The WRF model version 3.8.1 has been used in the present study. WRF is a mesoscale, non-hydrostatic, numerical weather prediction (NWP) model with advanced physics and numerical schemes for simulating meteorology and dynamics. WRF-ARW uses an Eulerian mass-based dynamical core with terrain-following vertical coordinates (Skamarock et al., 2008). ERA-Interim reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF), available at a temporal resolution of 6 h and a horizontal resolution of 0.75∘×0.75∘ with 37 vertical levels from the surface to the top at 1 hPa (Dee et al., 2011), has been used to provide the initial and lateral boundary conditions to the WRF model. Static geographical data from the Moderate Resolution Imaging Spectroradiometer (MODIS), available at 30 s horizontal resolution, are utilized for land use and land cover.

The Goddard scheme is used for shortwave radiation (Chou and Suarez, 1994), while longwave radiation is simulated by the rapid radiative transfer model scheme (Mlawer et al., 1997). For resolving the boundary layer processes the first-order Yonsei University (YSU) scheme based on non-local closure (Hong et al., 2006) is used, including an explicit entrainment layer with the K-profile in an unstable mixed layer. The planetary boundary layer (PBL) height is determined from the Richardson number (Rib) method in this PBL scheme. Convection is parameterized by the Kain–Fritsch (KF) cumulus parameterization (CP) scheme, accounting for sub-grid-level processes in the model such as precipitation, latent heat release, and the vertical redistribution of heat and moisture as a result of convection (Kain, 2004). With the increase in model grid resolution to less than 10 km (known as “grey area”), the CP scheme is usually turned off, and cloud and precipitation processes are resolved by the microphysics (MP) scheme (Weisman et al., 1997). In the present study, the CP scheme is used for d01, while it is turned off for d02 and d03. The Thomson microphysics scheme containing prognostic equations for cloud water, rainwater, ice, snow, and graupel mixing ratios is used (Thompson et al., 2004). Parameterization of surface processes is done with the MM5 Monin–Obukhov scheme and unified Noah land surface model (LSM) (Chen and Dudhia, 2001; Ek et al., 2003; Tewari et al., 2004). The Noah LSM includes a single canopy layer and four soil layers at 0.1, 0.2, 0.6, and 1 m within 2 m of depth (Ek et al., 2003).

The model is configured with three domains of 15 km (d01), 5 km (d02), and 1 km (d03) horizontal grid spacing using Mercator projection centering at Manora Peak (79.46∘ N, 29.36∘ E; ∼ 1936 m a.m.s.l.) in the central Himalaya. The topography within the model domains is highly complex, as evident from the ridges (Fig. 1). The outer domain d01 includes the northern part of the Thar Desert, part of Indo-Gangetic Plain (IGP), and the Himalayan mountains, while the innermost domain d03 consists of mostly mountainous terrain. The model has 51 atmospheric vertical levels with the top at 10 hPa. For d01, 100 east–west and 86 north–south grid points are used to account for the effect of synoptic-scale meteorology, e.g., the Indian summer monsoon. The d02 has 88 east–west and 76 north–south grid points covering a sufficient spatial region around the observational site to consider the effects of mesoscale dynamics, e.g., changes in wind pattern due to orography. The innermost domain, d03, has 126 east–west and 106 north–south grid points to resolve local effects, e.g., convection, advection, turbulence, and orthographic lifting.

For d01, boundary conditions are provided from the ERA-Interim reanalysis, as explained earlier. Model simulations have been performed for the 4 months of the summer monsoon: 1 June 2011 to 30 September 2011 (JJAS). This simulation period is chosen considering the availability of continuous observations from 11 June 2011 and to allow a sufficient spin-up time of 10 d for the model to achieve its equilibrium state (Angevine et al., 2014; Seck et al., 2015; Jerez et al., 2020). Only the outer domain d01 is nudged with the global reanalysis for temperature, water vapor, and the zonal and meridional (u and v) components of the wind using a nudging coefficient of 0.0006 (6 × 10-4) at all vertical levels (e.g., Kumar et al., 2012). Several of the configuration options, e.g., physics and meteorological nudging, are selected following earlier applications of this model over this region (e.g., Kumar et al., 2012; Ojha et al., 2016; Singh et al., 2016; Sharma et al., 2017).

We utilize observations during an intensive field campaign – the Ganges Valleys Aerosol Experiment (GVAX) – to evaluate model simulations. The GVAX campaign was carried out using the Atmospheric Radiation Measurement (ARM) Climate Research Facility of the U.S. Department of Energy (DOE) from 10 June 2011 to 31 March 2012 at ARIES, Manora Peak in Nainital (e.g., Kotamarthi, 2013; Singh et al., 2016; Dumka et al., 2017). This observational site (79.46∘ N, 29.36∘ E; 1940 m above sea level) is located in the central Himalaya, as shown in Fig. 1. The surface-based meteorological measurements of ambient air temperature, pressure, relative humidity, precipitation, and wind (speed and direction) were made using an automatic weather station at 1 min temporal resolution. The instantaneous values of the observations are compared with hourly instantaneous model output at the nearest grid point.

The vertical profiles of temperature, pressure, relative humidity, and horizontal wind (speed and direction) were obtained by four launches (00:00, 06:00, 12:00, and 18:00 UTC) of the radiosonde each day during the campaign (Naja et al., 2016). The continuous vertical profiles of the meteorological parameters except wind speed and direction were available from the end of June 2011 through the entire study period, whereas valid and quality wind data were available only for September 2011. Hence, in this study, radiosonde measurements from 1 July 2011 onwards are used for the model evaluation of meteorological parameters, except wind speed and direction, which are evaluated only for September. A total of 309 valid profiles of temperature and relative humidity and 104 profiles of wind are used. Statistical metrics such as the mean bias (MB), root mean square error (RMSE), and correlation coefficient (r) are used for the model evaluation, and a description of these metrics is given in the Supplement.

Here, we have used the ERA-Interim data for comparison with WRF output. We first compare the WRF-simulated spatial distribution of meteorological parameters (surface pressure, 2 m air temperature, 2 m RH, and 10 m WS) with the ERA-Interim reanalysis over the common region of all the domains averaged for the entire simulation period (Fig. 2). The three contours of the topographic height of 500, 1500, and 2000 m are used to relate the meteorological features to the resolved topography in three domains. The common area in all domains includes the low-altitude IGP region in the south (elevation of less than 400 m; Fig. 1) and elevated mountains of the central Himalaya in the north. Also, for a consistent comparison, model-simulated values are taken at the same time intervals as in ERA-Interim data (i.e., every 6 h). From the comparison presented in Fig. 2, it is evident that the meteorological parameters simulated by the model are dependent on the model grid resolution. The existence of the sharp-gradient topographic height (SGTH) of about 1600 m from the foothill of the Himalaya to the observational site modifies the wind pattern and moisture content differently at different grid resolutions, indicating the critical role of mountain orography. The surface pressure explicitly depends upon the elevation of a location from mean sea level. The contour of the pressure parameter from ERA-Interim data shows the surface pressure of about 900 hPa for the observational site Manora Peak, and it varied from 550 to 975 hPa within this region, while WRF-simulated pressure is 869, 835, and 827 hPa for d01, d02, and d03, respectively. WRF-simulated surface pressure ranges from 821.9 hPa over the high-altitude CH region to 977.0 hPa in the IGP region within d01. Simultaneously, the range of variation in the surface pressure is 788.1–977.5 and 760.4–977.7 hPa within d02 and d03, respectively, and the minimum pressure decreases from d01 to d03, which is attributed to the improvement in resolved topography on increasing model grid resolution. However, the effects of the SGTH are not observed for temperature, wind, and RH in ERA-Interim contours due to the unresolved topographic features. Simulated maps show the spatial homogeneity of meteorological parameters over the flat terrain of IGP in the foothills of the Himalaya compared to the elevated central Himalayan region.

The effect of spatial resolution is clearly observed over the mountainous region of the Himalaya, where the size of the mountains changes abruptly, with the modeled output showing increasingly distinct features with increasing grid resolution. On the other hand, there are minimal differences in the topography of the IGP, and hence the meteorological features associated with the topography are well captured in the model even at a coarser resolution of 15 km.

Model simulations show the topography-dependent spatial variation in 2 m temperature in the ranges of 20.0–29.5 ∘C in d01,17.3–29.6 ∘C in d02, and 15.5.0–29.9 ∘C in d03, with the lowest values simulated over the elevated mountain peaks and higher values over the temperate IGP region. The contours in three model domains show an explicit dependency of 2 m temperature on the grid resolution over the mountainous region. With increasing model resolution, the topography is resolved to a greater extent, and lower temperature is simulated at higher surface elevations, as expected. Further, the estimation of water vapor is essentially needed for both climate and numerical weather prediction (NWP) applications. Figure 2 shows that the simulated relative humidity is above 70 % in all three domains for the monsoon season. The variations (minimum–maximum) in the relative humidity in the ERA-Interim (80 %–85 %) dataset over the domains d01 (77 %–93 %), d02 (74 %–95 %), and d03 (70 %–96 %) are generally comparable. The mountain slopes provide uplift to the moist monsoonal air that on ascent subsequently saturates and increases the relative humidity to about 90 % as observed over the grid encompassing the site. Contour lines (1500 and 2000 m in Fig. 2) depict the low pressure and temperature with the higher relative humidity feature of the peaks, and these features are sharper as the resolution increases from d01 to d03.

The wind speed is highly dependent upon the model grid resolution and orography-induced circulations during different seasons (Solanki et al., 2016, 2019), and this is reflected in Fig. 2. As mentioned earlier, although the topography of the IGP region does not vary abruptly, the magnitude of the wind speed over this region and over the complex Himalayan region is found to change significantly at different model resolutions, thereby indicating that the wind speed is very sensitive to both model resolution and topography. The wind speed in d01 varies from 1.3 to 2.8 m s-1, while the wind variations in domains d02 (1.2–3.4 m s-1) and d03 (1.3–4.2 m s-1) show higher variability than that in ERA-Interim (1.2–2.3 m s-1) due to finer resolution of the WRF. Overall, the impact of the topography resolved at higher resolution in WRF shows the contrasting differences in surface pressure, temperature, relative humidity, and wind speed compared to the coarse-resolution ERA-Interim dataset.

The intensive radiosonde observations made during the GVAX field campaign at Manora Peak (79.46∘ N, 29.36∘ E; ∼ 1936 m a.m.s.l.) in the central Himalaya (shown in Fig. 1) are used for the evaluation of model resolutions. The comparison of the model-simulated profiles of temperature, relative humidity, and wind speed against the radiosonde observations is shown in Fig. 3.

The inversion of temperature at the top of the troposphere occurred at ∼ 90 hPa (∼ 16 km) in observations (Figs. 3d, S1), whereas radiosonde profiles show that temperature decreases with pressure from 15.5 ∘C at 750 hPa to -78.0 ∘C at ∼ 90 hPa. As evident from the simulated temperature profiles, the WRF model captured these features well and was found to show a reduction from 15.1 to -76.6 ∘C in these pressure levels. Further, the differences between model (d01) and radiosonde observations (Fig. 3g) range from -4 to 4 ∘C. The mean RH values from the radiosonde observations (model d01) also show a decrease from 82.3 % (76.7 %) at 750 % to 25.2 % (32.0 %) at 90 hPa. The mean RH difference between observations and the model (Fig. 3h) shows that the model simulates a more humid atmosphere at higher altitudes while showing a low humidity bias at lower altitudes. The wind data from radiosonde measurements available for September 2011 were utilized to compare the model output. Observations and modeled winds are ≤10 m s-1 within the altitude region of the surface to about 400 hPa (∼ 7 km) until the middle of September (day of the year 258). Wind increases (≥15 m s-1) above 400 hPa and attains maximum values (≥25 m s-1) between 250 and 100 hPa after 258 d of the year (15 September 2011). However, simulated winds are slightly lower and less widespread compared to observations. The comparison of the wind profiles with the same x axis (shown in Fig. S2) as other meteorological parameters shows that lower relative humidity (<30 %) is observed along with higher wind speed during this period. In general, the vertical profiles and variation of simulated wind speed agree well with the observations. The Taylor diagram (Taylor, 2001) in Fig. 7a is used to express the statistical comparison between model simulations and observations. In the diagram, the comparison is summarized with the correlation coefficient (r), normalized root mean squared difference (RMSD), and standard deviation normalized by that of the observations (SD). In most cases, the model simulates less variability in meteorological parameters, as shown by the normalized standard deviation, which is less than 1. For temperature and wind speed, the model shows better correlation (r) with the observations at 250 hPa (r>0.80) than that at lower altitudes, i.e., 750 hPa (r<0.40). On the other hand, the model captures variability in humidity relatively well at 500 hPa (r = 0.71) but shows poor correlation at 75 hPa (r = 0.17) near the model top.

Lower correlations for temperature and wind speed near the surface (750 hPa) could be due to terrain-induced effects, which are most significant in the local boundary layer. The surface-level winds and turbulence are some of the boundary layer features affected mainly by the surface and terrain characteristics. The vertical profiles of these parameters up to 500 hPa in all three model domains are shown in Fig. 4. Differences between the simulated vertical profile of temperature and radiosonde observations are in general similar in all the domains. Except for the relative humidity in d01, other meteorological parameters (temperature and wind speed) do not reveal strong dependencies on the model resolution. However, the model overestimates the relative humidity near the 500 hPa level in d02 and d03 on some days. In the case of the wind speed, the model underestimates the magnitude of the wind in the first few days up to 500 hPa, though by and large the model is able to capture the vertical profiles.

Figure 5 shows the vertical profiles of the following statistical metrics for the three simulations (d01, d02, d03): mean bias (MB), root mean square error (RMSE), and correlation coefficient (r) for temperature, relative humidity, and wind speed. The magnitudes of the MB values throughout the troposphere are estimated to be within about 1 ∘C, 12 %, and 2.5 m s-1 for temperature, relative humidity, and wind speed, respectively. Additionally, RMSE values are about 1 ∘C, 15 %–30 %, and 2.5–5 m s-1 for temperature, relative humidity, and wind speed, respectively. As discussed earlier, correlations between model results and observations are found to be stronger in the middle and upper troposphere than in the lower troposphere. For temperature, the r values are higher than 0.75 between 600 and 200 hPa, whereas they decrease to 0.4 at lower altitudes, i.e., near 800 hPa. Correlations in the lower troposphere are notably weaker (r = ∼ 0.25) in the case of wind speed. The results suggest that the model captures the day-to-day variabilities well in the meteorological parameters in the middle and upper troposphere and to a minor extent in the lower troposphere. Relatively weaker correlations in the lower troposphere are suggested to be associated with more pronounced effects of the uncertainties caused by the underlying complex mountain terrain and resulting unresolved local effects. Wind fields near the surface are strongly impacted by interactions between the terrain and boundary layer in addition to orographic drag in a modeling study over the Tibetan Plateau (Zhou et al., 2018) and in measurements over the Himalaya (Solanki et al., 2019). An increase in bias with altitude was reported by Kumar et al. (2012) for dew point temperature. Besides the model physics, the higher uncertainties in radiosonde humidity observations might have also contributed to these differences. The effect of model resolution is not very significant for temperature and wind profiles above 800 hPa; nevertheless, the mean bias for RH is lower (∼ 5 %) in the 800–600 hPa altitude range and higher in the 450–300 hPa altitude range in the d02 and d03 simulations. This might be due to deep convection in the model at a higher resolution. Overall, the model captured the vertical structures of meteorological parameters; however, a better representation of complex terrain is insufficient to improve the model performance aloft. On top of the better representation of topography as considered here, it highlights the need for future studies evaluating various physics scheme. Nevertheless, model biases have been significantly reduced for surface-level meteorology with higher resolution, and the details are discussed in the subsequent section.

The model-simulated 2 m temperature (T2), 2 m relative humidity (RH2), and 10 m wind speed (WS10) for the observational site, Manora Peak, are compared with the ground-based measurements made during the GVAX campaign in Fig. 6 and summarized in Table 1. The diurnal variations in T2, RH2, and WS10 simulated by the WRF model are compared with observations, whereas the surface pressure does not show a significant diurnal variation (not shown here). Model simulation d01 shows a positive bias of 68 hPa in surface pressure, with a strong correlation (r = 0.97) with observations (mean = ∼ 801 hPa). A significant improvement is achieved (MB = 26 hPa) in d03 as a result of the finest-resolution simulation (Fig. 6 and Table 1). The WRF-model-simulated T2 shows a warm bias in all three domains. The simulated T2 for d01 varies from 16.2 to 28.7 ∘C, with a higher mean value of 22.3 ± 2.1 ∘C compared to the observed mean value of 19.5 ± 1.6 ∘C and a correlation of r = 0.75 between d01 and observations. This warm bias is seen to decrease from d01 (2.8 ∘C) to 0.2 ∘C with increasing model resolution in the d03 simulation (Table S1). The mean value of the RH2 in d01 is about 88.2 ± 9.7 %, which is 6.4 % lower than the observed value of 94.7 ± 9.5 % with a correlation of about 0.45 (Fig. 7b). MB and RMSE values of RH2 show a decrease with increasing model resolution (Table S1). As relative humidity also depends on temperature, the diurnal variation in 2 m specific humidity (Q2; g kg-1) has also been analyzed (Fig. S3). Q2 is observed in the range of 5.5–21.5 g kg-1 with a mean value of 16.8 ± 2.0 g kg-1. It is found that the agreement in Q2 is relatively better (MB = -0.7 g kg-1; r = 0.77 in d03) when the statistical metrics are compared with those for RH2 (Table S1). The wind speed plays a vital role in transport processes and controls the dynamics of the atmosphere at different temporal and spatial scales. The average 10 m wind speed (WS10) during the monsoon over the measurement station is about 2.1 ± 1.4 m s-1, which is quite comparable to that simulated in d01 (2.1 ± 1.1 m s-1), whereas it is overestimated in d02 by 0.9 m s-1 and in d03 by 0.5 m s-1 (Tables 1 and S1). In the case of the WS10, the correlation is 0.18 for d01 and d02, which improves to 0.24 in d03. The diurnal variation of WS10 (Fig. 6c) is not well captured, especially during noontime.

Due to the complex terrain and the grid size of the model, the simulated altitude of the observational site could differ from reality. In this study, the model underestimated station altitude by about 588, 480, and 270 m in d01, d02, and d03. We performed an additional evaluation to explore and achieve possible improvement by linearly interpolating the vertical profile of meteorological parameters to the actual altitude of the station (Fig. 6d–f), as done in a few previous studies (e.g., Mues et al., 2018). The altitude adjustment was made as per the equation of linear interpolation given in the Supplement (Eq. 4). The analysis shows that the correlation coefficient values between the model and observations do not show any clear improvement in model output (e.g., for T2 the correlation coefficient is 0.35) on adjusting the altitude except for WS10. After adjusting the altitude, the temperature variability is suppressed by the model at diurnal (Fig. 6a, d) and day-to-day timescales; i.e., r drops from 0.67 to 0.36 in d03. The comparison of temperature among the three domains for additional model layers has also been analyzed (Fig. S4). The diurnal amplitudes are seen to be smaller at higher model layers. Additionally, the differences among different simulations (d01, d02, and d03) also decrease at higher layers. As expected, the altitude adjustment does reduce the bias in pressure. Nevertheless, reductions in mean biases are not achieved (Table S1); instead, the absolute values of biases increase from 0.2 to 1.2, 2.4 to 3.1, 0.5 to 1.9, and 0.7 to 1.6 in T2, RH2, WS10, and Q2 in simulation d03. Besides thermal and mechanical interactions of mountain surfaces with the atmosphere, local processes such as evaporation and transpiration affect the near-surface meteorological conditions. A reduction in wind speed during the daytime is associated with the competing effects of mountain–valley circulations due to heating of the slopes versus synoptic-scale flows (Solanki et al., 2019). To resolve such sub-grid-scale processes, we emphasize that very high-resolution simulations are needed, as conducted in this study, in order to simulate the meteorological variability in a satisfactory way. The analysis further highlights the need for accurate representation of the complex topographical features rather than altitude-adjusted estimations, which led to very limited improvements in this case. However, we will discuss the evaluation without altitude adjustment unless stated otherwise.

We evaluate the MB values (Tables 1 and S1) in model simulations considering the benchmarks suggested by Emery et al. (2001). In the d03 simulation, MB values for both T2 (0.2 ∘C) and Q2 (-0.7 g kg-1) are found to be well within the range of benchmark values: ±0.5 ∘C for T2 and ±1.0 g kg-1 for Q2. It is important to note that biases in T2 in the coarser simulations d01 (2.8 ∘C) and d02 (0.9 ∘C) are, however, higher compared to the benchmarks. MB values in T2 estimated for this representative Himalayan site are found to be slightly lower (+0.2 ∘C) (-1.2 ∘C with altitude adjustment) than those over the Tibetan Plateau (-2 to -5 ∘C) (Gao et al., 2015) and over mountainous regions in the Europe (Zhang et al., 2013). The warmer bias in our case is due to underestimation of the Himalayan altitude, whereas the model overestimated terrain height over the Tibetan Plateau region, giving contrasting results. Further, the RMSE in wind speed is lower (1.6–2.0 m s-1) than that over the Kathmandu valley (2.2 m s-1; Mues et al., 2018) and similar to the benchmark (2.0 m s-1). Mar et al. (2016) also reported a similar bias (2 m s-1) in the 10 m wind speed over Europe and an average correlation of ∼ 0.4–0.6 over the Alps. Nevertheless, simulating the diurnal variation of near-surface winds still remains challenging over complex terrain, but the bias was reduced after including effects of the turbulent orographic form drag (Zhou et al., 2017, 2019). Besides turbulent orographic form drag, it is suggested that wind speed is sensitive to boundary layer schemes (Yver et al., 2013; Zhou et al., 2019) and that more studies are needed to explore these aspects over the central Himalaya.

The wind direction is strongly influenced by the surrounding topography over mountainous regions, and the evaluation of the wind direction at horizontal resolution is depicted in Fig. 8. The winds varying between meteorological directions 337.5 and 22.5∘ are considered to be northerly and are represented by N in the frequency distribution and so on for other directional flow, referring to the clockwise meteorological convention. The wind flow dominance over the observational site is easterly (30 %) and southeasterly (26 %), while 26 % of wind occurrences are from the west and northwest. The percentage of southerly (21 %) and southwesterly winds (SW, 19 %) is relatively higher in d01 compared to the observations (S: 4 % and SW: 5 %) and decreases to 4 % and 9 % in d03. The model is able to simulate the northerly and northeasterly winds in d01 and d02, while the model simulates a larger contribution of northeasterly and westerly winds in d03, which is not seen in the observations. The easterly component of the model-simulated wind shows better agreement with observations on increasing the model resolution. In addition, the model is able to simulate the westerly and northwesterly wind contribution in d02, whereas the westerly component is overpredicted by 10 % in d03. The observations show that winds blowing from the north, northeast, south, and southwest are very weak (<2 m s-1) and amount to about ∼ 15 % of the total occurrences. The diurnal variation of the wind direction is not investigated here; however, the impact of mountain topography on the near-surface flow under low wind conditions has been discussed elsewhere (Solanki et al., 2019). Overall, the simulated wind field in d03 is in relatively better agreement with observations than d01 and d02. This is further assessed in Sect. 3.5 using a finer-resolution simulation by implementing SRTM3s terrain data.

In the preceding section, the simulations were carried out without any feedback (WRF-WF) from the finer-resolution domain to its parent domain, and results have been discussed. This WRF-WF experiment was conducted in such a way that it could explicitly account for the grid resolution effects on the model performance. The simulated meteorology with this model setup (with feedback) depicted different model performance in the outermost coarse-resolution domain d01 compared to d02 and d03. The model performance depends upon the boundary and initial conditions. Another model simulation is carried out in this section using the same configuration but with two-way interactive nesting and feedback (WRF-F) from the nested domain to its parent domain. The simulated meteorological parameters in the higher nests are fed back to their parent domains, and the boundary conditions are replaced there. The model results over the CH region in the regional-scale simulation (d01) show better agreement with the observations because of the feedback from the high-resolution nested simulation. The comparison of the simulated meteorological parameters (T2, RH2, and WS10) for the outermost domain with surface observations is presented in Fig. 9 for both WRF-WF and WRF-F simulations, thus showing the effect of feedback within the outermost domain.

The comparison of mean values (Table 2) shows a decrease in model bias for T2, RH2, and Q2 by 0.5 ∘C, 0.3 %, and 0.2 g kg-1, respectively, due to feedback from finer-resolution simulations. Additionally, correlations are found to show improvements for RH2 and Q2 by 0.15 and 0.12, respectively, due to feedback; hence, the diurnal variation of relative humidity is closer to the observations (Fig. 9). Nevertheless, smaller changes were seen in correlations for WS10 (by 0.05) and T2 (by -0.02) (Fig. S5). Variations in wind speed and direction also show significant improvements, especially in the dominant flow direction, e.g., east, west, and northwest (Fig. S6).

Effects of the feedback on surface pressure, 2 m temperature, and relative humidity in the domain d01 are shown by Fig. 10. Feedback effects are seen to be more pronounced over the mountainous region than over the flat terrain of the IGP. The feedback from the nested domain to the parent domain mostly modifies the meteorology over the mountainous region, as shown by the topography contours in Fig. 10. The analyses of biases and correlations suggest an improvement in the model-simulated pressure, temperature, and humidity through feedback from well-resolved nests. This further underpins the fact that better representations of the Himalaya over local scales can be adopted to simulate meteorology at the regional scale with lower biases over complex terrain in the given domain. Nevertheless, further modeling studies along with more observations are needed to improve the model performance. We extended the efforts to improve the wind speed and direction simulated over the complex topography by implementing a high-resolution (3 s) topographical input in the model to evaluate finer-resolution features over the Himalaya in the following section.

Simulations described in previous sections were performed using the 30 s (∼ 1 km) topographic data from the GMTED2010 (Danielson and Gesch, 2011), which is comparable to the highest resolution of the WRF simulation (d03). In order to evaluate the influences of topographical features on the wind flow at finer scales, the topography input available at very high resolution (3 s or ∼ 90 m) from the Shuttle Radar Topography Mission (SRTM3s) (Farr et al., 2007) has been utilized. However, retaining the same model configuration as earlier, an additional innermost nest d04 having a resolution of ∼ 333 m, as depicted in Fig. 1 (bottom right panel), is included. Simulations with SRTM data at 1 km resolution did not differ significantly from the simulation with a similar resolution using GMTED2010 (GMTED hereafter). For this experiment, model simulation is performed only for September 2011. This simulation is carried out without feedback and compared with the observation to check the effect of implementing high-resolution topography.

The comparison of the topographic height between GMTED and SRTM3s in Fig. S7 shows that the differences are larger over the mountainous region, which vary from -100 to +100 m. The differences are suppressed within d02 and d03 as the topography input is changed from the GMTED to SRTM3s datasets. The topography in d04 (Fig. 11c) becomes resolved better and is marked by sharp variations of mountain ridges and valleys using the SRTM3s compared to d03, which are smoothed out with GMTED (Fig. 11a) or with SRTM3s interpolated to 1 km (Fig. 11b). After including the SRTM3s topography, the MB value for wind speed in d03 is found to show a slight reduction (∼ 0.04 m s-1). Also, the flow from various directions exhibited an improvement of 1 %–2 % with the use of SRTM3s (Table S2).

In d04, surface pressure is seen to be simulated more realistically (809 hPa), and the dry bias in 2 m relative humidity is improved by ∼ 2 % (Fig. S8 and Table S2). Simulations of diurnal wind variations remain challenging (not shown here) even at the finest resolutions considered (d04) utilizing the updated topographic data (SRTM3s). Further, to understand the effect of SRTM3s data on d04, the wind direction is compared with that in d03 and observations. The variations in winds are analyzed and shown by a wind rose (Fig. 12a–d) and frequency distribution (Fig. 12e–h). The fraction of the northeasterly component in d04 with SRTM3s (5 %) is found to be comparable with observations (6 %), which was overestimated by 19 % (17 %) in d03 with GMTED (SRTM3s). The frequency of the southerlies improves with the increasing resolution of topography and better matches the observations. The observations show the prevalence of northwesterly (19 %), easterly (24 %), westerly (18 %), and southeasterly (20 %) winds, and these are also seen to be the dominant directions in the simulation d04, with the exception of southeasterly winds.

Simulation of the wind directions improved from d03 to d04 by using the SRTM3s topography, except certain wind directions such as southeasterly. An improvement is noticed in simulated surface pressure, 2 m relative humidity, and 10 m wind speed using the SRTM3s topography. Topographical data at different resolutions are found to show RH differences in the range of -1 % to 1 % in d02 and -3 % to 3 % in d03 (Fig. S9). Differences in simulated RH could be associated with the multi-scale orographic variations, which are found to be the key factors in meteorological simulation over complex terrain (e.g., Wang et al., 2020). The effects of the SRTM3s topographic static data have been studied previously over other regions of the world (e.g., Teixeira et al., 2014; De Meij and Vinuesa, 2014). However, the lower daytime wind speed and the transition phases during morning and evening hours still remain a challenge, even after using the high-resolution (333 m × 333 m) nest. Such discrepancies between the model and observations over the Himalayan region are suggested to be associated with still unresolved terrain features, in addition to the influences of input meteorological fields and the model physics on simulated atmospheric flows (e.g., Xue et al., 2014; Vincent and Hahmann, 2015).