Verification of Chemical Transport Models for PM 2 . 5 Chemical Composition Using Simultaneous Measurement Data over Japan

Evaluation of models for simulating temporal and spatial variations of PM2.5 chemical composition in Japan has been limited by the lack of observational data. In this study, we used PM2.5 chemical composition data measured simultaneously over several regions of Japan in winter, spring, and summer 2012 to evaluate three sensitivity simulations, one based on a secondary organic aerosol (SOA) yield model and two based on a volatility basis set (VBS) model. Concentrations of sulfate (SO4), nitrate (NO3), and ammonium (NH4) were well reproduced by all the simulations in summer. However, in winter and spring, SO4 concentrations were underestimated and NO3 concentrations were overestimated by the standard simulation. NO3 concentrations were better reproduced by a model with dry-deposition velocities of nitric acid and ammonia enhanced by a factor of 5, as was done in a previous study. Observed concentrations of elemental carbon and organic aerosol (OA) were higher at urban sites than at the surrounding remote sites, and this behavior was not adequately reproduced by models with a grid size of 15 km. Further refinements of emission inventories and models are necessary for better simulations of PM2.5 chemical compositions. OA concentration was greatly underestimated by the simulation based on the SOA yield model over all the seasons but was better reproduced by the simulations based on the VBS model in spring and summer because aging reactions were considered in the VBS model-based simulations. The VBS model-based simulations reproduced the observations that primary OA predominated in winter and that the contribution of SOA was higher than that of primary OA in spring and summer. These contributions should be validated by means of observation-based source contributions of OA in future studies.


INTRODUCTION
Environmental standards for PM 2.5 were enacted in Japan in 2009 (annual average, 15 µg m -3 ; daily average, 35 µg m -3 ), but in 2012 there were many observational sites in western Japan and the Tokyo metropolitan area (TMA) at which these standards were not attained (Ministry of Environment, 2014).Chemical transport models (CTMs) are useful tools for obtaining quantitative information about PM 2.5 emission sources.In general, the performance of CTMs depends strongly on the region and the season, so models must be evaluated for specific target areas and time periods.In Japan, several studies have evaluated PM 2.5 models (Morino et al., 2010a;Ikeda et al., 2014;Shimadera et al., 2014a).However, in those studies, the temporal and spatial coverage of measured PM 2.5 chemical compositions was limited, and thus the performance of the models with regard to PM 2.5 chemical composition was not sufficiently evaluated over the whole of Japan.In 2012, PM 2.5 chemical compositions were measured simultaneously over several areas in Japan in winter, spring, and summer, and this data thus became available for evaluation of PM 2.5 models.
Previous reports have pointed out that concentrations of organic aerosol (OA) were greatly underestimated by CTMs with secondary organic aerosol (SOA) yield models in western Japan and in the TMA (Morino et al., 2010a;Ikeda et al., 2014).Morino et al. (2010b) showed that underestimation of OA in the TMA was due to underestimation of concentrations of SOA from both fossil and biogenic sources.Recently, a volatility basis set (VBS) model that can deal with direct emissions of semivolatile and intermediate-volatility organic compounds (SVOCs and IVOCs, respectively), with the oxidation of these compounds, and with SVOC aging was developed (Donahue et al., 2006;Robinson et al., 2007).The VBS model has been widely used in the United States (Lane et al., 2008;Murphy and Pandis, 2009;Tessum et al., 2015), Europe (Bergstrom et al., 2012;Zhang et al., 2013), and Asia (Matsui et al., 2014;Han et al., 2015;Morino et al., 2015), and has improved simulations of OA concentrations.Morino et al. (2014) compared the results obtained with five SOA models over the TMA and found that the VBS model reproduced the observed SOA concentrations well, whereas other models, including SOA yield models, underestimated SOA.
In this study, we evaluated the performance of a CTM for simulation of PM 2.5 over Japan by using PM 2.5 chemical composition data that were measured simultaneously over several regions of Japan.In addition, we compared the results obtained with a VBS model with the results obtained with a SOA yield model.Our evaluation of the VBS model using simultaneously measured data can be expected to improve our understanding of the advantages and disadvantages of the VBS model.

Model System and Sensitivity Simulations
We simulated the distributions of gaseous and particulate species by using a three-dimensional CTM, the Models-3 Community Multiscale Air Quality (CMAQ, v5.0.2) modeling system developed by the U.S. Environmental Protection Agency (Byun and Schere, 2006).The model setups were similar to those of Morino et al. (2015).The chemical mechanism was based on the Carbon Bond Mechanism 05 (CB05) model of Yarwood et al. (2005).We used the sixthgeneration aerosol module of CMAQ (AERO6, which is based on a SOA yield model), as well as AERO6 coupled with a VBS model (AERO6VBS).Compared to the SOA yield model in AERO6, the VBS model in AERO6VBS has two advantages.First, the VBS model considers direct emissions of SVOCs and IVOCs from combustion sources and calculates the partitioning of primary OA (POA), SVOCs, and IVOCs by classifying these species on the basis of effective saturation concentrations ranging from 10 -2 to 10 6 µg m -3 .Emission profiles of POA, SVOCs, and IVOCs in the VBS model were taken from Shrivastava et al. (2011).Second, the VBS model treats chemical aging of directly emitted SVOCs and IVOCs, as well as chemical aging of SVOCs and IVOCs generated by oxidation of volatile organic compounds (VOCs).In this study, the VBS model included as chemical aging only gas-phase oxidation of organic compounds by OH radicals (Donahue et al., 2006).The rate of the reaction between directly emitted SVOCs and OH radicals was 4 × 10 -11 cm 3 molecule -1 s -1 , and the rate of reaction between OH radicals and SVOCs generated from anthropogenic VOCs was 2 × 10 -11 cm 3 molecule -1 s -1 .
Owing to chemical aging, the saturation concentrations of SVOCs and IVOCs decreased by a factor of 10 with a small net increase in mass (7.5%) to account for added oxygen.
One sensitivity simulation was conducted with the AERO6 model, and two sensitivity simulations were conducted with the AERO6VBS model (Table 1): a standard version and a revised version.In the standard version (AERO6VBS-std), aging of SVOCs from biogenic VOCs was not considered, as in Murphy and Pandis (2009).The revised sensitivity simulation (AERO6VBS-rev) considered an aging reaction between OH radicals and SVOCs produced from biogenic VOCs with a reaction rate of 2 × 10 -11 cm 3 molecule -1 s -1 .In addition, in the AERO6VBS-rev simulation, dry-deposition velocities for nitric acid (HNO 3 ) and ammonia (NH 3 ) were enhanced by a factor of 5, as was done by Shimadera et al. (2014a).
Meteorological fields were calculated with the Weather Forecast Research (WRF) model version 3.3 (Skamarock et al., 2008).The WRF hourly output files were processed for the CMAQ input files using the Meteorology-Chemistry Interface Processor (Byun and Schere, 2006).In the WRF simulation, analysis nudging was conducted using the threedimensional meteorological fields from the National Centers for Environmental Prediction Final Analysis datasets available with 1° × 1° horizontal resolution every 6 hours.
As for the lateral boundary conditions, monthly averaged data were obtained from the global chemical climate model Chemical AGCM for Study of Atmospheric Environment and Radiative Forcing (CHASER) (Sudo et al., 2002).
Data for emissions from anthropogenic and natural sources in the simulation domains are summarized in Table 2.We used two simulation domains: Domain 1 covers East Asia with a horizontal resolution of 60 km, and Domain 2 covers Japan with a horizontal resolution of 15 km (Fig. 1).Considering the spatial resolution of the emission inventories, we chose a grid size of 15 km.The limitation associated with this grid size is discussed in the section of Results and Discussion.Simulation periods were 1 January-29 February 2012 (winter), 1 April-31 May 2012 (spring), and 1 July-31 August 2012 (summer), with a spin-up calculation of 10 days.

PM 2.5 Observation Sites and Sampling Periods
Filter sampling of PM 2.5 was conducted by the National Institute for Environmental Studies and local government institutes at 13 sites (3 urban, 4 suburban, and 6 remote; Table 3 and Fig. 1) with a low-volume air sampler (FRM2025; Thermo Fisher Scientific Inc., Waltham, MA, USA).Observations were conducted at both urban/suburban Dry-deposition velocities of HNO 3 and NH 3 were enhanced by a factor of 5 Aging of SVOCs produced from BVOCs was introduced (k OH = 2 × 10 −11 cm 3 molecule −1 s −1 ) a Abbreviations: SVOCs, semivolatile organic compounds; BVOCs, biogenic volatile organic compounds.(JATOP 2012a, b).b Regional Emission Inventory in Asia (Kurokawa et al., 2013).c Global Fire Emissions Database (van der Werf et al., 2006).d Aerosol Comparisons between Observations and Models (Diehl et al., 2012).e Japan Meteorological Agency (Kazahaya et al., 2001).f Model of Emissions of Gases and Aerosols from Nature (Guenther et al., 2012).g Extrapolated from datasets for 2010 using statistical activity data.3. Rishiri Hokkaido Remote 45.12 141.20 0 a Urban fraction in each model grid of the measurement site.Urban fractions were estimated from Land Use Fragmented Mesh in Urban Area data (version 1.0) of the National Land Numerical Information database (Ministry of Land, Infrastructure, Transport and Tourism, 2013).b Urban area surrounding the measurement site in 2000 (Ministry of Land, Infrastructure, Transport and Tourism, 2012).and remote sites in several Japanese regions (i.e., Kyushu, Chugoku, Kansai, Chubu, and Hokkaido), so the contributions of urban pollution and regional background could be roughly estimated.Sampling periods were 6 h at the urban/suburban sites and 12 h at the remote sites.Measurement data in winter (11-25 January 2012), spring (9-23 May 2012), and summer (26 July-9 August 2012) were used in this analysis.Concentrations of ionic species were analyzed by means of ion chromatography and elemental carbon (EC) and organic carbon (OC) concentrations were determined with a thermal/ optical carbon analyzer (DRI model 2001A; Atmoslytic Inc., CA, USA or Sunset Laboratory, Inc., OR, USA) on the basis of the IMPROVE protocol (Chow et al., 1993).The ratio of molecular weight and carbon weight (i.e., OA/OC ratio) was assumed to be 1.4 (Turpin and Lim, 2001).

Temporal and Spatial Variation of PM 2.5 Chemical Composition
Fig. 2 shows time series of PM 2.5 chemical compositions at three sites in the Kansai region of Japan (Osaka, Shiga, and Kyotango) during winter, spring, and summer.Model performance averaged over all the sites is summarized in Table 4. Time series of PM 2.5 chemical composition and gaseous species at all 13 sites are shown in Figs.A1 and A2 of the Supporting Information.) concentrations at most of the sites.For example, observed SO 4 2-concentration increased during 13-14 January at many sites, but the magnitude of these peaks was not quantitatively reproduced by the models.In addition, SO 4 2-concentrations on other winter days were also greatly underestimated; on average, the models underestimated observed SO 4 2-concentrations by a factor of 5 (Table 4).Note also that during 13-14 January, both observed and simulated SO 2 concentrations increased (Fig. A2(a), Supporting Information), suggesting that SO 4 2-and SO 2 originated from similar emission sources.Analysis of the spatial distributions of the simulated SO 4 2-and SO 2 concentrations indicated that both species were transported from eastern China (Fig. A3, Supporting Information).However, the observed SO 4 2-/SO 2 ratios (0.2-1.0) were greatly underestimated by all three simulations (0.02-0.1) during this period (Fig. 3).In addition, in winter the observed SO 4 2-concentrations were clearly higher in western Japan than in eastern Japan (Fig. 4).These spatial variations of SO 4 2-concentration indicate that important sources of SO 4 2are located in western Japan or on the Eurasian continent, as previously suggested by Aikawa et al. (2010).This 2-concentration during 8-9 May and 25-30 July were also underestimated by the simulations.However, SO 4 2-concentrations were generally reproduced well on other days during the spring and summer.In spring, SO 4 2-concentrations were substantially higher in western Japan than in eastern Japan (Fig. 4), and this behavior was reproduced by the simulations.In summer, observed SO 4 2-concentrations were distributed homogeneously from west to east, and this distribution was roughly reproduced by the models.

(b) Nitrate and Ammonium
In contrast to SO 4 2-concentrations, nitrate (NO 3 -) concentrations were substantially overestimated by both the AERO6 and the AERO6VBS-std simulations in winter and spring (Figs.2(a), 2(b), A1(a), A1(b) and Table 4).The simulated results showed NO 3 -peaks that were not actually observed; these peaks occurred mostly during nighttime, when the temperature and the boundary layer height were low.98.9 a Model/Obs indicates ratios of simulated to observed concentrations.FA2 and FA10 indicate the proportions (%) of simulated data that reproduced the observed data within factors of 2 and 10, respectively.The minimum cutoff of observed concentration was 0.5 µg m −3 .Abbreviations: EC, elemental carbon; OA, organic aerosol.3.

Many factors control NO 3
-concentration, including gaseous and heterogeneous reactions, gas-particle conversion, and deposition.Shimadera et al. (2014a) reported overestimation of NO 3 -concentrations by a CTM in the Kanto region in winter 2010 and summer 2011.Those investigators conducted sensitivity analyses of meteorological factors, emission rates, dry-deposition velocities, and heterogeneous reaction rates and found that higher dry-deposition velocities of HNO 3 and NH 3 led to better simulation of NO 3 -, whereas the model results were less sensitive to other factors.Neuman et al. (2004) indicated that dry-deposition velocities of HNO 3 estimated from aircraft measurements of power plant plumes  11 000 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0 0 12 12 Fig. 4. Observed and simulated concentrations of SO 4 2-, NO 3 -, NH 4 + , EC, and OA in winter (left), spring (center), and summer (right).The numbers on the lower axes correspond to the measurement sites listed in Table 3. in the United States (Texas) were 4 times as high as reported dry-deposition velocities.Thus, dry-deposition velocity is a source of large uncertainty, and like Shimadera et al. (2014a), we enhanced dry-deposition velocities of HNO 3 and NH 3 by a factor of 5 in the AERO6VBS-rev simulation.As a result, the AERO6VBS-rev simulation reproduced the observed NO 3 -concentrations much better than did standard simulations, which overestimated average NO 3 concentrations by a factor of 3 to 15 (Table 4).However, in winter, observed NO 3 -concentrations were higher at urban/suburban sites than at remote sites (Fig. 4).This behavior was not well captured by the simulations, even the AERO6VBS-rev simulation; this problem is further discussed in the next subsection.Observed NO 3 -concentrations were low in summer (Figs.2(c) and A1(c)), and this behavior was reproduced by all the simulations.
Both observed and simulated equivalent molar concentrations of ammonium (NH 4 + ) were close to the sum of the SO 4 2-and NO 3 -concentrations in all three seasons.Thus, the model performance for NH 4 + can be represented by combination of the model performances for SO 4 2-and NO 3 -.

(c) EC
In all three seasons, EC concentrations were generally reproduced well by the models at an urban site (Osaka) and were underestimated at the Shiga and Kyotango sites (Fig. 2).
Temporal patterns were generally reproduced well at these sites.Observed EC concentrations were higher at urban/ suburban sites than at remote sites in all three seasons (Fig. 4).In several regions, this behavior was not adequately reproduced by the models (e.g., Kyushu and Chugoku).Grid resolution is a critical factor in simulating EC.A previous study showed that a grid size of 15 km may not be of sufficiently high resolution to resolve urban air pollution (Stroud et al., 2011).In suburban cities, urban zones are generally smaller than our grid size (15 × 15 km 2 ), as shown in Table 3.Most of the urban/suburban sites were located in urban zones (except for the Niigata site), but urban fractions in the model grids containing the measurement sites were occasionally less than half.Thus, particularly in suburban areas, primary emission sources were inhomogeneously distributed in each model grid.Simulations with finer emission inventories and a finer grid size are necessary for examination of the sensitivity of the simulated results to the grid resolutions of the model.As noted in the previous section, the simulations did not reproduce the observed contrast in NO 3 -concentrations between the urban/suburban sites and the remote sites in winter.This problem can also have been due to the coarseness of the model grid.

(d) OA
The model performance for OA depended strongly on the season and differed substantially between simulations: in winter, all the simulations showed similar OA concentrations and greatly underestimated the observed concentrations (Figs.2(a) and A1(a)).All the simulations showed that POA dominated OA in winter (Fig. 5) and that the OA distributions were similar to each other over Japan, with enhancement in western Japan and around urban areas (Fig. 6).As was the case for EC, it is likely that urban pollution by POA could not be resolved at the grid size of the model simulations.Simulated SOA concentrations were low over all of Japan for all the simulations (Fig. 7).
In spring and summer at all the sites, the AERO6VBS-rev simulation showed the highest OA concentrations, followed by the AERO6VBS-std simulation (Figs.2(b), 2(c), 4 and Table 4).Differences between the three simulations with regard to OA were mostly associated with the simulations' consideration of aging processes of SVOCs from anthropogenic and biogenic VOCs.The AERO6VBS-rev simulation generally reproduced the observed increases of OA concentration during 7-10 May and 26-30 July 2012.Statistical analyses also indicated that in most cases, the AERO6VBS-rev simulation was the best at reproducing the observed data (Table 4).However, diurnal variation of OA at Osaka in summer was not adequately reproduced by any of the simulations: the observational data showed a sharp increase during the daytime, and this increase was not reproduced by the simulations (Fig. 2(c)).It has been suggested that in several urban areas, daytime increases in OA are due to increases in fossil SOA (Morino et al., 2010b(Morino et al., , 2015)).It is possible that the AERO6VBS-rev simulation also underestimated the daytime increase of anthropogenic SOA (ASOA) at Osaka.
showed lower POA concentrations than the AERO6 simulation because the former two simulations consider POA evaporation.However, the effect of this difference in POA concentration on total OA concentration was compensated for by the much higher concentrations of ASOA and biogenic SOA (BSOA) predicted by the AERO6VBS-rev simulation (Fig. 5).On average, the AERO6VBS-rev simulation showed that POA, ASOA, and BSOA had similar contributions to total OA in spring.All of these species showed higher concentrations in western Japan than in eastern Japan (Fig. 7).
In summer, the AERO6VBS-rev simulation showed that BSOA had the highest contribution to total OA, followed by ASOA (Fig. 5).BSOA concentrations were higher in summer than in the other seasons, most likely because of the higher photochemical activity and higher emissions of biogenic VOCs during the summer.In contrast, all of the simulations showed that the POA concentration was the lowest in summer.This result is due to the fact that the boundary layer height is highest in summer, and intrusions of clean maritime air from the Pacific Ocean due to the prevailing southerly wind are more frequent in summer.This is supported by the low POA concentrations on the Pacific side of Japan (Fig. 7).
The contributions of the aging processes of SVOCs derived from anthropogenic and biogenic VOCs to OA concentrations can be roughly estimated from the differences between the AERO6VBS-std and AERO6 simulations and between the AERO6VBS-rev and AERO6VBS-std simulations, respectively.Fig. 6 suggests that aging processes strongly affected the simulated OA concentrations over the whole of Japan in spring and summer.

Comparison of Model Performance with That of Other Models
In the standard simulations (AERO6 and AERO6VBS-std) in this study, SO 4 2-concentrations were underestimated and NO 3 -concentrations were overestimated in winter and spring.Previous studies in Japan (the TMA and western  results are similar to the results of a study by Morino et al. (2010a), in which the ensemble means of four CTMs in the TMA in summer 2007 reproduced observed concentrations (Morino et al., 2010a).In Beijing, China, PM 2.5 chemical compositions were measured during the period from December 2010 to January 2012, and the data were used to evaluate the performance of a CTM (CMAQ) (Lang et al., 2013).The model underestimated observed SO 4 2-concentrations in summer and generally reproduced them in winter.By contrast, observed NO 3 concentrations were greatly underestimated in winter and reproduced relatively well in summer.In the United States, PM 2.5 chemical compositions have been measured at hundreds of sites under several frameworks, and CTMs have been comprehensively evaluated.For example, Appel et al. (2008)  2-concentrations were underestimated by 5% to 21%, and NO 3 -concentrations were underestimated by 7% to 45%.These results clearly indicate that problems with model simulations differ between regions and seasons, owing to differences in the factors that control PM 2.5 concentrations, such as meteorological conditions, emission inventories, chemical/aerosol processes, and deposition processes (dry or wet).The problems in simulating SO 4 2-and NO 3 -concentrations found in this study are not always found in other countries; further studies to resolve these problems are necessary.
In the above-mentioned studies, OA concentrations were commonly underestimated by CTMs with SOA yield models in summer.Bergstrom et al. (2012) found that a VBS model reproduced observed OA concentrations over Europe in summer reasonably well.Matsui et al. (2014) reported that OA concentrations were better reproduced by a VBS model in western Japan and the TMA because the model included SVOC aging processes.VBS models clearly give better results in OA simulations than do SOA yield models.Nevertheless, VBS models have many sources of uncertainty, which are discussed in the next subsection.

Model Uncertainties and Future Directions
The relative contributions of POA, ASOA, and BSOA depend strongly on several setup parameters of the VBS model (see, e.g., Morino et al., 2014).As already noted, aging reactions contribute substantially to ASOA and BSOA concentrations.In the VBS model, only one or two aging reaction rates for ASOA and BSOA were used and the simulation results strongly depended on these rates.In up-to-date SOA models, aging reaction rates are estimated as a function of carbon number and O/C ratio by means of structure-activity relationships, and the rates of SVOC reactions vary by 1 to 2 orders of magnitude, depending on the species (Cappa and Wilson, 2012;Zhang and Seinfeld, 2013).In addition, several processes are not included in the current VBS model (e.g., branching between functionalization and fragmentation, particle-phase reactions).
Another uncertainty is in the emission profiles of SVOCs and IVOCs.OA concentrations depend strongly on the emission profiles (Morino et al., 2014).In Japan, PM 2.5 emission factors used in the emission inventories of the Japan Auto-Oil Program (JATOP) were measured with dilution chambers for vehicle sources and without dilution chambers for stationary combustion sources (JATOP, 2012a, b and references therein).PM 2.5 sampling conditions are critical in determining emission profiles of SVOCs and IVOCs, although these data were not sufficiently well organized.Development of detailed SVOC and IVOC emission inventories with careful reviews of sampling conditions is necessary in future studies.
Overall, OA models have been validated with observed total OA or OC concentrations.For better model evaluation, observation-based characterizations of OA obtained by the use of organic tracer species, highly time-resolved measurements combined with factor analyses, or radiocarbon have been shown to be useful (Hallquist et al., 2009).In particular, model evaluations of the respective contributions of ASOA, BSOA, and POA provide useful information regarding the use of CTMs to establish strategies for controlling PM 2.5 .
Finally, we should reiterate that the simulations conducted in this study did not capture the higher OA concentrations in the urban/suburban areas than in the remote areas.In addition, diurnal variation of OA concentration was not reproduced even by the AERO6VBS-rev simulation at urban sites in summer.For better simulation of both spatial and temporal variations of OA, finer emission inventories, as well as model simulations with finer grids, should be evaluated in future studies.

CONCLUSIONS
We used PM 2.5 chemical composition data measured simultaneously over several regions of Japan in winter, spring, and summer 2012 to evaluate three sensitivity simulations, one based on a SOA yield model and two based on a VBS model.In summer, the concentrations of SO 4 2-, NO 3 -, and NH 4 + were well reproduced by all the simulations.However, in winter and spring, SO 4 2-concentrations were underestimated and NO 3 -concentrations were overestimated by the standard simulation.NO 3 -concentrations were better reproduced when the dry-deposition velocities of HNO 3 and NH 3 were enhanced by a factor of 5, as was done in a previous study.The performances of these simulations in this study were different from those of previous studies in other geographical areas.Thus, the problems with the simulations reported in this study are not always found in other countries, and further refinement of the CTMs, the emission inventories, or both are necessary for better simulation of SO 4 2-and NO 3 -concentrations.Observed concentrations of EC and OA were higher at urban sites than at surrounding remote sites, and this behavior was not adequately reproduced by the model, which had a grid size of 15 km.Further refinements of emission inventories and models are necessary.
OA concentrations were greatly underestimated by the SOA yield model in all the seasons but were better reproduced by the VBS model in spring and summer because the contributions of aging reactions were simulated by the VBS model.The VBS model simulated that POA predominated in winter and that the contribution of SOA was higher than that of POA in spring and summer.This better simulation of OA by the VBS model in spring and summer is consistent with the results of previous studies.However, the VBS model still includes large uncertainties, and further model refinement, as well as model validation using observation-based source contributions of OA, should be conducted in future studies.

Fig. 1 .
Fig. 1.Model domains and sites of PM 2.5 chemical composition measurements.The circled numbers in the right panel correspond to the measurement sites listed in Table3.

Table 1 .
Setups of simulations used in this study a .

Table 2 .
Setups of emission inventories used in this study.

Table 3 . Sites of PM 2.5 chemical composition measurement.
a (%) Urban area (km 2 ) b

Table 4 .
Performance of the three simulations averaged over the measurement sites in three seasons a .