Air Quality Modeling with WRF-Chem v 3 . 5 in East Asia : Sensitivity to Emissions and Evaluation of Simulated Air Quality

We conducted simulations using the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) version 3.5 to study air quality in East Asia at a spatial resolution of 20 km × 20 km. We find large discrepancies between two existing emissions inventories: the Regional Emission Inventory in Asia version 2 (REAS) and the Emissions Database for Global Atmospheric Research version 4.2 (EDGAR) at the provincial level in China, with maximum differences of up 5 to 500 % for CO emissions, 190 % for NO, and 160 % for primary PM10. Such differences in the magnitude and the spatial distribution of emissions for various species lead to 40–70 % difference in surface PM10 concentrations, 16–20 % in surface O3 mixing ratios, and over 100 % in SO2 and NO2 mixing ratios in the polluted areas of China. WRF-Chem is sensitive to emissions, with the REAS-based simulation reproducing observed concentrations and mixing ratios better than the 10 EDGAR-based simulation for July 2007. We conduct additional model simulations using REAS emissions for January, April, July, and October of 2007 and evaluate simulations with available ground-level observations. The model results illustrate clear regional variations in the seasonal cycle of surface PM10 and O3 over East Asia. The model meets the air quality model performance criteria for both PM10 (mean fractional bias, MFB 6 ±60 %) and O3 (MFB 6 ±15 %) in most of the 15 observation sites, although the model underestimates PM10 over Northeast China in January. The model predicts the observed SO2 well at sites in Japan, while it tends to overestimate SO2 in China in July and October. The model underestimates observed NO2 in all four months. Our study highlights the importance of constraining emissions at the provincial level for regional air quality modeling over


Introduction
Many Asian countries have faced deteriorating air quality since the late 1990s and early 2000s due to rapid economic development and population growth.According to the latest World Health Organization (WHO) ambient air pollution database (WHO, 2014), air quality in China and India was ranked 14th and 9th, respectively, out of the 91 most polluted countries.Since these countries have the largest population in the world, exposure to air pollutants poses health risks to billions of residents.For example, Chen et al. (2013) reported that outdoor air pollution in China alone caused approximately half a million premature deaths every year.A similar number of premature deaths was estimated in India in 2010 (HEI, 2013).Air pollution not only impacts human health, but also has important potential consequences for natural ecosystems, crop yields, visibility, and radiative forcing (Seinfeld and Pandis, 2006).In order to mitigate these negative consequences, it is essential to have a better understanding of air pollutant emissions sources and magnitudes, as well as atmospheric transport and chemical composition over the region.
Several modeling studies have applied the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) (Grell et al., 2005) to study air quality in Asia.Saikawa et al. (2011) analyzed the impact of China's vehicle emissions on air quality both within China and across East Asia.They found that stricter regulation of the road transport sector in China would reduce surface concentrations of fine particulate matter with an aerodynamic diameter of 2.5 µm or less (PM 2.5 ) and tropospheric ozone (O 3 ) mixing ratios in the region.Kumar et al. (2012) examined ground-level measurements and satellite observations in South Asia and reported that WRF-Chem could simulate O 3 and CO well, but large discrepancies were found for NO 2 due to uncertainties in emissions from biomass burning and anthropogenic NO x estimates.Wang et al. (2010) conducted sensitivity analyses of O 3 , NO x , and sulfur dioxide (SO 2 ) mixing ratios to temporal and vertical emissions; their results showed that air quality in East Asia was impacted by the diurnal and vertical distribution of anthropogenic emissions.Studies that have conducted WRF-Chem modeling for PM 2.5 and PM 10 have found that these surface concentrations were usually underestimated.For example, Saikawa et al. (2011) reported that modeled 4-month average PM 2.5 concentrations at Oki and Rishiri in Japan had a mean normalized bias (MNB) of −34 % compared to observations.Gao et al. (2014) compared simulated and measured PM 10 concentrations at six sites in Japan and found that the model underestimated the annual average PM 10 at all sites except one.
One of the possible reasons that models underestimate particulate matter (PM) concentrations is the uncertainty in emissions.Several emissions inventories for Asia have been developed by different groups, each with different purposes and characteristics (Kurokawa et al., 2013;JRC and PBL, 2010;Streets et al., 2003;Klimont et al., 2011).Comparison of the emissions inventories has revealed large differences in these emissions estimates.Kurokawa et al. (2013) compared different emissions inventories for several provinces in China and found that the difference in primary organic carbon emissions can be as high as 140 %.The possible causes of such discrepancies among emissions inventories are differences in estimates of (1) activity level, (2) level of technologies implemented, and (3) emissions factors.Since it is hard to measure emissions factors of each individual source at the scale of a province or a country, uncertainties arise when emissions factors from one place are applied to another.Activity data or emissions factors are often not available at the level of detail required for making insightful comparisons across emissions inventories.
While comparison of emissions inventories has revealed notable differences in the emissions estimates, few studies have addressed to what extent uncertainty in the emissions inventories really matters for the outcome of air quality modeling studies.Ma and van Aardenne (2004) compared simulated surface O 3 mixing ratios over China using three different emissions inventories as model inputs, and found that surface O 3 differed as much as 30-50 % among different model simulations.They also demonstrated that the differences in NO x and non-methane volatile organic compounds (NMVOCs) among different inventories were dominant factors for the discrepancies in simulated O 3 mixing ratios.Amnuaylojaroen et al. (2014), on the other hand, studied the effect of different anthropogenic emissions inventories on air quality over Southeast Asia and found only a small difference in simulated O 3 (about 4.5 %) and CO (about 8 %) mixing ratios.However, these studies did not investigate the impact of emissions inventories on other pollutant species such as PM.Unlike the previous studies, which focused on uncertainties of simulated O 3 , CO, and NO x , this study provides quantitative information on how emissions inventories impact PM and other pollutants including SO 2 .
The first objective of this paper is to study the sensitivity of regional air quality to emissions.We select two commonly used anthropogenic emissions inventories for comparison: the Regional Emission Inventory in ASia version 2 (REAS) (Kurokawa et al., 2013) and the Emissions Database for Global Atmospheric Research version 4.2 (EDGAR) (JRC and PBL, 2010).By comparing the 2-week model simulations using these two emissions inventories and observations from July 2007, we select the REAS inventory to perform air quality simulations over East Asia in different seasons.The second objective is to evaluate the simulated PM 10 concentrations, as well as O 3 , SO 2 , and NO x mixing ratios from four 1-month WRF-Chem runs against ground-level observations to build confidence in its ability to simulate future air quality over this region.WRF-Chem is an online-coupled meteorology and chemistry model, simulating meteorological quantities and air pollution concentrations simultaneously and allowing two-way interactions between meteorological and chemical constituents.In regions with high PM loading, meteorology-chemistry interaction significantly improves model performance in simulating air pollutant concentrations (Kong et al., 2015).So far, many of the WRF-Chem studies that focused on China conducted limited model evaluation due to the scarcity of observations in the region.This study compares the model simulations to observations from more than 70 sites in China to evaluate the model.There are some studies that have compared simulation results using a different chemical transport model (i.e., the Community Multi-scale Air Quality Model), but as far as we are aware, few studies have used as extensive a network of PM 10 observations for WRF-Chem validation in this region as ours has.
This paper is organized as follows.Section 2 explains the regional air quality model (WRF-Chem) configuration, emissions used for the model, observations used for validation, and data analysis methods.Section 3 analyzes the differences in emissions inventories and the sensitivity of simulated pollutant concentrations to the inventory used.Section 4 evaluates model performance by comparing observations with model simulations.Section 5 presents a summary of results and suggestions for future research.

Model and observations description 2.1 Model description
We use the fully coupled "online" regional chemical transport model WRF-Chem version 3.5 (Grell et al., 2005) in this study.The Regional Acid Deposition Model version 2 (RADM2) atmospheric chemical mechanism (Stockwell et al., 1990) is used for gas-phase chemistry.Aerosol chemistry is represented by the Model Aerosol Dynamics for Europe with the Secondary Organic Aerosol Model (MADE/SORGAM) (Schell et al., 2001;Ackermann et al., 1998) with some aqueous reactions.This aerosol mechanism is widely used in regional atmospheric chemistry models (Saikawa et al., 2011;Gao et al., 2014;Tuccella et al., 2012;Kumar et al., 2012).It predicts the mass of seven aerosol species (sulfate, ammonium, nitrate, sea salt, black carbon (BC), organic carbon (OC), and secondary organic aerosols), using three log-normal aerosol modes (Aitken, accumulation, and coarse).Aerosol dry deposition is simulated following the approach of Binkowski and Shankar (1995) and the wet removal approach follows Easter et al. (2004) and Chapman et al. (2009).Photolysis rates are obtained from the Fast-J photolysis scheme (Wild et al., 2000).We include the aerosol-radiative feedback in our simulation.The rapid radiative transfer model (RRTM) scheme (Mlawer et al., 1997) is used to represent both shortwave and longwave radiation.The horizontal winds, temperature, and moisture are nudged to 2007 meteorological fields at all vertical levels.The 2007 meteorological data are obtained from the National Center for Environmental Prediction (NCEP) Global Forecast System final gridded analysis data sets.We use the Lin et al. (1983) microphysics scheme and the Grell-3d ensemble cumulus parameterization (Grell and Dévényi, 2002).
The model domain, shown in Fig. 1, covers most of the East and South Asia region with 398 × 298 grid cells, using a 20 km spacing and a Lambert conformal map projection centered on China at 32 • N, 100 • E. There are 31 vertical levels from the surface to 50 mb.The initial and lateral boundary conditions are taken from a time-slice simulation of Geophysical Fluid Dynamics Laboratory (GFDL) coupled chemistry-climate model AM3 (Donner et al., 2011;Naik et al., 2013) for the year 2010 following the configuration described by Naik et al. (2013).This AM3 simulation was driven by climatological mean sea surface temperature and sea ice distributions for the 2006-2015 time period derived from the transient GFDL coupled model (GFDL-CM3) simulations following Representative Concentration Pathway 8.5 (RCP8.5)(John et al., 2012).Concentrations of wellmixed greenhouse gases and ozone-depleting substances, and emissions of short-lived pollutants (ozone precursors and aerosols) were set to year 2010 values in RCP8.5.We simulate air pollutant concentrations for the central month of each season (January, April, July, and October) in 2007, to assess seasonal variability in air quality.The model is spun up for 7 days before the beginning of each monthly simulation.This is sufficient to ventilate our regional domain.

Emissions
The anthropogenic emissions of gaseous pollutants (CO, NO x , NH 3 , SO 2 , and NMVOCs) and particulate matter (BC, OC, PM 2.5 , and PM 10 ) are taken from REAS (Kurokawa et al., 2013).REAS covers most of the model domain (see Fig. 1, regions in blue).For the areas of our domain that are not covered by the REAS emissions inventory, we use the RCP8.5 emissions data set for the year 2010 (Riahi et al., 2011).The RCP8.5 emissions data set has been used in many studies for air quality simulations (Gao et al., 2013;Colette et al., 2013;Fry et al., 2012).For emissions from biomass burning, we use the year 2007 from the Global Fire Emissions Database version 3 (GFED) (Randerson et al., 2013).For biogenic emissions of CO, NO x , and NMVOCs, as well as aircraft emissions of CO, NO x , and SO 2 , we use the Pre- cursors of Ozone and their Effect on the Troposphere version 1 (POET) emissions inventory (Granier et al., 2005).Dust and sea salt emissions are calculated online using the dust transport model (Shaw et al., 2008) and sea salt (Gong, 2003) schemes, respectively.
To study the influence of anthropogenic emissions inventories on air quality simulation, we conducted a sensitivity simulation using the EDGAR (European Commission Joint Research Centre, 2010) inventory, as described in Sect.3. EDGAR does not provide BC, OC, and PM 2.5 emissions, and thus this study only compares simulated O 3 and PM 10 .NMVOCs in EDGAR are also not speciated, so we di-vided them into 17 chemical species, using weighting factors calculated from REAS.The total anthropogenic emissions of each air pollutant within the model domain as estimated in REAS and EDGAR for July 2007 are listed in Table 1.We apply the same diurnal variation to both REAS and EDGAR based on the East Asian Air Pollutant Emission Grid Database (EAGrid2000, http://www.cger.nies.go.jp/db/eagrid/eagrid_index_e.html).The ratios used to create hourly emissions for different pollutants are presented in Table S1 in the Supplement.The REAS emissions inventory provides monthly emissions for each pollutant, while the EDGAR emissions inventory provides only yearly emissions estimates.

Observations
The surface concentrations of PM 10 in China are derived from the Air Pollution Index (API) from the website of the Ministry of Environmental Protection of the People's Republic of China.When PM 10 is reported as the primary pollutant with a maximum pollutant index, daily PM 10 concentrations are calculated from the API, using the following equation: where C is the daily concentration of PM 10 , I is the API reported, I low and I high are the lower and upper API breakpoints that I falls within, and C low and C high are the PM 10 concentrations corresponding to I low and I high .Values of I low , I high , C low , and C high are described for different API levels, as shown in Table S2.Qu et al. (2010) have shown that API-derived PM 10 concentrations are generally comparable to those from filter sampling, although the latter tends to be approximately 10 % higher than API-derived PM 10 .As mentioned earlier, the derived concentrations from API have been used for the evaluation of a different chemical transport model in previous studies (Wang et al., 2009;Liu et al., 2010).
The observed PM 10 concentrations in Nepal are obtained from the Godavari station, located at the southern edge of the Kathmandu Valley (Ramanathan et al., 2007;Stone et al., 2010).We are unable to evaluate PM 2.5 against measurements for 2007 since PM

Data analysis method
We assess the model performance using the correlation coefficient (r), the normalized mean bias (NMB), the mean fractional bias (MFB), the mean fractional error (MFE), and the normalized mean square error (NMSE) between the observed (Obs) and modeled (Model) concentrations.The performance evaluation is based on monthly and yearly statistics using the daily mean values at each site, each region, and all sites.Following Boylan and Russell (2006), we set the performance goals of PM 10 as MFB less than or equal to ±30 % and MFE less than or equal to 50 %.The performance criteria of PM 10 are MFB ±60 % and MFE 75 %.
For O 3 , we use the performance benchmark MFB ±15 % and MFE 35 %, as recommended by Morris et al. (2005).

Sensitivity to emissions
To better understand the effect that anthropogenic emissions have on regional air quality simulations, we conducted two simulations in which REAS and EDGAR are used as separate inputs.In the following sections, we compare the major pol-lutant emissions estimated in REAS and EDGAR, followed by comparisons of resulting air quality simulations.

Emission comparisons
Table 1 summarizes the total emissions of major air pollutants over the model domain in July 2007 for air pollutant precursors.Both REAS and EDGAR estimate similar total SO 2 emissions of 4.62 Tg month −1 .We note that this similarity is purely coincidental and depends on the domain.In certain parts of the domain the REAS estimate is higher than that of EDGAR, while the opposite is true for other parts of the domain.When averaged over the whole domain, both inventories produce similar estimates (Fig. 2).We, however, find large discrepancies between REAS and EDGAR estimated emissions for total NH 3 (53 %) and NO x (27 %).For CO, NH 3 , and NO x , REAS estimates are higher than those of EDGAR, while for PM 10 and NMVOCs, the opposite is the case.Figure 2 illustrates the difference in the spatial distribution and magnitude of emissions between REAS and EDGAR for PM 10 , CO, SO 2 , and NO x in our model domain.
Although the total emissions within the domain for many of the species are comparable between the two inventories, the national and regional differences are large.REAS estimates are uniformly higher than those of EDGAR in northern, eastern, and southern China for all four species and in most parts of India for NO x and CO.For PM 10 and CO, EDGAR estimates are higher in most areas of South and Southeast Asia, as well as in Japan and South Korea.Table S4 compares the differences in provincial emissions between REAS and EDGAR in China.For example, we find that REAS estimates 150 % higher PM 10 and 548 % higher CO emissions than EDGAR in Hebei province.

Simulation comparisons
For the convenience of discussion, we refer to the simulation with REAS emissions as WRF-Chem-REAS and the simulation using EDGAR emissions as WRF-Chem-EDGAR.Figure 3 illustrates the differences in the 14-day mean PM 10 , O 3 , SO 2 , and NO 2 simulated from 1 to 14 July 2007.The difference is presented as the percentage difference in concentrations or mixing ratios relative to those simulated in WRF-Chem-EDGAR.The pattern of the difference for these species is similar to that of the emissions difference.WRF-Chem-REAS simulates 40-70 % higher surface PM 10 in most areas of the North China Plain (Beijing, Tianjin, Hebei, Henan, Shandong provinces).This difference, around 35 µg m −3 or higher, is comparable to the PM 10 levels in many sites in Japan (Table 3).The highest difference (70 %) occurs in Shandong province and the lowest difference (less than ±5 %) is found in southwestern and northwestern China (Table S4).WRF-Chem-EDGAR simulates higher PM 10 than WRF-Chem-REAS around Cambodia, Vietnam, and Thailand.For surface O 3 , a moderate difference of 16-20 % (approximately 12-16 ppbv) is found over the North China Plain, the Yangtze River Delta, central China, and eastern Pakistan.WRF-Chem-REAS also results in higher SO 2 and NO 2 (more than 10 ppbv) in these areas than WRF-Chem-EDGAR.The largest discrepancies, over 100 %, occur in Table 2 summarizes the statistical measures of model simulations using these two anthropogenic emissions inventories against observations.Both simulations reproduce the temporal variation of O 3 , SO 2 , and NO 2 well, with the value of r between 0.64 and 0.83.The temporal correlation of PM 10 for WRF-Chem-REAS (r = 0.38) is higher than that calculated for WRF-Chem-EDGAR (r = 0.2).In terms of bias, both simulations produce a similar NMB and MFB for O 3 .For PM 10 , NO 2 , and SO 2 , WRF-Chem-REAS has a smaller MFB than WRF-Chem-EDGAR.In terms of error, MFE and NMSE from the two simulations are comparable for O 3 , but WRF-Chem-REAS results in less MFE and NMSE for PM 10 and NO 2 .According to the model performance goals and criteria of PM 10 suggested by Boylan and Russell (2006), WRF-Chem-EDGAR meets the performance criteria, while WRF-Chem-REAS achieves the stricter performance goals.We have conducted additional sensitivity simulations using REAS and EDGAR in January and July and compared the simulated air pollutants and observation.The results of these 2-month simulations (not shown here) agree with what we find here.
Based on the above performance analyses, we choose REAS as the anthropogenic emissions inventory to conduct further simulations for 4 months to explore the seasonality of air pollutant concentrations.In this paper, we focus on validating the WRF-Chem model with REAS.

Spatiotemporal variations of pollutants and model evaluation
In this section, we analyze the spatial variability of simulated and observed monthly mean PM 10 concentrations, as well as O 3 , SO 2 , and NO x mixing ratios (Figs. 4, 7, 9, and 10).A color-filled circle overlaid on a model-simulated monthly average surface concentration map represents the observed monthly average value at each site.Tables 3-6 describe yearly statistics for PM 10 concentrations, as well as O 3 , SO 2 , and NO 2 mixing ratios at individual stations, respectively.Before analyzing the model performance in simulating air pollutants, we evaluate the simulated meteorological fields, including daily mean 2 m temperature, 2 m relative humidity, and 10 m wind speed against observations from the National Climate Data Center of China Meteorological Administra-tion for the year 2007 (Table S5).The model reproduces 2 m temperature with a correlation of 0.97 and a negative NMB of −14.57%.Relative humidity is simulated with a correlation of 0.71 and a positive NMB of 7.02 %.Compared to temperature and relative humidity, the 10 m wind speed has a relatively lower correlation of 0.52 and a higher positive NMB of 59.35 %.Overall, the model performance in simulating these meteorological data is similar to that reported for regional air quality models (Tuccella et al., 2012;Tessum et al., 2015;Zhang et al., 2015).

PM 10
We obtain ground-level measurements from 1 site in Nepal, 7 sites in Japan, and 71 sites in China.China is divided into seven geographical regions and measurements are analyzed based on these regions (Table 3).The coverage of each geographical region in China is shown in Fig. S1.In China, the highest 4-month average PM 10 is observed in northwestern China (126 ± 94 µg m −3 ), followed by northeastern (119 ± 65 µg m −3 ) and central China (117 ± 48 µg m −3 ), while the lowest observed PM 10 is in southern China (82 ± 28 µg m −3 ).In Japan, the observed 4-month average PM 10 concentration is 27 ± 33 µg m −3 , which is more than 3 times lower than those observed in China.
The model simulates high PM 10 concentrations (over 200 µg m −3 ) near the Gobi Desert in northwestern China and in the border area near Iran, Afghanistan, and Pakistan (Fig. 4).In these areas, dust emissions are the predominant source of PM 10 and the anthropogenic primary PM 10 is negligible, as shown in Fig. S2.Besides these areas, the model simulates high PM 10 concentrations (up to 100 µg m −3 ) over the North China Plain, the Yangtze River Delta region, and the Sichuan Basin.The model simulates relatively low PM 10 concentrations (lower than 60 µg m −3 ) in most of southern, southwestern, and northeastern China, most of India, and other countries in the model domain.Unlike northwestern China, where the maximum PM 10 concentrations are simulated in spring, other regions of China are simulated to have high concentrations in January and October, with low concentrations in April and July.This is because in winter, reduced precipitation leads to higher PM 10 concentrations, while the monsoon circulation brings in clean marine air and dilutes the PM 10 surface concentrations in eastern China in summer.Moreover, aerosols in summer are removed by wet scavenging due to more frequent precipitation (Zhao et al., 2010).High concentrations are also simulated in an area surrounding Lhasa in Tibet in January.Since primary anthropogenic emissions in Tibet are low, dust emissions from local soils on the plateau are the main reason for high PM 10 concentrations.The previous study of tracer element analyses has shown that local dust is the major source of total particulate matter (PM) over Tibet (Zhang et al., 2001).
For 4-month averaged PM 10 , the model meets the performance criteria at 84 % of observation sites in China.The model tends to underestimate observations at the rest of the sites, which are mainly located in northeastern and southwestern China.Analyzing model-observation comparison by region, we find better model performance in central, eastern, northern, and southern China (Table 3).However, north- eastern and southwestern China have a higher correlation (r > 0.35) than others.For sites outside of China, the model underestimates observations in both Japan (MFB = −32 %) and Nepal (MFB = −48 %).
The seasonal statistics (Table S6) and Figs.5-6 indicate that the model meets the performance criteria in all 4 months (January, April, July, and October) in central, eastern, northern, and southern China.In the remaining regions in China and Japan, the model meets or is close to the criteria in April, July, and October, but has more difficulty reproducing PM 10 concentrations in January.Previous research has suggested that poor model performance in winter is common among air quality models and may be caused by difficulty in simulating stagnant weather conditions that lead to high winter PM concentrations (Tessum et al., 2015).In Nepal, model performance in both January and April is poor when the observed PM 10 is high.The time series comparison plots (Fig. S3) reveal distinct air pollution episodes occurring in middle January and early April at the Godavari site, which the model fails to simulate.One of the possible reasons for this is that the model is unable to reproduce the local meteorology due to the complicated topography that is not well resolved at  the current horizontal resolution.The temporal correlations of all sites in each month are similar (0.37-0.39) as shown in Table S7 and we do not observe obvious trends of temporal correlation change with seasons.

O 3
Similar to PM 10 , the simulated O 3 over the model domain also exhibits a seasonal variability that varies by region.Figure 7 illustrates that the highest O 3 mixing ratio (over 70 ppbv) occurs in northern and eastern China in July.This is because biogenic NMVOC emissions are relatively high and active photochemical reactions constitute favorable conditions for the build-up of O 3 mixing ratios in summer.On the other hand, a low monthly mean mixing ratio (below 40 ppbv) is found in the same region in January.In the Tibetan Plateau, the surface O 3 mixing ratio reaches a maximum (over 70 ppbv) in April due to high elevations and downward transport of O 3 from the stratosphere, while the minimum O 3 (40 ppbv) is found in July because the upward transport of air to the stratosphere in the summer suppresses the downward transport of O 3 (Gettelman et al., 2004;Randel et al., 2010).This simulated seasonal variability of O 3 in our model over the Tibetan Plateau is consistent with the findings of Ma et al. (2014).
The model performs well for simulating O 3 at all sites in Japan, and both MFB and MFE of these sites are within or close to the model benchmark (MFB < ± 15 % and MFE < 35 %).The model overestimates O 3 at Lulin in Taiwan.MFB at Lulin (41 %) is more than 2 times higher than that of any sites in Japan.Statistical analysis of O 3 in different seasons at the Lulin site (Table S8) reveals that such high bias is mainly caused by overestimation in October (MFB = 63 %).A previous study by Ou Yang et al. (2012) suggested that Lulin has more pronounced mountain valley circulation in fall, which leads to low observed O 3 mixing ratios in October.Our model with a horizontal resolution of 20 km × 20 km may not be able to capture such local meteorology.The model reproduces the overall daily temporal variation of O 3 well (r = 0.57) and the value of temporal correlation is also high for each site (0.47-0.93), except at Rishiri.This is partly due to the lateral boundary conditions, since this site is located close to the northeastern boundary of the model domain.The model predicts the seasonal variability well, as shown in Fig. 8 and Table S7.The modeled and observed monthly mean O 3 has a maximum in April and a min-imum in July.The same seasonal characteristics of O 3 levels were reported before (Yamaji et al., 2006).The MFB and MFE of all sites in each month are in the acceptable range.Among the 4 months, the model tends to underestimate the highest observations in April, while it overestimates observations in the other 3 months.

SO 2 and NO 2
Figure 9 illustrates that the model simulates a high monthly mean SO 2 mixing ratio (higher than 20 ppbv) over urban areas in northern China (including Beijing, Tianjin, Hebei, and Shanxi) and some provinces in eastern China (including Shandong and Henan), where emissions are also the highest.In these areas, the mixing ratios are the highest in January, followed by October, April, and July (Fig. 9).The lowest mixing ratios in our model simulation are found in July due to more active oxidation of SO 2 by hydroxyl radical (OH) and O 3 in the gas phase, as well as frequent precipitation that favors aqueous-phase oxidation of SO 2 (Feichter et al., 1996).Overall, the model predicts SO 2 well, with a MFB of 9 % and r of 0.64.The model performs better in predicting observed SO 2 mixing ratios at sites in Japan (MFB = −12 to 29 %, r = 0.52-0.82)than in China (MFB = −70 to 63 %, r = 0.14-0.5).The lowest overall MFB value of all sites occurs in April (8 %), while the highest happens in July (31 %).Although MFB values are acceptable, both the MFE and NMSE in July and October are high.The site that contributes most to high errors is Beijing, with a MFE of more than 115 % in these 2 months.The model largely overestimates SO 2 in Beijing (Fig. 11), probably because the REAS emissions inventory did not take into account the local emissions control policies for the Beijing Olympics.In 2007, the Chinese government reduced anthropogenic emissions by shutting down many polluting industries, banning high-emission vehicles, and restricting the number of on-road vehicles in Beijing (Zhang et al., 2012).It is likely that our emissions were overestimated in Beijing, which caused a large discrepancy between modeled and observed SO 2 mixing ratios.
The spatial and seasonal distribution of NO 2 is similar to SO 2 as shown in Fig. 10.A high NO 2 mixing ratio is found over northeastern, northern, and eastern China due to high emissions from the power plant, industry and transportation sectors in these regions.Outside China, several hotspots are identified, such as Seoul (South Korea) and New Delhi (India).The modeled NO 2 mixing ratios have a summer minimum and a winter maximum.The lifetime of NO 2 in winter is relatively longer (18-24 h) than that in summer (6 h) because the concentration of hydroxyl radical (OH) in the atmosphere is low (Beirle et al., 2003).Consequently, the re-moval reaction of NO 2 with OH radicals to form HNO 3 is less active in winter than in summer.Among the four sites in China, the model performs well in predicting observed NO 2 mixing ratios at the Shanghai site (MFB = −9 %); however, it underestimates at the other three sites (MFB > −53 %).WRF-Chem captures the seasonal variability of NO 2 , but underestimates the monthly average of NO 2 , with a MFB between −41 and −68 % for all 4 months.Underestimation of NO 2 has also been reported in the South Asian region using WRF-Chem (Kumar et al., 2012) and a possible reason was proposed as the underestimation of NO x emissions from biomass burning or anthropogenic sources.Another potential reason is that the removal of NO x was overestimated through the heterogeneous reaction of N 2 O 5 to form nitric acid in the RADM2 WRF-Chem chemical mechanism (Yegorova et al., 2011), used in this study.

Conclusions
We performed WRF-Chem simulation of air quality over East and South Asia using two different anthropogenic emissions inventories and evaluated the model performance for PM 10 concentrations, as well as O 3 , SO 2 , and NO 2 mixing ratios, using ground-level observations for the year 2007.We find that large discrepancies exist between the extensively used EDGAR global anthropogenic emissions and the REAS regional inventory at national and provincial scales.The discrepancies between these inventories can lead to large differences in simulated surface PM 10 concentrations (40-70 %) and moderate differences in O 3 mixing ratios (16-20 %) in most areas of the North China Plain, as well as more than 100 % differences in SO 2 and NO 2 mixing ratios, found in several provinces in China.Our study demonstrates that WRF-Chem is sensitive to emissions inventories and improvements in emissions inventories are important for accurately simulating regional air quality.Further studies are needed to assess model performance differences due to different emission inputs.
On the basis of lower bias and error values versus observations we found for our WRF-Chem-REAS simulations, we chose the REAS inventory to conduct four 1-month simulations for the purpose of model evaluation.The model results indicate clear regional variations in the seasonal cycle of surface PM 10 and O 3 over East and South Asia.In northwestern China, maximum PM 10 occurs in April, while in Nepal and other regions of China, the highest PM 10 mainly occurs in January.For surface O 3 mixing ratios, the peak values are simulated in July for northern and eastern China, and in April for Tibet and Japan.Comparisons between model simulations and observations show that the model performs well in simulating surface PM 10 and O 3 , meeting air quality model performance criteria for both PM 10 and O 3 at most sites, although the model underestimates PM 10 at some sites in China in January.The model predicts SO 2 better at sites in Japan than in China, where overestimation is large at the Beijing site in July and October.The model underestimates most observed NO 2 in all 4 months.
Quantifying uncertainties of simulated air quality at the provincial level due to emission inputs reveals that the uncertainty in emissions inventories leads to significant differences in simulated levels of air pollutants, especially PM 10 , SO 2 , and NO 2 .For O 3 , on the other hand, different emissions inventories lead to only a moderate variability, showing agreement with the findings of previous studies (Ma and van Aardenne, 2004;Amnuaylojaroen et al., 2014).Our study highlights the importance of better constraining emissions at the provincial level for regional air quality modeling over East Asia, where anthropogenic emissions are high and air pollution is a major environmental and public health challenge.Model evaluation results also indicate that emissions inventories that do not consider local emissions control policies could cause large discrepancies.Our results suggest that future work should focus on better constraining the provinciallevel emissions especially estimating the primary PM, SO 2 , and NO x .

Figure 1 .
Figure 1.WRF-Chem model domain and observation sites.Blue shading indicates locations where the REAS emissions inventory is used.Gray shading indicates where the RCP8.5 emissions are used.For the entire model domain, biomass burning emissions from GFED v3 and biogenic emissions from POET v1 are used.Redfilled circles denote the observational sites with PM 10 ; orange triangles sites with O 3 ; purple crosses sites with SO 2 ; and green squares sites with NO 2 .
2.5 measurements in China started in late 2012.The observed PM 10 , O 3 , and SO 2 in Japan and SO 2 and NO 2 in China are taken from the Acid Deposition Monitoring Network in East Asia (EANET).The surface mixing ratios of O 3 in Mt.Lulin are taken from the Lulin Atmospheric Background Station (LABS, 2862 m above mean sea level) in central Taiwan (Ou Yang et al., 2012).The description of each site is listed in Table S3a-b; the locations of these sites are shown in Fig. 1.

Figure 4 .
Figure 4. Simulated and observed monthly average surface PM 10 in 2007 using WRF-Chem-REAS.The filled circles indicate the observed monthly average values.

Figure 5 .
Figure 5. Comparisons of simulated and observed daily mean PM 10 (µg m −3 ) in northeastern, northern, northwestern, and central China in each month.The model to observation ratios of 2 : 1, 1 : 1, and 1 : 2 are represented by orange lines.Monthly average performance statistics (r, MFB, and MFE) are listed.

Figure 6 .
Figure 6.Comparisons of simulated and observed daily mean PM 10 (µg m −3 ) in eastern, southwestern, and southern China, and Japan in each month.The model to observation ratios of 2 : 1, 1 : 1, and 1 : 2 are represented by orange lines.Monthly average performance statistics (r, MFB, and MFE) are listed.

Figure 7 .
Figure 7. Simulated and observed monthly average surface O 3 in 2007 using WRF-Chem-REAS.The filled circles indicate the observed monthly average values.

Figure 8 .
Figure 8. Comparisons of observed (blue dots) and modeled (red lines) daily mean O 3 (ppbv) at seven sites in Japan and one site in Taiwan.

Figure 9 .
Figure 9. Simulated and observed monthly average surface SO 2 in 2007 using WRF-Chem-REAS.The filled circles indicate the observed monthly average values.

Figure 10 .
Figure 10.Simulated and observed monthly average surface NO 2 in 2007 using WRF-Chem-REAS.The filled circles indicate the observed monthly average values.

Figure 11 .
Figure 11.Comparisons of observed (blue dots) and modeled (red lines) daily mean SO 2 (ppbv) at six sites in China and Japan and NO 2 (ppbv) at three sites in China.

Table 1 .
List of total emissions for major pollutants from REAS and EDGAR over the model domain in July 2007.Unit is Tg month −1 .
Figure 2. Monthly emissions difference of PM 10 , CO, SO 2 , and NO x between REAS and EDGAR in July 2007 in our model domain.

Table 2 .
Statistical measures calculated for model simulations using REAS and EDGAR as emissions inputs for PM 10 , O 3 , SO 2 , and NO 2 .r is the correlation coefficient between observations and model simulations; NMB (%) is the normalized mean bias between observations and model simulations; MFB (%) and MFE (%) are the mean fractional bias and mean fractional error; NMSE is the normalized mean square error between the observations and model.

Table 3 .
Statistical performance of WRF-Chem-REAS simulations for PM 10 in 2007.Count is the total number of observations for calculation; Obs (µg m −3 ) and Model (µg m −3 ) are 4-month mean daily average value of observations and model simulations, respectively.Other indicators and associated units are described in Table2.

Table 4 .
Statistical performance of WRF-Chem-REAS simulations for O 3 in 2007.The unit of Obs and Model is ppbv.Other statistical indicators and associated units are described in Table 2.

Table 5 .
Statistical performance of WRF-Chem-REAS simulations for SO 2 in 2007.The unit of Obs and Model is ppbv.Other statistical indicators and associated units are described in Table 2.

Table 6 .
Statistical performance of WRF-Chem-REAS simulations for NO 2 in 2007.The unit of Obs and Model is ppbv.Other statistical indicators and associated units are described in Table 2.