The gridded EDGAR HTAP v2.2 air pollutant emission data combine the latest available regional information within a complete global data set. HTAP uses nationally reported emissions combined with regional inventories. The emission data are complemented with EDGAR v4.3 data for those regions with missing data. The global data set is a joint effort of the US Environmental Protection Agency (US-EPA), the MICS-Asia group, EMEP/TNO, the Regional Emission inventory in Asia (REAS) and the EDGAR group for scientific studies of hemispheric transport of air pollution. The EDGAR HTAP v2.2 data set provides emissions of CH4, CO, SO2, NOx, non-methane volatile organic compounds (NMVOCs), NH3, PM10, PM2.5, BC and organic carbon (OC) on a 0.1∘ × 0.1∘ grid for the years 2008 and 2010 with a monthly time resolution. In the region considered in this study, the emissions are based on data from REAS which have a resolution of 0.25∘ × 0.25∘.

In , a method is presented to estimate black carbon emission fluxes for the Kathmandu Valley from mixing layer height data, derived from ceilometer measurements, and black carbon concentrations measured during SusKat-ABC at the Bode station (number 0017) located within the valley (Table  and Fig. ). These estimated emission fluxes are based on measurement data from March 2013 to February 2014 and calculated for each month. The emission estimates are based on the assumptions that (i) black carbon aerosols are horizontally and vertically well mixed within the mixing layer, (ii) the variation of the mixing layer height is only small at night (as frequently observed in the ceilometer measurements used in the study), (iii) the vertical mixing between the mixing layer and the free atmosphere is small (consistent with a stable mixing layer height), and (iv) the horizontal transport of air pollutants into and out of the valley is small (consistent with low nocturnal wind speeds).

The use of these observationally based black carbon emission fluxes is motivated by the finding that the emission fluxes in the EDGAR HTAP inventory for the Kathmandu Valley are rather small compared to other big cities such as Delhi and Mumbai, where black carbon concentrations are measured that are similar to the black carbon measurements in the Kathmandu Valley. Table  summarizes the main differences between the two emission data sets for the Kathmandu Valley for February and May. In the WRFchem_BC simulation, these monthly means were used as black carbon emission fluxes for the grid cells representing the valley. For all other grid cells, the EDGAR HTAP emissions are used. For a more detailed description of the estimation of the black carbon emission fluxes, we refer to .

The SusKat project started with a 2-month long intensive measurement campaign (December 2012 to February 2013), which was extended until June 2013, providing detailed observations of a large number of chemical compounds and meteorological parameters. From December 2012 to June 2013, more than 40 scientists representing nine countries and 18 research groups deployed more than 160 measurement instruments for intensive ground-based monitoring at the urban supersite Bode and a network of 22 additional satellite and regional sites in the Kathmandu Valley and other parts of Nepal . SusKat-ABC was so far the second largest international air pollution measurement campaign conducted in south Asia, following the Indian Ocean Experiment during 1998 to 1999 . SusKat-ABC provides the most detailed air pollution data for the foothills of the central Himalayan region available to date. Hourly data of the following meteorological parameters are available: near-surface temperature, wind direction and speed, relative humidity and precipitation. Furthermore, data on the mixing layer height derived from ceilometer measurements are available . Black carbon measurements at the Bode site are used in this study for comparison with the WRF-Chem simulations. The black carbon concentrations were measured with a dual-spot aethalometer (Aethalometer AE33, Magee Scientific, USA) with a time resolution of 1 min. For the model evaluation, all data are used with a time resolution of 1 h calculated as means from the original data. In contrast to the densely built-up center of the Kathmandu Valley, the surroundings of the Bode site are characterized by a mixed residential and agricultural setting in a suburban location with only light traffic and scattered buildings.

Besides Bode, BC concentrations were also measured at Pakanajol, a site near the center of the Kathmandu metropolitan city about 9 km (aerial distance) to the northwest of the Bode site. The BC concentrations at both sites were found comparable in all seasons and have therefore not been used in this study.

The Department of Hydrology and Meteorology (DHM) of the Ministry of Population and Environment of the Government of Nepal hosts a network of meteorological stations. Data from five stations within this network were used in order to compare the meteorology simulated with WRF to observations. Hourly data of 2 m temperature and 10 m wind speed and direction were used (Table ).

ERA-Interim is a reanalysis data set compiled by the European Centre for Medium-Range Weather Forecasts . Zonal and meridional wind fields at 500 and 800 hPa are used for comparison with the modeled wind fields, as a general consistency check of the model results. As observations in this region are scarce, the reanalysis data for this region are expected to have larger uncertainties than in regions with a higher coverage of observations.

No radiosonde data are available for the Kathmandu Valley, but radiosonde data from the Integrated Global Radiosonde Archive (IGRA) at two locations (Table ) within the modeling domain D01 can be used for comparison with the model results . Both of these two radiosonde stations are located in northern India (Fig. ), and only one of the stations lies within the highly resolved model domain D02. For station 42182 (New Delhi/Safdarjung), observations are available at around 00:00 and 12:00 UTC between January and June 2013. As launch time of the radiosondes varied, observations for 00:00 UTC also include 23:00 and 01:00 UTC observations, and profiles for 12:00 UTC also include observations for 11:00, 13:00 and 14:00 UTC. In total, 174 profiles were available at around 12:00 UTC and 180 profiles were available at around 00:00 UTC. For station 42379 (Gorakhpur), observations are available only at around 00:00 UTC, which also includes observations at 01:00 and 02:00 UTC due to varying launch times. In total, 77 profiles were available. The processing of the radiosonde observations is further described in Sect. .

Tropical Rainfall Measuring Mission (TRMM)-based precipitation estimates are used to analyze the geographical distribution of the simulated precipitation fields . TRMM is a joint mission of NASA and the Japan Aerospace Exploration Agency (JAXA) to measure tropical rainfall for weather and climate research. The TRMM precipitation data are widely used and contributed to improving the understanding of, for instance, tropical cyclone structure and evolution, convective system properties, lightning–storm relationships, climate and weather modeling, and human impacts on rainfall. For the analysis in this study, daily precipitation rates with a spatial resolution of 0.25∘ × 0.25∘ were used (TRMM product 3B-42).

As a first assessment of the model's performance in reproducing the large-scale wind pattern, the model results are compared to the 500 and 800 hPa wind fields from the ERA-Interim reanalysis. It should be kept in mind that because of the sparsity of available observations in this region, the reanalysis data for this region are expected to have larger uncertainties than in better observed regions. The spatial distributions of the zonal and meridional wind components at 500 and 800 hPa from WRF and the ERA-Interim reanalysis averaged over February and May 2013 are shown in Figs.  and S1 (in the Supplement), respectively. The overall pattern of the zonal wind component is qualitatively similar in both data sets for February, with lower values over India in the model simulation. Differences of up to 5 ms-1 are found in the 500 hPa zonal wind component in February, south of the Himalayas, extending in the east–west direction throughout the whole model domain. At 800 hPa, the zonal wind component exhibits differences up to 5 ms-1 in the north of Bangladesh. In May, the zonal wind speed at 500 hPa simulated with the model is much lower compared to ERA-Interim data as shown by the domain-averaged mean bias of 2.9 ms-1. ERA-Interim shows here a stronger westerly wind component. At 800 hPa, the zonal wind is up to 20 ms-1 higher at the bottom of the Himalayas in regards to ERA-Interim data, while the meridional wind is over most of Indian territory up to 5 ms-1 lower than ERA-Interim. The spatial distribution of the meridional wind component simulated by the model is also qualitatively similar to the ERA-Interim fields in both months, with some difference in the southeast of domain D01 in February and over India in May 2013. The domain-averaged mean bias of the monthly mean meridional (zonal) wind fields is 0.1 ms-1 (2.2 ms-1) for February and 0.3 ms-1 (2.9 ms-1) for May, and the spatial correlations of the meridional and zonal wind distributions are 0.9/0.8 and 0.9/0.8 for February and May, respectively.

In order to evaluate the ability of the model to correctly represent the vertical structure of the atmosphere, measurements from radiosondes for temperature and relative humidity are compared to the model results (Figs.  and ). This comparison only provides a limited quality check of the model, since there is only a single radiosonde station available within D02. The comparison shows that WRF is able to capture the basic features of the vertical profiles of temperature and relative humidity with the modeled vertical profiles being within the variability estimated by the standard deviation (shaded areas), with the largest differences typically between about 900 and 700 hPa and near the surface.

The measured vertical profiles of temperature at station 42182 show an inversion layer that is also captured by the model. At 00:00 and 12:00 UTC, the observed mean near-surface temperatures are 15 and 24 ∘C; at 960 hPa, they are 22 and 31 ∘C, respectively. The modeled temperatures are about 3 ∘C higher at 00:00 UTC and about 1 ∘C lower at 12:00 UTC compared to the measurements. On average, the modeled mean temperature is overestimated by more than 1 ∘C over the entire column compared with the radiosonde data. The largest differences between model and measurements are up to 6.5 ∘C at 700 hPa (00:00 UTC) and up to 10 ∘C at 890 hPa (12:00 UTC). At 740 hPa, the model shows an underestimation of up to 3 ∘C. At station 42379, the modeled mean temperatures are overestimated by less than 2 ∘C compared with the observations. The standard deviation of the simulated temperature profiles in the lower half of the troposphere is typically around 6 ∘C, which is similar to observed one of around 7 ∘C.

The measured vertical relative humidity profiles are well reproduced by the model within the first five vertical layers with the model bias at each model level ranging between -4 and 4 %. Between 900 and 740 hPa, the model overestimates relative humidity by up to 20 %. As for temperature, the model reproduces the observed standard deviation of relative humidity quite closely at all heights investigated, with the standard deviation ranging from 16 % at surface up to 25 % at about 570 hPa.

The daily mean 2 m temperature increases during the simulation period at all stations shown in Fig. , from about 5–10 ∘C in January to 20–30 ∘C in June, which is also shown by the model (WRF_ref_D02). While the observed temporal evolution of the daily mean near-surface temperature is well reproduced by the model (correlation above 0.9; Fig. ), the absolute values are systematically over- or underestimated at several stations. The mean bias for WRF_ref_D02 ranges between -1.9 and 2.2 K (Fig. ). This MB is larger than the benchmark value of ±0.5 K from . Here, it should be noted that at several stations the over- or underestimation of measured temperature is associated with a difference between the actual elevation of the measurement station and the elevation of the model grid cell the station is located in. For example, at station 1206, the elevation of the grid cell in the domain D02 is 149 m lower than the elevation of the measurement station (1720 m); given a typical atmospheric vertical temperature gradient of 6–7 Kkm-1, one would expect a bias of about 1 K, which is close to the actual mean temperature bias of 0.8 K. In order to correct for the temperature biases caused by differences in elevation, a height correction has been applied to the model data by linearly interpolating the modeled vertical temperature profile to the elevation of the measurement station. For the stations, the mean bias was reduced by 1 K (station 0014) to 0.2 K (station 1206) (Fig. ) when considering this height correction. Table  summarizes the statistics averaged over all available stations and the whole simulated time period based on 3 h data. On average, the model overestimates the observed mean temperatures by 0.7 K, which is slightly larger than the benchmark value from but still suggesting that the model performance is acceptable given the challenging topography of the Himalayas. The mean daily minimum and maximum temperatures are overestimated by 1 K and underestimated by 0.5 K, respectively. The main features of the average diurnal cycle of the 2 m temperature (Fig. ) are reproduced by the model but the daily temperature amplitudes (difference between the daily minimum and maximum temperature) are often smaller in the model simulation than in the measurements. This is mainly caused by a high bias in the simulated values in the morning hours. In contrast, the daily variability of the 2 m temperature shown by the 25th and 75th percentiles in Fig.  is reproduced quite well by the model.

The temperature biases found at stations in the present study are in the same range as the ones found in other regions with WRF , particularly when considering that the reported 2 m temperature biases in these studies tend to be higher in mountainous terrain than in other regions. For example, found a mean bias in the 2 m temperature of -1.5 to 1 K at stations in east Asia, while at single stations the mean bias can range between -5 and +5 K in January and July 2005, respectively. found a good agreement between WRF-Chem simulations for south Africa and ERA-Interim reanalysis data 2 m temperature in 2010 (mean bias 0.4 and -0.03 K, spatial correlation 0.93 and 0.91, for September and December, respectively). found that the spatial variability in measured 2 m temperature is well reproduced by WRF-Chem in all seasons in 2007 over Europe with values of the absolute mean bias of generally less than 1 K. Both and found the largest biases in 2 m temperature in the Alps. describes an overprediction by more than 1 K in this region, whereas found a cold bias of -5 to -2 K.

The wind speed has an essential impact on the horizontal transport of pollutants. For example, low wind speeds favor an accumulation of pollutants close to their sources, whereas higher wind speeds lead to the transport of pollutants away from their source. The average measured wind speed over all stations and over the 6 months based on hourly data is 1.7 ms-1 (Table ), which is overestimated by the model by 1 ms-1. The averaged RMSE value over all stations of 2.2 ms-1 is close to the benchmark value of 2 ms-1 proposed by . In fact, at most individual stations, the RMSE values are within the benchmark range. At individual stations where wind speed data are available, the biases range between 0 and 1.7 ms-1. The temporal correlation coefficient of hourly wind speed is on average 0.4 with a range of 0.1 to 0.6 at these individual stations (Table ). The overestimation in wind speed in the WRF_ref_D02 simulation can probably be attributed to a large extent to an overestimation of the maximum wind speed during daytime, which is on average biased positively by 2 ms-1. In contrast, the daily minimum wind speed is close to the observation (MB of 0.2 ms-1) (Table ). This is also clearly seen in the frequency distributions of the wind speeds (Fig. S1), which typically have a much broader distribution with higher wind speeds and a maximum shifted to larger values for the model compared to the observations.

This performance of WRF in reproducing the observed mean 10 m wind speed is consistent with biases reported in the literature, especially when considering stations in mountain regions. For example, found an overestimation of the modeled wind speed over Europe, especially during winter and fall with a bias of 2 ms-1 and more. Regions with a larger bias include the mountain region of the Alps, indicating the challenges of simulating wind accurately over complex terrain. The temporal correlation of the modeled 10 m wind speed in Europe is typically above 0.7 but lower (0.4–0.6) over the Alps and close to the Mediterranean , which is still higher than the correlation found at some stations in this study. describe a significant overprediction at almost all sites investigated in Europe (MB of 2.1 ms-1) with the largest biases over several countries in low-lying coastal areas and over the Alps as well as the Carpathian Mountains. They argue that these results indicate the difficulty of the WRF model in simulating wind patterns and mesoscale circulation systems (such as sea breeze and bay breeze) and their interaction with land over complex terrain. Furthermore, they state that this high bias in 10 m wind speed can be mainly attributed to a poor representation of surface drag exerted by the unresolved topography in WRF. tested different planetary boundary layer (PBL) schemes in their model setup and also found an overestimation of wind speed at stations in California in all cases, although of different magnitude (about 0.5 to 3 ms-1). found a significant overprediction of 10 m wind speed at stations in east Asia with a mean bias of 1.9–3.1 ms-1.

An evaluation of the 10 m wind speed and especially the wind direction at the individual measurement stations (not shown) strongly suggests that these parameters are highly dependent on the stations' locations and the topography of their surroundings, especially in mountain areas. The measurements at some of these sites are therefore probably only representative of a rather small area around the station. Because of the complex topography in this region, a horizontal resolution of 3×3 km2 is too coarse to represent the near-surface wind at sites strongly influenced by small-scale features such as individual mountains. Therefore, the main focus of the evaluation of the 10 m wind is on the Kathmandu Valley. The Kathmandu Valley, with a diameter of about 30 km, is starting to be large enough to be resolved at the model resolution of 3×3 km2. The relatively flat valley floor further facilitates a comparison of the 3×3 km2 model grid cells with observational data, as measurements inside the valley are expected to be less influenced by small-scale topography than at most stations outside the valley.

The frequency distribution of wind speed per wind direction based on 3-hourly data for the whole simulation period is shown in Fig.  as wind roses for all available stations in the valley. The main wind directions in the east of the valley (station 1015) are north–northwest, east–southeast and south, with wind speeds of typically up to 6 ms-1. In contrast to the observations, the model shows wind directions from north–northwest to south–southeast. Wind speeds are similar as observed. The main wind direction at stations in the west of the valley (0014 and 0017) is less clearly dominated by particular sectors more than that in the east of the valley but rather characterized by predominately westerly winds. This pattern is reproduced by the model, although the wind speed is generally overestimated. The observed diurnal cycle of wind speed at the Bode station (Fig. a) shows very low median values between 0 and 1 ms-1 during the night and a maximum median wind speed during daytime of about 4 ms-1. As discussed before, the low wind speed during night is well reproduced by the model but the maximum wind speed during daytime is overestimated. The main wind direction during nighttime is from the east–southeast (around 100∘) in the observations (Fig. b), while it is from around 180∘ in the model. For such low wind speeds, however, the measured wind direction is expected to be affected by small-scale dynamics such as turbulence and thus not expected to be directly comparable to a 3×3 km2 model grid cell. In the transition phase from low to high wind speed during morning hours (09:00–11:00 LT) and from high to low wind speed in the evening (19:00–21:00 LT), the model does not reproduce the wind direction correctly. In contrast, the main wind direction during daytime is west–southwest (around 250∘) which is reasonably well reproduced by the model.

A key parameter for air quality is the depth of the mixing layer which is a part of the planetary boundary layer and characterized by a strong gradient in parameters such as potential temperature and aerosol concentration, and by an unstable layer and strong mixing due to turbulence during daytime and a rather stable layer during nighttime. Thus, the mixing layer has an important impact on the dispersion or accumulation of pollutants at the ground level. In the WRF model, the mixing layer height is a diagnostic variable which is calculated based on the Richardson number . The model output is compared to the values derived from ceilometer measurements obtained during SusKat-ABC . In Fig. , the diurnal cycle of the mixing layer height calculated from data covering the two time periods (January–February and March–June 2013) is shown for the model (WRF_ref_D02) in comparison with the ceilometer data. Both model and observations show a distinct diurnal cycle with low mixing layer heights during the night and morning hours and higher values during the day. While the lowest measured nocturnal values are around 160 m in January–February and around 200 m in March–June, the modeled values typically go down to less than 50 m in all seasons. The maximum mixing layer height values are measured at around 16:00 LT in the afternoon with a median of 1060 m in winter and 1053 m in the pre-monsoon season. The simulated values are higher during the day, with a median of 1132 m at 15:00 LT in winter and of 1512 m during the pre-monsoon season. These over- and underestimations of the maximum and minimum in the diurnal cycle are also shown for individual months, for instance, a high/low bias for the maximum/minimum mixing layer height of +244/-76 m in February and +280/-122 m in June. The mean bias in the modeled maximum/minimum diurnal mixing layer height is +81/-32 m in January–February and +438/-372 m in March–June.

A similar pattern was also found by for WRF-Chem simulations over Germany in summer, with a mean bias of -113 m for the daily minimum and 287 m for the daily maximum mixing layer height. This is also in agreement with , who showed an overestimation up to 204 m in February 2012, and up to 584 m in March 2012 of the modeled boundary layer height predicted at a measurement site in the central Himalayas.

Furthermore, the simulated diurnal cycle of the increase in mixing layer height during daytime is shifted by about 2 h to earlier times compared to the measurements. During the day, convection is an important process for determining the mixing layer height. A premature onset of convection found in many models is a long-standing issue and has been identified in numerous previous modeling studies, including studies with WRF e.g.,.

A good representation of the precipitation in the model is important for the calculation of wet deposition of air pollutants such as particulate matter including black carbon. The domain-averaged daily precipitation totals from the model (WRF_ref_D02) and TRMM are shown as a time series in Fig. . The near absence of strong rain events in the dry season (January through April) is reproduced well by the model, and also the timing of the single rain events between January and March is reproduced relatively well, although the total amount of precipitation is overestimated by the model. This overestimation is particularly strong during the dry season, as can also be seen for February in Fig. . Here, the strongest overestimation of precipitation by the model is seen in the northern part of Nepal and seems to be related to the particularly high and steep orography in this part of the country. The transition to and start of the rainy season in late April/early May as seen in the TRMM data are reproduced reasonably well by the WRF simulation. The geographic distribution of the observed average precipitation rates during the rainy season which is shown in Fig.  for May shows rain rates between 6 and 15 mmday-1 particularly over Nepal, and rain rates between 0 and 5 mmday-1 over northern India and the part of the Tibetan Plateau covered by model domain D02. This is qualitatively well reproduced by the model.

The statistics summarized in Table  represent the skill of the model (WRF_ref_D02) to reproduce precipitation events at a single station in the valley (Bode). It shows that 62/57 % (H) of the observed precipitation days are correctly captured by the model when using the Bode station measurements and the TRMM data, respectively, as reference data. The ratio of days when precipitation was present in the model data but not measured relative to all forecasted precipitation days (FAR) is relatively high: 32 % for the station measurements and 36 % for the TRMM data. Other than the hit rate, the CSI also considers false alarm and missed forecast, but it is not influenced by correctly forecast no-precipitation days. The CSI score indicates that 48 % of the forecast and observed precipitation days are correct. When using the TRMM data as observational reference, the score is a bit smaller (43 %). Hit rate and CSI are both lower for the model if considering TRMM as reference. Differences between the two observational data sets (station measurement and TRMM data) are shown in Table . The hit rate for the station measurements and the TRMM data (station measurement/TRMM) indicates that 71 % of the measured precipitation days at the Bode station are also visible in the TRMM data. The differences obtained when using the two different observational data sets also show the uncertainties and limitations particularly of the TRMM data for this kind of comparison. Since some of the precipitation events can be rather localized (e.g., convective rain) and can thus not be expected to be fully reproduced by a 3×3 km2 model simulation, they might also be missed in the rather coarse spatial and temporal (satellite overpass times) resolution satellite data.

In order to test that the simulated large-scale circulation does not drift or deviate from the observed synoptic condition, a sensitivity simulation in which a grid nudging technique was employed for horizontal winds, temperature and water vapor above boundary layer has been performed. In this simulation, we obtained similar results as in the reference simulation; for example, the RMSE of temperature is 3.0 K using the nudging approach compared to 3.1 K in the reference run. The model performance for wind speed does not change. In the upper troposphere, the differences in the simulated meteorological variables in the reference and the sensitivity runs were not statistically significant, suggesting that the WRF model results in this altitude range are mostly driven by the prescribed boundary conditions. In a second sensitivity simulation, we have analyzed the impact of using MODIS land use data instead of the default USGS data set. In this simulation, the impact of using the MODIS data together with applying the nudging technique on WRF results is tested for temperature and wind speed parameters. As in the first sensitivity simulation, the RMSE of temperature does not deviate much from the one obtained from the reference simulation, i.e., using USGS land use data and no nudging, leading to a RMSE of 2.9 K compared with 3.1 K in the WRF_ref_D02 simulation. In contrast to temperature, the model performance for wind speed worsens with a RMSE of 3.2 ms-1 and an average correlation coefficient of 0.21. Since the relatively small number of measurement stations in the evaluation domain might not be representative of the whole domain, we have also compared the results from the sensitivity simulations with the reference simulation. When applying the nudging technique, the domain-averaged mean bias between the sensitivity and the reference simulation is -0.03 K for temperature and 0.08 ms-1 for wind speed. For the MODIS land use sensitivity simulation, the domain-averaged mean bias when compared to the reference simulation is 0.08 K for temperature and 0.2 ms-1 for wind speed. This suggests that the changes in temperature and wind speed when applying the nudging technique and using the MODIS land use data set are rather small and not expected to be important factors in explaining the differences between the model results and observations found.

Two WRF-Chem simulations have been performed with an identical model configuration but using different black carbon emissions. The WRF-Chem reference simulation uses the EDGAR HTAP emissions (WRFchem_ref); the second simulation uses the same emission data but with black carbon emission fluxes over the Kathmandu Valley replaced by emission estimates based on SusKat-ABC measurements (WRFchem_BC) (see Sects.  and for details on the emission data sets). The black carbon emission fluxes used in both WRF-Chem simulations are shown in Fig. .

Monthly mean black carbon concentrations measured in the Kathmandu Valley at the Bode station are 27 µgm-3 in February 2013 and 11 µgm-3 in May 2013. These values are strongly underestimated in the reference simulation WRFchem_ref_D02 (using EDGAR HTAP emissions), which average only 3 µgm-3 (89 % underestimate) in February and 2 µgm-3 (82 % underestimate) in May. The WRF-Chem sensitivity simulation using the black carbon emission fluxes inside the Kathmandu Valley estimated from observations (WRFchem_BC_D02) shows significantly reduced biases, averaging 12.5 µgm-3 (54 % low bias) in February and 6 µgm-3 (45 % low bias) in May. These results from WRFchem_BC_D02 are in much better agreement with the measurements at the Bode site, even though black carbon is still underestimated by the model. The improvement of the simulated black carbon concentrations when using the observationally based estimated fluxes can also be seen in the time series of daily mean black carbon concentrations (Fig. ). Measured daily black carbon concentrations reach values of up to 35 µgm-3 in February and up to 28 µgm-3 in May, with a pronounced variability within the same month (e.g., 2–5 May vs. 6–8 May). The daily mean black carbon concentrations from the reference simulation WRFchem_ref_D02 are below 5 µgm-3 in both months. The differences between the two months as well as the large daily variability are not reproduced by the reference simulation. In contrast, the time series of the WRFchem_BC_D02 sensitivity simulation shows values of up to 20 µgm-3 in February and up to 8 µgm-3 in May. In addition, the observed differences between February and May as well as the daily variability are better reproduced than in the reference simulation WRFchem_ref_D02. In order to compare the spatial variability of the simulated black carbon concentration in the valley, also the daily mean concentrations simulated in the grid cells with the highest and lowest values of all neighboring grid cells of the Bode grid cell are shown in Fig. . The spatial variability of the simulated black carbon concentration is higher (in absolute and in relative terms) in the WRFchem_BC_D02 simulation compared to WRFchem_ref_D02. This figure also shows that the grid cell with the Bode station is not an outlier but generally at the upper end of the range of minimum and maximum concentrations of its neighbors.

The histogram of the measured hourly black carbon concentrations (Fig. ) shows values of up to 90 µgm-3 and a maximum of the distribution between 0 and 10 µgm-3. These values of the measured frequency distribution are not reproduced by the reference simulation WRFchem_ref_D02, in which the black carbon concentrations range only between 0 and 6 µgm-3 with a maximum frequency between 1 and 1.5 µgm-3. The histograms of the WRFchem_BC_D02 simulation for February and May show a wider frequency distribution compared to the reference simulation WRFchem_ref_D02 with maximum concentrations of up to 40 and 20 µgm-3, and maximum frequencies in the interval 0 to 10 µgm-3 and around 5 µgm-3 (in February and May, respectively).

The pollution roses in Fig.  show the measured and simulated black carbon concentrations coinciding with each specific wind direction at the Bode station and the frequency of the occurrence of the corresponding wind direction in percent. The figure shows that the observed main wind direction in February is from the west and west–southwest, but high black carbon concentrations are found for all wind directions. Simulated main wind directions span a wider range than in the observations (west–northwest, southwest and south) but the model reproduces the observation that high black carbon concentrations are found independent of the actual wind direction. In May, the observed main wind direction is from the west (and slightly north and south of west), and the highest concentrations are measured for winds from the north and east–southeast (Fig. d). Again, the model does not fully reproduce the main wind directions (here northwest to south) and underestimates black carbon concentrations in all wind directions.

These findings strongly suggest that the EDGAR HTAP emissions of black carbon in the valley are underestimated and that there is a need for further improvements of the local emissions in the Kathmandu Valley. Despite this improvement in the simulated black carbon concentrations in the Kathmandu Valley when using the black carbon emission fluxes estimated from observations, the measured concentrations are still significantly underestimated by the model.

Two possible reasons for the abovementioned underestimation of the observed black carbon concentrations in the WRFchem_BC_D02 simulation are an overestimation of the dispersion of the black carbon aerosols away from the ground and too small observation-based black carbon emission estimates. Even though the model tends to overestimate the observed near-surface wind speed, the model bias of about 1 ms-1 is not expected to be large enough to explain the large differences in simulated and observed black carbon concentrations through an overestimated horizontal dispersion. The observed and simulated mixing layer heights (Fig. ) are quite similar, suggesting that the model is able to produce a reasonable vertical dispersion. Furthermore, particularly at nighttime, the smaller-than-observed simulated mixing layer height would rather lead to an overestimation of the observed black carbon concentrations by the model. This suggests that biases in the modeled dispersion (horizontal and vertical) alone are unlikely to be able to explain the large differences in modeled and observed black carbon levels. This, in turn, suggests that the top-down emissions determined by based on the observed black carbon concentrations and mixing layer heights might be underestimated – despite the fact that they are several times as high as the values in the state-of-the-art EDGAR HTAP v2.2 data set (for further discussion of the observation-based emission estimates for BC in comparison with available emission inventories, we refer to ).

There are various possible reasons why the top-down emissions derived from measurements at the Bode station might be underestimated or not fully representative of the entire Kathmandu Valley as assumed in the sensitivity study WRFchem_BC. One main reason is that the Bode station is not located in the urban center. Thus, throughout most of the year, during the months when the brick kilns near Bode are not operating, several important urban emission sources such as traffic, cooking and open burning of trash might be underestimated due to applying the top-down method to determine the black carbon emission flux based on the semi-urban Bode site data. Future development of high-resolution (e.g., 1×1 km2) emission data sets (Sadavarte et al., 2018) may help to resolve this possible discrepancy. An important first step for such work is the measurements of various fuel-based emission factors obtained during the Nepal Ambient Monitoring and Source Testing Experiment (NAMaSTE, ).

The other main possible reason for the top-down emissions to underestimate actual emissions is that the method currently only considers sources that are active at night, when the mixing layer height is stable and the increase in black carbon concentrations can be directly attributed to emissions during that time period. It is assumed that the average emissions during the rest of the day are the same as during this period. This can lead to either an over- or underestimation, depending especially on the extent to which the morning food preparation and rush hour traffic occur during the period of the stable nocturnal boundary layer. It is possible that the contribution of black carbon sources which are mainly active during daytime, after the nocturnal boundary layer begins to break up, exceeds the nighttime emissions. Since the daytime-specific emissions such as rush hours throughout much of the year and the generally heavier daytime traffic are not taken into account by the top-down computation, this could lead to an underestimation in the black carbon emission fluxes. This is consistent with the statement by that the top-down emission estimate is “likely a lower bound” and thus strongly supports the indication of an underestimation of the values in current emission data sets. Unfortunately, no technique has yet been found to apply the top-down method for the full diurnal cycle in the situation of the Kathmandu Valley, so it will be left to emission inventory developers to improve their estimates based on updated emission factors and activity data for the region, in order to hopefully determine what is missing according to the top-down analysis.

Despite that offset that is apparently due to the emissions, the temporal correlation coefficient between daily data of the WRFchem_BC_D02 results and the Bode observations is relatively high (0.7) in February, while it is much lower (0.2) in May 2013. There are likely two factors that contribute to this difference. Firstly, in May, the day-to-day variability of the emission strength from different sources can expected to be higher because brick kilns, which emit black carbon relatively constantly throughout the day and night, are no longer running, and emission sources with a much clearer diurnal cycle like cooking, traffic and trash burning take on a greater relative importance. Secondly, the meteorology in May is more difficult to simulate than that in February as convective precipitation becomes more frequent. The correct simulation of the occurrence of daily precipitation events is particularly important in this context. Although the transition from the dry season in winter to the wet season in summer is captured well by the model, there are several days when precipitation was observed and not simulated in the model and the other way around (Table ), which has an important impact on the simulated day-to-day variability of black carbon. In addition to particles being removed by wet deposition, also certain emission sources such as burning of trash and biomass, can be affected by precipitation.