literature suggest that the use of finer resolutions may raise precipitation biases in WRF simulations over North America, but the implementation of convective-permitting processes 470 and other atmospheric parameterizations could reduce this bias.

Conclusions
This study has shown the effect of changes in horizontal resolution and LSM choice on the simulation of energy fluxes at the surface and temperature and water conditions in the near-surface.
The effect of both model choices intensifies in summer, due to the increase in energy and water ex- Information provided by downscaling studies are used for building climate change policies, through the information collected in assessment reports (IPCC, 2013;Mbow et al., 2017;Reidmiller et al., 2019). Thus, sensitivity studies like the one presented here, are crucial to understand and ultimately restrict uncertainties in climate simulations, with direct benefits to society and environment. Par-500 ticularly, these results should be considered for downscaling studies over North America aimed at projecting future or past conditions and informing policy-makers.
Code and data availability. The source code of the Weather Research and Forecasting model (WRF v.3.9 Giorgi, F. and Francisco, R.: Uncertainties in regional climate change prediction: a regional analysis of ensemble resolution in global and regional climate simulations for European climate extremes, Geosci. Model Dev. Lucas-Picher, P., Laprise, R., and Winger, K.: Evidence of added value in North American regional climate Table 1. Summary of the regional simulations performed in this analysis.

NAME LSM Resolution Vegetation Mode Simulation Time Step Radiation Time
Step        Eastern North America, ENA; Alaska, ALA; and Greenland, GRL.