Model Inter-Comparison for PM2.5 Components over urban Areas in Japan in the J-STREAM Framework

: A model inter-comparison of secondary pollutant simulations over urban areas in Japan, the ﬁrst phase of Japan’s study for reference air quality modeling (J-STREAM Phase I), was conducted using 32 model settings. Simulated hourly concentrations of nitric oxide (NO) and nitrogen dioxide (NO 2 ), which are primary pollutant precursors of particulate matter with a diameter of 2.5 µ m or less (PM 2.5 ), showed good agreement with the observed concentrations, PM 2.5 indicated some problems in the simulated local meteorology such as the atmospheric stability. This model inter-comparison suggests that these deviations may be owing to a need for further improvements both in the emission inventories and additional formation pathways in chemical transport models, and meteorological conditions also require improvement to simulate elevated atmospheric pollutants. Additional accumulated observations are likely needed to further evaluate the simulated concentrations and improve the model performance. the model simulations. of global simulations.

. Dates of enhanced simulation periods for model evaluations including a simulation spin-up. Updated from the overview of Japan's study for reference air quality modeling (J-STREAM) [5].

Season Dates
Spring 2013  Four nested model domains, d01, d02, d03, and d04, on a Lambert conformal map projection were employed in the J-STREAM project [5]. The finest domains, d03 and d04, with a 5 × 5 km grid, cover the major city clusters in western Japan, including Osaka, Kobe, Kyoto, and Nagoya, and the Tokyo metropolitan area, respectively. Simulated concentrations in the d03 and d04 domains were used for model evaluations, and the results are discussed in following sections. Figure 1 shows the d03 and d04 domains, including the locations of ambient APMSs (AAPMSs), for which simulated concentrations were evaluated via comparisons with observations.
Atmosphere 2020, 11,222 3 of 26 model frames and/or model settings, including boundary and inputted conditions and physical and chemical mechanisms. Detailed model settings are described below. Furthermore, these including an introduction of J-STREAM can be found in previous research for the overview [5] and the performance on ozone [13]. The main target of the first phase of J-STREAM (J-STREAM Phase I) is to evaluate the general performances of participant models on secondary atmospheric concentrations over urban areas in Japan. Daily concentrations of PM2.5 components in each season among others were treated as subjects of evaluation in this paper. The enhanced simulation periods of J-STREAM Phase I were the spring of 2013, 27 April-26 May 2013, the summer of 2013, 12 July-10 August 2013, the autumn of 2013, 11 October-9 November 2013, and the winter of 2014, 10 January-8 February 2014, which corresponded to the seasonal periods of the national observation frame for PM2.5 components ( Table  1). The detailed evaluations and additional experiments for individual participant models can be found in [14]. Table 1. Dates of enhanced simulation periods for model evaluations including a simulation spin-up. Updated from the overview of Japan's study for reference air quality modeling (J-STREAM) [5]. Four nested model domains, d01, d02, d03, and d04, on a Lambert conformal map projection were employed in the J-STREAM project [5]. The finest domains, d03 and d04, with a 5 × 5 km grid, cover the major city clusters in western Japan, including Osaka, Kobe, Kyoto, and Nagoya, and the Tokyo metropolitan area, respectively. Simulated concentrations in the d03 and d04 domains were used for model evaluations, and the results are discussed in following sections. Figure 1 shows the d03 and d04 domains, including the locations of ambient APMSs (AAPMSs), for which simulated concentrations were evaluated via comparisons with observations.

Baseline Meteorological Model Configurations
The baseline meteorological simulation for J-STREAM Phase I was performed by the Weather Research and Forecasting (WRF) model, using the Advanced Research WRF (ARW) Version 3.7.1 [15]. The WRF inputted data were acquired from the National Centers for Environmental Prediction Final Operational Model Global Tropospheric Analyses (ds083.2) with a 1 × 1 degree resolution [16] and the Real-Time, Global Sea Surface Temperature High-Resolution (RTG_SST_HR) analysis with a 1/12 × 1/12 degree resolution [17] and a temporal resolution of 6 h for the initial and boundary conditions. The horizontal configurations of the one-way nested model domains, d01, d02, d03, and d04, are 220 × 170 grids with a 45-km horizontal resolution, 154 × 160 grids with a 15-km resolution, 82 × 61 grids with a 5-km resolution, and 64 × 70 grids with a 5-km resolution, respectively. The vertical grid structure consists of 31 layers from the surface to the model top (100 hPa). Five grids were trimmed off each of the four lateral boundaries for the offline CTMs. The physics parameterizations applied in this model included the WRF Single-Moment 5-class scheme [18], the Radiative Transfer Model (RRTM) [19] for a longwave radiation scheme, the Dudhia scheme [20] for a shortwave radiation scheme, the Noah Land Surface Model [21], the Mellor-Yamada Nakanishi and Niino surface layer scheme level 2.5 [22], and the Kain-Fritsch convective parameterization [23] for d01 and d02. No convection parameterization was used for the 5-km domains. The grid-nudging four-dimensional data assimilation technique was employed for wind, temperature, and water vapor from level 11 (approximately 2 km) to the top of the model at 100 hPa with the nudging coefficients of 1.0 × 10 −4 and 0.5 × 10 −5 s −1 for d01 and d02, respectively. Most of the participant CTMs employed baseline meteorological fields, while others employed the meteorology based on different model settings. The differences in the model settings in some participant models are described in Section 2.3.
The baseline meteorological fields were compared with hourly observations of the Japan Meteorological Agency (JMA) for the observation stations within d03 and d04 ( Figure 1

Chemical Transport Model Configurations
A total of 32 simulations were performed using three types of regional CTMs in J-STREAM Phase I: Community Multiscale Air Quality (CMAQ) [24], Comprehensive Air quality Model with eXtensions (CAMx) [25], and Weather Research and Forecasting-Chemistry (WRF-Chem) [26]. Table  2 presents the configurations of the employed models. All participants conducted J-STREAM simulation under their own usual simulation conditions. The CMAQ group (M01-M28) included The WRF using the baseline setting can generally simulate the observed meteorological conditions well. Meanwhile, WRF tended to overestimate the observed wind speeds. This was likely affected by the sparse horizontal resolution and coarse land information. The simulation performance of wind patterns was slightly better for d04 than that for d03 (Figures 2 and 3). However, simulated precipitation timing and their amounts were consistent with the observations (Figures 2 and 3).

Chemical Transport Model Configurations
A total of 32 simulations were performed using three types of regional CTMs in J-STREAM Phase I: Community Multiscale Air Quality (CMAQ) [24], Comprehensive Air quality Model with eXtensions (CAMx) [25], and Weather Research and Forecasting-Chemistry (WRF-Chem) [26]. Table 2 presents the configurations of the employed models. All participants conducted J-STREAM simulation under their own usual simulation conditions. The CMAQ group (M01-M28) included several versions, i.e., chemical mechanisms: Statewide Air Pollution Research Center mechanism (SAPRC) 99 [27], SAPRC07 [28], Carbon Bond (CB) 05 [29], and Regional Atmospheric Chemistry Mechanism (RACM) 2 [30], and three types of CMAQ aerosol calculation techniques [31]: aero5, aero6, and aero6 with the volatility basis set (VBS) approach [32]. These CMAQ aerosol calculation techniques employed ISORROPIA Version 1 [33,34] as an aerosol thermodynamic model and the second version of ISORROPIA (ISORROPIA Version 2) for updating the crustal species thermodynamics, the speciation schemes, and the SO 4 2− formation pathway [35] for versions after 5.0. The basic techniques of aero5 and aero6 include secondary organic aerosol (SOA) formation processes based on empirical parameters for SOA yields [36]. Major or minor updates were reflected in the chemical and aerosol mechanisms in the later versions. One CAMx model (M29) applied in J-STREAM used the SAPRC07 chemistry and the coarse and fine aerosol scheme treating both static coarse and fine mode aerosols [37]. The WRF-Chem group (M30-M32) included two Versions (3.8.1 and 3.7.1) that employed RADM2: the aerosol module of the Modal Aerosol Dynamics Model for Europe (MADE) [38] and the SOA Model (SORGAM) [39].
As described in detail in an overview article on J-STREAM [5], participants were requested to run CTM simulations during the enhanced target periods of four seasons for d03 or d04. As shown in Table 2, the simulations for some participant models began at d01 (M02, M03, M07-M15, M20, and M30-M32), but others began from the more inner domains. Fifteen participant models (M01-07, M14, Initial concentrations on the first day of each season and boundary concentrations throughout the entire target period were generated in the simulation using the M15 setting via CMAQ Version 5.0.2 with the SAPRC07-aero6 mechanisms for d01 and d02. Boundary concentrations for d01 of M15 were obtained from results for a chemical atmospheric general circulation model designed for studying atmospheric environment and radiative forcing, CHASER [40] for the Hemispheric Transport of Air Pollution (HTAP) Version 2 [41]. In J-STREAM Phase I, model-ready mosaic emission data corresponding to all participant chemical-aerosol mechanisms involving multiple emission inventories and results from an emission model for biogenic volatile organic compounds were provided: HTAP Version 2.2 [42] and Global Fire Emissions Database Version 4.1 [43] for Asian anthropogenic emissions, the Japan Auto-Oil Program (JATOP) emission inventory database (JEI-DB) [44], the updated JEI-DB [5], and Sasakawa Peace Foundation emissions for ships for Japanese anthropogenic emissions, volcanic emission data from Aerosol Comparisons between Observations and Models (AeroCom) [45] and JMA [46], and estimations obtained by using Model of Emissions of Gases and Aerosols from Nature Version 2.1 [47]. Most participant CTMs used model-ready input data; however, some participant CTMs performed simulations in their own emission frames. M03 used EAGrid2010-JAPAN [48], and M20 and M27 used EAGrid2000-JAPAN [49] for anthropogenic emissions in Japan. For the Asian scale anthropogenic emissions, M20 employed NASA INTEX-B [50] instead of HTAP Version 2.2. Additionally, some CTMs employed different emission injection heights. The Model for Ozone and Related chemical Tracers Version 4 (MOZART-4) [51], for instance, was used as boundary conditions in some model settings.
As mentioned in Section 2.2., most of the participants employed the baseline meteorological fields; however, other CTMs (M07, M20) used WRF-ARW outputs based on their own conditions, including physical options, parameterizations, and a fine input meteorological analysis data, which is the grid point value derived from the mesoscale model (GPV MSM) data by JMA.   [42]. "E2" uses EAGrid2000-JAPAN [49] and NASA INTEX-B [50]. "E3" uses EAGrid2000-JAPAN [49]. 3 Boundary concentration. "o" indicates that the baseline boundary concentration is used. "M" uses MOZART-4 [51]. "D" uses CMAQ defaults. 4 Meteorological condition. "o" indicates that the baseline metrological condition is used. "W" uses the meteorology simulated using WRF-ARW with own conditions, including physical options, parameterizations, and meteorological reanalysis. "WC" indicates the meteorology simulated using WRF-Chem with own conditions including physical options and parameterizations. 5 "o" indicates data submitted for d03 and d04 in each season. "su" was submitted for only summer. 6 NH 4 + and total PM 2.5 were not submitted.

Observational Data for Model Evaluation
A monitoring framework of ambient PM 2.5 components was initiated in the fiscal year 2011 under the Japan government initiative [4]. Over a period of at least two weeks set for each season, 1-day accumulated concentrations of PM 2.5 components, including ions (e.g., SO 4 2− , NO 3 − , and NH 4 + ), inorganic elements (e.g., Na, Al, K, and Ca), and carbonaceous aerosols (EC and OC), were monitored using the filter pack method at selected stations from three types of APMSs, including AAPMSs, roadside APMSs (RAPMSs), and background monitoring stations (BGMSs). PM 2.5 mass concentrations determined gravimetrically by weighing the filters were employed as the PM 2.5 mass concentration in this paper. Monitoring data from valid AAPMSs that obtained data for each PM 2.5 component over a period of at least eight days (53%) from each target period (up to 15 days) per each station were used to evaluate the performances of the participant CTMs. The number of valid AAPMSs was 16-22 stations for each domain and season. The data acquisition rate was highest in the summer, while a poor data acquisition rate was found for NO 3 − in autumn. Observed gaseous pollutants at these AAPMSs were also used to evaluate the simulated nitric oxide (NO), nitrogen dioxide (NO 2 ), and sulfur dioxide (SO 2 ). Figures 4 and 5 present observed daily concentrations for PM 2.5 components, i.e., SO 4 2− , NO 3 − , NH 4 + , EC, and OC, and total PM 2.5 mass for each 12-to 15-day seasonal period at the AAPMSs for d03 and d04, respectively. The box-and-whisker and black dots (outliers) means the differences between AAPMSs in each domain. In general, the concentrations of the total PM 2.5 and its components within a single domain exhibit similar day-to-day variabilities for each season. However, the frequency distributions of daily concentrations between the AAPMSs in each domain were enhanced, particularly for elevated concentrations (Figures 4 and 5). Therefore, the spatially averaged concentrations obtained from daily monitoring data for different AAPMSs within each domain were used for time series analysis hereafter.
For d03, i.e., western Japan, the seasonal-average total PM 2.5 concentrations were 17.3, 23.1, 20.3, and 19.6 µg/m 3 , with maximum daily concentrations of 31.9, 37.6, 36.3, and 34.3 µg/m 3 for spring, summer, autumn, and winter. The summer PM 2.5 concentration was slightly higher than those for the other seasons; however, the seasonal characteristics of the PM 2.5 concentration are unclear. SO 4 2− was a dominant PM 2.5 component, accounting for approximately 40% (9.1 µg/m 3 ) of the total PM 2.5 mass concentration in the summer. Meanwhile, from autumn to winter, the ratios of NO 3 − and OC to total PM 2.5 mass increased. The ratios of the five major PM 2.5 components were similar, with values of 12%-19% in the winter. On the dates when PM 2.5 was elevated, the AAPMS differences in PM 2.5 concentration levels increased, and considerably high PM 2.5 was found at AAPMSs placed at major cities: Osaka and Nagoya. These results were compared with those from rural areas.

Hourly Concentrations of Primary Pollutants
Major gaseous pollutants were also monitored at the AAPMSs. Figures 6 and 7 present the spatial averages of observed and simulated hourly concentrations of NO, NO2, and SO2 from different AAPMSs for d03 and d04, respectively. Table 3 summarizes the ensemble performances of the participant CTMs at each AAPMS for each season. For d04, the Tokyo metropolitan area, which is several hundred kilometers east of d03, the day-to-day changes in the concentrations of total PM 2.5 and its components were similar to those for d03; however, the seasonal-average concentrations: 15.2, 18.6, 17.9, and 19.3 µg/m 3 for spring, summer, autumn, and winter, were slightly lower than those for d03; whereas the maximum daily concentrations were 26.3, 35.4, 41.9, and 46.8 µg/m 3 . The elevated daily concentrations were obviously higher than those for d03 in the autumn and winter. Wintertime PM 2.5 concentrations were slightly higher than those in the other seasons, with increased daily concentrations; however, the seasonal characteristics of the PM 2.5 concentration were unclear for d04.

Hourly Concentrations of Primary Pollutants
Major gaseous pollutants were also monitored at the AAPMSs. Figures 6 and 7 present the spatial averages of observed and simulated hourly concentrations of NO, NO 2 , and SO 2 from different AAPMSs for d03 and d04, respectively. Table 3 summarizes the ensemble performances of the participant CTMs at each AAPMS for each season.      38 20 In general, most CTMs showed good agreement with the observed concentration levels of NO and NO 2 for each season, with regular diurnal patterns of NO in the warmer seasons. However, none of the models fully reproduced the elevated concentrations, e.g., for 19-20 May and 3 November (NO and NO 2 ) and 30 January (NO), with differences of 50%-200% between the observations and models, among others; the models tend to overestimate the observed daily maximums of NO: around 10-20 ppbv in spring and around 10-30 ppbv in summer, by a factor of 2.
For midnight on 30 January, all participant CTMs could not simulate the considerably increased level of NO before the rapid NO decrease associated with airmass changes, although all participants reproduced the NO decrease well. This suggests that CTMs successfully simulated the concentration change owing to meteorological changes in the synoptic scale but failed to simulate an increase in the amounts owing to local scale meteorological changes such as the strong atmospheric stability, especially during colder seasons. All models tended to overevaluate the daytime NO reduction. In particular, two WRF-Chem types (M30 and M31) and M05 produced strikingly low constant values, 0.001 or 0.000 ppbv, during the daylight hours in summer and autumn. The normalized mean bias (NMB) for both domains produced a strong underestimation of NO (approximately −40% to −50%), except during the spring. Underestimates of NO at remote stations in Japan have been observed for regional CTMs, as reported by MICS-Asia III results [9], and the correlations and index of agreement (IoA) values ranged from 0.18 to 0.43 and 0.41 to 0.51, respectively. The performance levels of each model exhibited substantial differences between both domains and seasons. The differences between seasons are likely related to meteorology simulation abilities, but the reasons for the differences appearing between domains are unclear in this stage.
The differences for NO 2 in each model were large. Among these models, M31, M32, and M30 tended to overestimate elevated NO 2 levels. The lower levels of NO 2 obtained by M30 were often comparable to the NO 2 concentration obtained by M03, which provided considerably lower NO 2 concentrations compared to other models. These results suggest that the differences in meteorological conditions and NO x chemistry in each model produced the NO 2 discrepancy between the models. Most of the models produced better results for NO 2 than for NO, with ensemble averages of seasonal statistics, e.g., correlation values, of 0.56 (d03) and 0.55 (d04), 0.72 (d03), and 0.71 (d04), particularly in the winter.
Over the year, most models obviously overestimated the observed SO 2 , with an ensemble bias of 1.7-4.2 ppbv (NMB: 120%-350%) for d03 and 1.5-2.5 ppbv (NMB: 160%-470%) for d04. In addition, relatively high SO 2 levels were found for M30, M31, and M32. Meanwhile, M03 and M20 tended to produce lower concentrations compared to the other models, with a negative bias of −1.3 ppbv (M03) and −0.2 ppbv (M20) recorded especially in the spring; and exhibited better performances (IoA: 0.58-0.59) over the other models (IoA: 0.30-0.39), especially in the winter. The input SO 2 emissions into two CMAQ simulations (M03 and M20) differed from SO 2 emissions of J-STREAM. For example, SO 2 emissions in both total and bottom layers of J-STREAM were more than twice those of M03 for d03, respectively. Meanwhile, for d04, including active volcanos, although the total SO 2 emissions of J-STREAM were half those of M03, the bottom layer SO 2 emissions of J-STREAM were 1.3 times those of M03. The differences in divided SO 2 emission amounts in the lower layers possibly affected the simulated atmospheric SO 2 concentrations. The second-best model setting, M03, performed slightly better (IoA: 0.41) than other models, which suggests that atmospheric SO 2 concentrations were considerably affected by the input emission conditions, including the injection heights. Although modifications of emission conditions help to produce better SO 2 simulation, using modifications alone to resolve the overestimation of SO 2 (up to 470%) is not realistic.
The differences among models with respect to emissions, chemistries, and meteorological conditions led to major differences in simulated primary pollutant concentrations; moreover, the simulated differences between similar model settings increased in the winter.  Tables 4 and 5 for each domain. The goal and criteria levels for CTM performance statistics, NMB, normalized mean error (NME), and correlation were recommended by Emery et al. [52], and the fractional bias (FB) and fractional error (FE) were recommended by Boylan and Russell [53], which is listed in Table A1. Individual model performance reports of each CTM are shown in Tables A2 and A3.      and IoA (0.54). However, M11 also produced low concentrations for SO 4 2− and NH 4 + . As observed for d04 in autumn, all models exhibited better performance for the daily concentration levels and day-to-day changes in NO 3 − . For example, M30 has a minimum bias of 0.14 µg/m 3 (NMB: 11%), which passed the goal NMB level for 24-h NO 3 − . Some deviations in NO 3 − between observations and the models were attributed to NH 4 + and potentially NH 4 NO 3 . In winter, most models reproduced day-to-day changes in both domains but tended to underestimate elevated NO 3 − levels, with ensemble mean biases of −0.89 µg/m 3 (NMB: −18.9%) and −2.36 µg/m 3 (NMB: −42.8%). A previous model inter-comparison study for the Tokyo metropolitan area, UMICS, concluded that the participant models overestimated NO 3 − levels in both summer and winter [11,12], although available observations included only one winter and three summer stations. In our validations, most models produced higher NO 3 − levels in spring and summer, lower NO 3 − levels in winter, and moderate NO 3 − levels in autumn, compared with accumulated observation data for d03 and d04. This result is expected to be more accurate than previous reports because a greater number of observations (for 18-22 stations) were included. As mentioned above, the day-to-day variations in NH 4 + were consistent with those of SO 4 2− and NO 3 − . Therefore, most CTMs showed good agreement with daily concentration levels and day-to-day changes in both domains for each season, with the exception of some elevated peaks. Above all, the ensemble performances indicators, FE and FB, were −27.9%-8.9% and 34.3%-41.1%, thus passing the goal level in both domains for all seasons except winter. Notably, the differences among models increased in summer. Two WRF-Chem models (M32, M31) predicted higher NH 4 + levels, with biases of 1.96-3.03 µg/m 3 (NMB: 84%-130%) and 1.72-2.71 µg/m 3 (NMB: 85%-61%) for d03 and d04, respectively. The M20 model, which employed EAGrid for emissions and an original configuration for meteorology, also produced relatively high NH 4 + levels in d03, with a bias of 1.68 µg/m 3 (NMB: 51%). These overpredictions were likely associated with those of SO 4 2− and NO 3 − in summer. Meanwhile, relatively larger negative biases were found for M11, at −1.18 µg/m 3 (NMB: 35%) for d03 and −0.81 µg/m 3 (NMB: 33%) for d04. The EC levels simulated by most CTMs were considerably lower than the observations in both domains for all season. The model ensemble biases were −0.90 to −0.20 µg/m 3 (NMB: −46% to −22%) and −2.77 to −0.39 µg/m 3 (NMB: −58% to −40%) for d03 and d04, respectively, with larger values for Tokyo. Both models employing EAGrid2000-JAPAN (M20 (d03) and M27 (d04)) produced higher EC values than other CTMs with different emission settings, and relatively better NMB values were obtained, at −20% to −3% and −35% to 42%, respectively. This trend suggests that the EC emissions of J-STREAM might be underestimated.
The CTMs reproduced some of elevated OC levels in the warmer seasons, but clearly underestimated the observed OC levels for autumn and winter, with model ensemble biases of −1.78 to −0.01 µg/m 3 (NMB: −42% to 7%) and −2.77 to −0.81 µg/m 3 (NMB: −59% to −39%) for d03 and d04, respectively, which are similar to the EC values. Additionally, as observed for the EC, the negative biases of OC for the Tokyo area were larger than those for western Japan. However, the negative biases of all participant CTMs have been clearly moderated compared with the UMICS cases [11,12]. Among the models, M02, M03, and M11 predicted relatively higher OC levels and overestimated the summer OC concentrations. Full-domain nesting simulations were performed via M02 and M03 using a relatively recent CMAQ model (Version 5.1), which includes updates for some chemical and aerosol mechanisms, such as POA aging, SOA mass yields with new pathways from isoprene, alkanes, and PAHs, and SOA formation reactions in the aqueous-phase chemistry. Continual nesting simulations for the Asian scale (d01) performed by CMAQ Version 5.1 exhibited higher regional-scale OC levels, leading to higher OC levels in urban areas in Japan compared with previous versions. Thus, an empirical SOA yield model can predict the same OC concentration level as the VBS model M11. It should be noted that effect of the updated SOA yield mechanisms was not clear at the urban scale when using CMAQ Version 5.1 or higher (e.g., M01, M04-05). Additionally, to evaluate simulated OC concentrations, more observational data are needed.
Overall, most CTMs showed good agreement with observed concentration levels of total PM 2.5 mass in both domains for each season. These results are likely associated with the reproducibility of some dominant components, e.g., SO 4 2− and NH 4 + . Moreover, CTMs tended to fail at reproducing some heavily polluted situations and underestimated the considerably high PM 2.5 concentrations (approximately 40-50 µg/m 3 ). A considerable underestimation (≈30 µg/m 3 ) of total PM 2.5 associated with PM 2.5 components, except for SO 4 2− , was observed for d04 in the winter season, 25 January and 2 February; during that time, the nighttime simulated surface temperature was clearly lower than that in the observations (Figure 3). This implies that the simulated higher surface temperature compared with that in the observations formed weaker atmospheric stability, which produced weaker accumulations of particulate pollutants at nighttime, especially during colder seasons. The model ensemble biases were −8.66 to −0.99 µg/m 3 (NMB: −43% to −5%) for d03 and −2.91 to −11.98 µg/m 3 (NMB: −55% to −19%) for d04. The largest negative biases are found in winter due to underestimations of NH 4 NO 3 , particularly for d04. M31 and M32 tended to overpredict the total PM 2.5 due to overestimates of inorganic compounds. Of the model ensemble statistics for d03, the NMB (−5%, 13%) NME (22%, 26%), FB (−9%, −17%), FE (26%, 29%), and correlation (0.81, 0.78) passed the goal level for 24-h total PM 2.5 mass in spring and summer, respectively. In addition, the majority of the other statistical indicators passed the criteria levels as well.

Summery
A model inter-comparison of secondary pollutant simulations over urban areas in Japan, J-STREAM Phase I, was performed, in which a total of 32 simulations were conducted by combining CMAQ, CAMx, and WRF-Chem.
Simulated hourly concentrations of the primary pollutants NO and NO 2 , which are precursors of PM 2.5 , generally showed good agreement with the observed concentrations, at the same level as the MICS case. However, some differences between observations and simulations and CTMs may be considered to be caused by the differences in meteorological conditions and NOx chemistries of each CTM. Furthermore, most of the CTMs using the same input emissions tended to overestimate SO 2 concentrations, although the models showed good performance for PM 2.5 SO 4 2− . The different emission inventory, EAGrid produced better results for SO 2 ; therefore, it appears that the emission input can be improved. However, it was likely to be unrealistic that just the modifications of the emissions could fully resolve the overestimation of SO 2 . Simulated concentrations of PM 2.5 and its components were evaluated via a comparison with daily observed concentrations by using the filter pack method at selected AAPMSs for a period of at least two weeks for each season in this project. In general, most of the models showed good agreement with the observed concentration of total PM 2.5 mass for each season, within goal or criteria levels of model ensemble statistics especially in warmer seasons. This agreement was associated with the reproducibility of some of dominant particulates.
Among  [11,12]. This difference between two model inter-comparison studies is attributed to variations in the number of observations applied for verification. Thus, a sufficient amount of observation data on PM 2.5 components is needed to evaluate and improve CTMs. The EC levels simulated by most models were considerably lower than the observed levels for all seasons; however, some models employing EAGrid emissions produced higher EC levels than the other models. The models reproduced concentrations for some elevated OC values in the warmer seasons, but clearly underestimated the OC levels in autumn and winter. In addition, some models employing the VBS model and the newly updated SOA yield mechanisms produced higher OC levels and even overestimated the observed OC concentration in some cases. This study has identified some effective approaches for improving PM 2.5 simulations for urban areas in Japan based on a model inter-comparison. First, improvements in emissions are expected to increase the reproducibility of primary pollutants that are precursors of PM 2.5 and EC concentrations. For SO 4 2− , NO 3 − , and OC, additional formation pathways can help to reduce underestimations. The recent model updates (e.g., CMAQ Version 5.3) improved the chemical pathways and are expected to simulate the secondary PM 2.5 components well. Simulated meteorological fields will be important for the Asian scale PM 2.5 concentration levels and the elevated PM 2.5 concentrations during the days with high amounts of pollution. In addition, special attention is needed for misjudgments in these models. Finally, additional accumulated observations are needed to evaluate the simulated concentrations. Future studies will include these modifications to realize reference air quality modeling in the next stages of J-STREAM.