Advantages of Continuous Monitoring of Hourly PM2.5 Component Concentrations in Japan for Model Validation and Source Sensitivity Analyses

Continuous monitoring of hourly PM2.5 component concentrations has been performed in Japan. The objective of this study was to evaluate the advantages of continuous monitoring to obtain data that can be useful for regional air quality simulations. Inclusion of transboundary transport in the simulations improved the correlation between the observed and simulated hourly concentrations of SO42−, SO3−, secondary organic aerosols (SOA), and metals in PM2.5. Black carbon was an exception, suggesting the overestimation of emissions in upwind countries. Including volcanic and dust emissions also improved the correlations between the observed and simulated hourly concentrations of SO42− and metals, respectively. However, despite the good correlation achieved by including transboundary transport, it also resulted in overestimated NO32− and SOA concentrations in western Japan during the winter. Further improvements are necessary, such as balancing with SO42− and the dry deposition of gaseous HNO3 for NO3−, and new treatment of the partitioning and aging of semivolatile organic aerosols, which have been incorporated into recent models for SOA. The differences in model performance with regard to simulating metal concentrations suggest imbalances in the speciation profiles used for countries other than Japan. Further, comparing the observed and simulated hourly concentrations helped identify the key processes driving air quality. This revealed evening peaks in black carbon concentrations, owing to the relatively stable atmosphere; and early morning peaks in NO3− concentration, owing to the low temperature and high humidity through thermodynamic equilibrium. This study demonstrated that continuous monitoring of hourly variations in PM2.5 composition is valuable for understanding the roles of the emission sources and for improving future models, both of which contribute to deriving effective PM2.5 suppression strategies.


INTRODUCTION
Fine particulate matter smaller than 2.5 μm (PM 2.5 ) adversely affects human health (Kojima et al., 2020).The Environmental Quality Standard for PM 2.5 was introduced in Japan in 2009.Despite various strategies that have been implemen ted, ambient PM 2.5 concentrations still exceed this standard at some of the monitoring stations in Japan (Ministry of the Environment, 2020).
PM 2.5 consists of multiple components: its primary components are emitted directly from the source as particulates, while secondary components form in the atmosphere from gaseous precursors via photochemical reactions.There are many different emission sources of primary components, and gaseous precursors of secondary components.Therefore, it is essential to clarify the PM 2.5 composition to develop effective mitigation strategies for reducing emissions from key contributing sources.
In addition to the routine monitoring of PM 2.5 mass concentrations at general monitoring stations, the Ministry of the Environment of Japan (MOEJ) has initiated a PM 2.5 composition monitoring campaign in cooperation with local governments.As part of this campaign, daily variations in the concentrations of PM 2.5 components collected on filters are measured for two weeks in each sea son.During the 2018 fiscal year, monitoring was conducted at 179 locations throughout Japan (Ministry of the Environment, 2020).These measurement data have been utilized in various scientific studies, including validation of regional air quality simulations (Itahashi et al., 2020b;Yamaji et al., 2020;Itahashi et al., 2018a;Uranishi et al., 2017).Such simulations are useful for investigating the relationships between emission sources and ambient PM 2.5 concentration, taking photochemical reactions into account.However, one of the disadvantages of this monitoring campaign is that it is restricted to only two weeks per season, with daily averaged samples.Therefore, the data may not reflect the typical air quality during the respective seasons.Moreover, the variations in the daily averages may be insufficient to fully understand the dyna mic behavior of individual PM 2.5 components.
To overcome these disadvantages, the MOEJ has begun automated continuous monitoring of hourly PM 2.5 component concentrations.It has several advantages.Hourly variations obtained in this monitoring are helpful for understanding the dynamic behavior of PM 2.5 components.Using a short sampling duration is effective at suppressing any loss of PM 2.5 components caused by evaporation from the samples.Continuous monitoring does not miss elevated concentrations of PM 2.5 components that may occur throughout the year.In addition, some of the components identified in this monitoring are not available in conventional monitoring data.Monitoring con ducted at multiple locations also enables the under-standing of spatial variations in PM 2.5 components.The objective of this study was to evaluate how these advantages of the continuous monitoring of PM 2.5 compositions are effective for regional air quality simulations.This data has previously been analyzed by only a few studies, and for limited time periods (Itahashi et al., 2020a).Herein, we conducted the first comprehensive analysis of this data for the entire year.We validated hourly PM 2.5 component concentrations simulated by three different regional chemical transport models in comparisons with those obtained in the monitoring.In addition, we investigated the influences of key emission sources on hourly variations in the observed and simulated PM 2.5 component concentrations.Through these investigations, we evaluated the effectiveness of the continuous monitoring of PM 2.5 compositions.

1 Model Configuration
Uncertainties are present in input data processing and in model execution.In addition, different capabilities emb edded in models could produce large variations in simulated PM 2.5 concentrations (Yamaji et al., 2020).Therefore, we used three different regional chemical trans port models, namely the Community Multiscale Air Quality (CMAQ) modeling system (Byun and Schere, 2006) v. 5.2.1 and 5.0.2 (hereafter CMAQv5.2.1 and CMAQv5.0.2), and the Comprehensive Air Quality Model with Extensions (CAMx) (Ramboll Environ, 2016) v. 6.40, to evaluate the influences of uncertainties and different capabilities embedded in the three models.The SAPRC07 (Carter, 2010) chemical mechanism is common to all three models.The AERO6 and coarse/fine (CF) aerosol schemes were used in CMAQ and CAMx, respectively.
Figure 1 shows the target study domains, with d02 nested inside d01.These domains are commonly used in model comparisons in Japan's Study for Reference Air Quality Modeling ( J-STREAM) project (Chatani et al., 2018).The larger domain (d01) covers Asian countries with a 45 km × 45 km grid and provides boundary concentrations for d02.The boundary concentrations for d01 were derived from the chemical atmospheric general circulation model for the study of the atmospheric environment and radiative forcing (CHASER) (Sudo et al., 2002).The nested domain (d02) covers most of Japan with a 15 km × 15 km grid.Thirty layers were set vertically from the ground to 5,000 Pa.Hourly variations in meteorological fields and emissions were considered in the simulations.Details of the meteorological and emission inputs are described in Chatani et al. (2020a).Simulations were performed for the period from March 1, 2017, to March 31, 2018.The results for the first month were discarded as spin-up.
The base simulation, using all emission inputs, was performed using the three selected models.In addition, the source sensitivities of the simulated hourly PM 2.5 component concentrations were investigated.Three emission sources, namely volcanoes, dust, and anthropogenic sources in countries other than Japan (hereafter referred to as transboundary transport), were selected for this study.These isolated or distant sources are appropriate for investigating the drivers of change in hourly PM 2.5 component concentrations because their effects are apparent only when the airmass that was present at the distant source travels over the monitoring site.Source sensitivities were calculated using the Brute-Force Method (Chatani et al., 2020a;Clappier et al., 2017).The emissions of the target sources were subtracted entirely in the additional sensitivity simulations performed using CMAQv5.2.1.Differences in the simulated concentrations between the base and sensitivity simulations were evaluated as sensitivities to the corresponding sources.

2 Monitoring Data
In the continuous monitoring conducted by the MOEJ, the hourly concentrations of PM 2.5 components are continuously measured using two automatic instruments, a Continuous Dichotomous Aerosol Chemical Speciation Analyzer (ACSA-14; KIMOTO ELECTRIC Co., Ltd) and a Continuous Particulate Monitor with X-ray Fluorescence (PX-375; HORIBA Ltd.).ACSA-14 provides hourly mass concentrations of particulate matter (PM) and its components, including H + , SO 4 2-, NO 3 -, and water-soluble organic compounds (WSOCs) in the fine and coarse fractions, and optically measured black carbon (OBC) in the fine fraction.PM 2.5 mass concentrations, equivalent to those used in the Federal Register Method (FRM), are also generated.ACSA-14 and its earlier versions have been used in studies conducted in China (Pan et al., 2019;Ye et al., 2019) and Japan (Itahashi et al., 2017;Uno et al., 2017a, b).ACSA-14 has been installed by MOEJ at ten locations in Japan (Fig. 1).PX-375 provides hourly mass concentrations of the metals in PM 2.5 , including Ti, V, Cr, Mn, Fe, Ni, Cu, Zn, As, Pb, Al, Si, S, K, and Ca.PX-375 has been used in scientific studies conducted in China (Li et al., 2017) and Japan (Asano et al., 2017).PX-375 has been installed by the MOEJ at four locations in Japan (Fig. 1).Flags have been given to the data which needs careful treatments.The data containing measurement errors, maintenance, calibration, and contamination, as well as independent outliers, were excluded in this study.
It is important to note that the methods used to classify particles into fine and coarse fractions are not consistent between ACSA-14 and PX-375, which use a virtual imp- actor and a FRM impactor, respectively.Although both are intended to classify PM 2.5 , Nakatsubo et al. (2018) reported differences between the PM 2.5 mass concentrations classified by ACSA-14 and FRM.Further, the methods used to represent particle size distributions are not consistent between CMAQ and CAMx.CMAQ uses three modal size distributions and PM 2.5 is represented by the Aitken and accumulation modes (Binkowski and Roselle, 2003).CAMx uses two discrete bins for coarse and fine particles, and the latter represents PM 2.5 .These different treatments and particle evolution may result in differences in PM 2.5 concentrations simulated by them.It is difficult to maintain strict particle size consistency among the observed and simulated values.We compared the simulated values of the Aitken and accumulation modes of CMAQ and the "fine" bin of CAMx, with the observed PX-375 and ACSA-14 (fine fraction) values.CMAQ-simulated values of the coarse mode were compared with the observed ACSA-14 coarse-fraction values.Although using inconsistent size classifications might result in data gaps, our discussion focused mainly on the correlation between the observed and simulated concentrations.

RESULTS AND DISCUSSION
Figure 2 shows the observed and simulated monthly mean concentrations of PM 2.5 , PM in the coarse fraction (PMc), OBC, SO 4 2-, NO 3 -, and WSOC in the fine fraction (OBC, fSO4, fNO3, and f WSOC, respectively), and SO 4 2-and NO 3 -in the coarse fraction (cSO4 and cNO3, respectively), along with the monthly mean source sensitivities simulated by CMAQv5.2.1.The values were averaged over all hours of each month at the ten ACSA-14 stations.The corresponding values at each ACSA-14 station are shown in Fig. S1 in the Supplementary Materials.The simulated concentrations of elemental carbon (EC) were compared with the observed OBC concentrations.WSOC is a marker of secondary organic aerosols (SOAs) (Kondo et al., 2007;Miyazaki et al., 2006).The simulated concentrations of secondary organic carbon (SOC) were compared with the observed WSOC concentrations.
Distinct seasonal variations in PM 2.5 concentrations were observed (Fig. 2a), with higher concentrations in the spring and lower concentrations in the summer and early autumn.Although these seasonal variations were reproduced well by the models, the absolute values were underestimated.Among the three models, the CMAQv5.2.1simulated values were the closest to the observed value.The underestimation was more evident in eastern Japan, as indicated by Chatani et al. (2020a).The PM 2.5 concentrations simulated by CMAQv5.2.1 were higher than the observed values at Fukuoka, which is located in western Japan (Fig. S1a).The simulated sensitivity to transboundary transport corresponded to 30-80% of the simulated PM 2.5 concentrations.
The models performed well in terms of simulating PMc concentrations, except in the spring, during which higher peaks were observed (Fig. 2b).The simulated sensitivities of PMc concentrations to volcanic emissions, dust, and transboundary transport were relatively small.The influence of other emission sources, including sea salt, seemed to be larger, given that Na and Cl were the most influential components in the simulated PMc concentrations (not shown).
The statistical model performances for the entire target period are presented in Table 1 including the annual mean observed and simulated concentrations, normalized mean bias (NMB), normalized mean error (NME), and cor rela tion coefficients (Emery et al., 2017).The NME and correlation coefficients were calculated from the daily and hourly concentrations (NMB is identical, by definition, whether calculated from the daily or hourly concentrations).The NME and correlation coefficients calculated from the daily PM 2.5 concentrations simulated by the three models satisfied the criteria proposed by Emery et al. (2017).However, using the hourly concentrations reduced the model performance.This indicates that it is more challenging for models to reproduce hourly concentrations of PM and its components.
Figure 3 shows monthly variations in the correlation coefficients calculated from the hourly concentrations of PM and its components at all ten ACSA-14 stations, derived from the CMAQv5.2.1 base and sensitivity simulations.The corresponding values at each ACSA-14 station are shown in Fig. S2 in the Supplementary Materials.The correlation coefficients of PM 2.5 and PMc were higher in the base simulation than in the sensitivity simulations (Fig. 3a, b).This implies that volcanic emissions, dust emissions, and transboundary transport improved the correlation between the observed and simulated hourly PM 2.5 and PMc concentrations.
The model performances and source sensitivities of PM 2.5 and PMc reflected those of their components.These are discussed in detail in the following subsections.

1 OBC
OBC concentrations were overestimated by the three models throughout the target period (Fig. 2c).Sensitivity to transboundary transport corresponded to 24-58% of the simulated OBC concentrations.It must be noted that the OBC concentrations obtained by ACSA-14 tend to be lower than the EC concentrations.Nakatsubo et al. (2018) reported that OBC concentrations obtained by ACSA-14 were approximately 40% lower than EC concentrations in their comparison.Therefore, the absolute magnitude of the overestimation in this study may have been affected by the performance of ACSA-14 for OBC.Regardless, the base simulation yielded lower correlation coefficients than the sensitivity simulation without transboundary transport (Fig. 3c).Including transboundary trans port reduced the correlation between the observed and simulated OBC concentrations.In contrast, lower correlation coeffic ients in the base simulation are not evident at each  station(Fig.S2c).Thus, the inclusion of trans boundary transport redu ced the spatial correlations among the ten ACSA-14 stations, whereas it contributed to better temporal correlations, particularly at stations located in remote areas of western Japan such as Oki and Goto.As shown in Fig. S1c, the base simulation had a better performance at the stations located in urban areas, including Sapporo, Tokyo, Nagoya, and Osaka.Overestimation mainly occu rred in remote areas when influences of transboundary transport were large.The inclusion of transboundary transport resulted in spatial inconsistencies in model performance between urban and remote areas.Figure 4 compares the observed hourly OBC concentrations with the simulated concentrations generated by the base (three models) and sensitivity simulations; as an example, the results in November 2017 and February 2018 are shown for Nagoya.The absolute values and hourly variations observed in November 2017 were reproduced well by the three models (Fig. 4a).Based on hourly concentrations, OBC peaked mainly in the evenings when atmospheric conditions are relatively stable.CAMx simulated higher OBC concentrations than CMAQ at peak hours.Chatani et al. (2014) indicated that simulated EC (OBC) concentrations were greatly affected by the treatment of vertical diffusion.Although both CMAQ and CAMx use the Asymmetric Convective Model v. 2 (ACM2) vertical diffusion scheme (Pleim, 2007), CMAQ forces vertical diffusion coefficients to be greater than 1.0 m 2 /s in urban areas, including Nagoya.The higher OBC concentrations simulated by CAMx were likely caused by the different lower limits of the vertical diffusion coefficients in urban areas under the stable atmosphere.To improve the model performance for OBC concentrations at peak hours, it is necessary to improve the treatment of vertical diffusion.
Throughout February 2018, the simulated OBC concentrations were slightly higher than the observed values (Fig. 4b).There were distinct gaps between the OBC concentrations obtained in the base simulation and the sensitivity simulation without transboundary transport, indicating that transboundary transport caused the overestimation, as well as a poorer correlation.This was particularly evident in the peak on February 24.The results suggest that BC emissions in countries other than Japan have been overestimated.Kanaya et al. (2020) indicated that their simulation tended to overestimate the BC concentrations at Fukue Island, located between Japan and the continent, when the airmass was transported from China and South Korea.et al., 2015), used in this study, overestimated the BC/ CO ratios of the emissions in China and South Korea.The results obtained in this study were consistent with their findings.However, it must also be noted that February 2018, which had larger discrepancies between the obser ved and simulated OBC concentrations (Fig. 2c), had the lowest correlation coefficients (0.48), even in the sensiti vity simulation without transboundary transport (Fig. 3c).Both the decrease in OBC originating from transboundary transport and the increase in OBC originating from local emissions are expected to improve spatial correlations.

2 SO 4 2-
The statistical model performance for fSO4 was comparable to that for PM 2.5 (Table 1), satisfying the criteria proposed by Emery et al. (2017).The observed fSO4 concentrations were higher during the first half of the target period (spring and summer; Fig. 2d).Their absolute values and monthly variations were reproduced well by the three models.In contrast, during the second half of the target period (autumn and winter), these values were underestimated by the three models.Previous studies (Itahashi et al., 2018b;Morino et al., 2015) also reported the underestimation of winter SO 4 2-concentrations around Japan.The models may need to be updated by including additional aqueous-and gaseous-phase oxidation pathways, to improve their performances for winter SO 4 2-values (Itahashi et al., 2019(Itahashi et al., , 2018a)).Itahashi et al. (2018b) indicated that differences between CAMx and CMAQv5.0.2 regarding how dry deposition is represented resulted in CAMx simulating higher SO 4 2-concentrations.The opposite pattern that we obtained in this study may be due to another difference: we used CMAQ v5.0.2 calculating photolysis rates online, whereas CAMx picked them up from a lookup table.Differences in photolysis rates could affect the secondary formation of relevant species in the atmosphere, including SO 4 2- (Chatani et al., 2020b;Itahashi et al., 2020b).
For fSO4, the simulated concentration sensitivities to transboundary transport and volcanic emissions were notable, at 32-67% and 2-25% of the simulated concentrations, respectively (Fig. 2d).Throughout the target period, the correlation coefficients in the sensitivity simulation without transboundary transport were notably lower than those in the base simulation (Fig. 3d).This clearly reflects the contribution of transboundary transport to improving the correlation between the observed and simulated fSO4 concentrations.The contribution of volcanic emissions was also evident from May to August.
Figure 5 compares the observed hourly fSO4 concentrations with the simulated concentrations derived from the base (three models) and sensitivity simulations, using Osaka in May and July 2017 as an example.Hourly variations in observed fSO4 concentrations were reproduced well by the three models.The fSO4 concentrations simulated in the sensitivity simulation without transboundary transport had distinctly lower peaks than those simulated in the base simulation on May 1, 12, 20, and 30 (Fig. 5a), clearly illustrating the influence of transboundary transport on these days.Likewise, the fSO4 concentrations simulated in the sensitivity simulation without volcanic emissions had distinctly lower peaks on July 3, 12, and 16 (Fig. 5b), clearly illustrating influence of volcanic emissions on these days.
In contrast, the models performed worse in simulating cSO4 concentrations (Table 1).While the observed and simulated annual cSO4 concentrations were comparable, the NME was large and the correlation coefficients were quite low (Fig. 3g).The simulated sensitivities of cSO4 concentrations to the three emission sources were small (Fig. 2g).This implies that the factors affecting cSO4 were entirely different from those affecting fSO4.Osada et al. (2016) reported low correlations between cSO4 concen trations measured using the ACSA and those mea sured by collection on filters.These findings suggest that there are large uncertainties in both the observed and simulated cSO4 concentrations.

3 NO 3
-There were distinct seasonal variations in the f NO3 concentration (Fig. 2e).Observed fNO3 concentrations were much higher during the winter than the summer.This seasonal variation was reproduced well by the models.However, the higher winter f NO3 concentrations were overestimated by the CMAQ models, particularly in western Japan, including Oki, Goto, and Fukuoka (Fig. S1e).Most of the overestimation was due to transboundary transport, even though it improved the correlation during the winter (Fig. 3e and Fig. S2e).The overestimation of summer NO 3 -concentrations, which has been previously reported (Shimadera et al., 2014), was not evident in this study.The observed cNO3 concentrations did not exhibit distinct seasonal variations (Fig. 2h).In contrast to fNO3, the models overestimated cNO3 throu ghout the target period.Transboundary transport imp-roved the cNO3 correlations (Fig. 3h).
Figure 6 compares the observed hourly fNO3 concentrations with the simulated concentrations derived from the base (three models) and sensitivity simulations, using February 2018 in Tokyo (eastern Japan) and Fukuoka (wes tern Japan), as examples.At Tokyo, while the observed f NO3 concentrations were relatively low on most days, there were peaks on some days (Fig. 6a).The models did not satisfactorily reproduce some of these peaks (those on February 3 and 11).Itahashi et al. (2020b) conducted a detailed analysis of the model performance for winter NO 3 -concentrations in Tokyo, finding that NO X and NH 3 emissions, heterogeneous reactions involving HONO, and photolysis rates were critical factors that affect model performance in terms of simulating NO 3 concentration peaks.Unlike Tokyo, at Fukuoka (western Japan) many of the peaks were higher, albeit flatter, and all were significantly overestimated by the CMAQ models (Fig. 6b).The fNO3 concentrations and correlation coefficients based on the sensitivity simulation without transboundary transport were substantially lower than those in the base simulation, indicating that they were exclusively and excessively affected by transboundary transport.Uno et al. (2020) indicated that the recent reduction in SO X emissions realized by the stringent emission controls impl e mented in China caused a paradigm shift and inc reased NO 3 -transport to western Japan, particularly during colder seasons (Itahashi et al., 2017).Although their findings are related to our simulated results, the influence of transboundary transport on the fNO3 concentration was too large.As mentioned, the winter fSO4 concentrations were underestimated by the models used in this study.When more fSO4 is available, the simulated fNO3 concen trations should decrease, according to the paradigm discussed by Uno et al. (2020).The possible impacts of imp roved model performance on winter fSO4 concentrations (discussed in section 3.2), are also important for addressing the problem of overestimated fNO3 concentrations.Further, the CAMx-simulated peak values were substantially lower than those simulated by CMAQ.It is known that the inaccurate estimation of gaseous HNO 3 dry deposition velocity can cause the overestimation of NO 3 -concentrations (Itahashi et al., 2017;Morino et al., 2015;Shimadera et al., 2014).CMAQ and CAMx use different dry deposition schemes, M3DRY (Otte and Pleim, 2010) and that of Zhang et al. (2003), respectively, which might have caused the differences in the fNO3 concentrations simulated by CMAQ and CAMx.More accurate methods of representing the dry deposition of gaseous HNO 3 are therefore required.
Figure 7a compares the observed hourly fNO3 concentrations and the simulated hourly concentrations derived from the base (three models) and sensitivity simulations, for Tokyo in July 2017, when the fNO3 concentrations and their correlation were remarkably low (Figs.2e and  3e). Figure 7b shows the observed and simulated hourly cNO3 concentrations, and Figure 7c shows the cumulative concentrations of cNO3, fNO3, and stoichiometrically equivalent gaseous HNO 3 , simulated by CMAQ v5.2.1, along with temperatures and relative humidities in the corresponding month and location.The observed fNO3 concentrations were mostly below 1 μg/m 3 .There were peaks in the simulated fNO3 values on some days.These overestimated peaks greatly reduced the correlation between the observed and simulated fNO3 concentrations.Although the observed and simulated cNO3 con centrations were higher at the corresponding hours, their variations were smaller than those of f NO3, and were reproduced relatively well by the models.
The simulated gaseous HNO 3 concentrations were higher during the daytime due to photochemical formation induced by solar radiation (Fig. 7c).However, the ISORROPIA phase equilibrium model (Fountoukis and Nenes, 2007) embedded in CMAQ and CAMx did not par tition the gaseous HNO 3 into f NO3 and cNO3 becau se of the higher daytime temperature and lower rela tive humidity.The simulated f NO3 peaks occured mostly in the early morning, when the temperature was low and humidity was high, and when ample gaseous HNO 3 remained.The strong dependency of the simulated partitioning of NH 4 NO 3 , a major component of f NO3, on temperature and humidity, as well as on the concentrations of the associated species, causes instability and rapid variations in the simulated f NO3 concentrations (Nenes et al., 2020).It is challenging for models to reproduce hourly variations in all of the related factors.Moreover, it is necessary to validate hourly variations of gaseous HNO 3 to check if the simulated dynamic behaviors surely occur in the real atmosphere.
However, the cNO3 concentrations were appropriately simulated.In the coarse fraction, NO 3 -appears to be present in more stable forms, including as NaNO 3 (Kajino et al., 2013;Dasgupta et al., 2007).Further, the dynamic mass transfer considered in the models requires a longer time for the coarse fraction particles to reach equilibrium (Kelly et al., 2010).These factors may have resulted in smaller variations and and overall better model performance regarding cNO3 concentrations.

4 WSOC
As SOAs form via photochemical reactions, the SOC concentrations simulated by CMAQv5.0.2 and CAMx were high during July (Fig. 2f) when photochemical reac tions are active.Although the July SOC concentrations simulated by CMAQv5.0.2 were comparable to the observed f WSOC concentrations, those in other months were significantly underestimated by the model.The SOA concentrations simulated by CMAQv5.2.1 were much higher than the observed f WSOC concentrations and the SOC concentrations simulated by CMAQv5.0.2 and CAMx in the corresponding months.The increase in SOC concentrations from December to March simulated by CMAQv5.2.1 was greatly affected by transboun-dary transport, which contributed to a better correlation between the observed and simulated values (Fig. 3f).The contribution of transboundary transport was smaller during the spring and summer.
The observed f WSOC concentrations were lower in summer, implying that their seasonal variations cannot be explained simply by the strength of photochemical reactions.Another potential factor is biomass burning emissions, which also contribute to WSOC (Ikemori et al., 2021;Yan et al., 2015).However, the absolute magnitude and seasonal variations of the observed WSOC concentrations require careful attention.Saito et al. (2020) compared the f WSOC concentrations measured by ACSA-14 and the conventional method.They found seasonal varia- tions in the slopes and intercepts of the regression lines between them.The discrepancies were larger in the summer than the winter, possibly caused by seasonal variations in the WSOC compositions.In addition, components including levoglucosan, which is a well known marker for biomass burning emissions, cannot be detected by ACSA-14 using ultraviolet spectrophotometry.Additional research is necessary to clarify the relationships among f WSOC concentrations measured by ACSA-14 and by the conventional method, as well as SOA concentrations simulated by the models.
CMAQ v. 5.2 and later versions have incorporated partitioning and aging of semi-volatile organic aerosols (Robinson et al., 2007;Donahue et al., 2006).Along with this treatment, they have introduced a new surrogate species, which is a potential SOA from combustion emissions (pcSOA).pcSOA represents the SOA that was missed in the preceding models for various reasons, including SOA formation from intermediate-volatility organic compound (IVOC) emissions, which are not included in current emission inventories (Murphy et al., 2017).In CMAQv5.2.1, an amount equivalent to 6.6 times the primary organic aerosol (POA) emissions is ingested as missing emissions of pcVOC (precursors of pcSOA), which are subsequently oxidized in the atmo-sphere to form pcSOA.In our simulations, pcSOA occupied a large fraction of the simulated SOC concentrations (not shown), leading to overestimation and differences in seasonal variations of the simulated SOC concentrations (Fig. 2f).
Figure 8 compares the observed hourly f WSOC concentrations with the simulated SOC concentrations derived from the base (three models) and sensitivity simulations, using Tokyo in March 2018 and July 2017 as an example.The SOA concentrations simulated by CMAQ v5.2.1 significantly overestimated the broad increases in the f WSOC concentrations on March 11-15 and 25-31 (Fig. 8a).For the same periods, the SOA concentrations simulated in the sensitivity simulation without transboundary transport were much lower, indicating that the overestimation was caused mainly by transboundary transport.It is worth noting that the correlation coefficient in the base simulation (0.79) was higher than that in the sensitivity simulation without transboundary transport (0.67).Part of the increase in observed f WSOC concentrations likely reflects SOA formed from IVOC emissions in upwind countries.In contrast, the SOC concentrations simulated in the base and sensitivity simulations without transboundary transport were similar; both simulations overestimated the increases on July 7-9 and 16-20 (Fig. 8b).The SOC concentrations simulated by CMAQv5.0.2 were closer to the observed values in the corresponding period.The later versions of CMAQ did not perform well for modeling summer pcSOA formed from IVOC emissions in Japan.
The emission ratio of pcVOC to POA is an uncertain parameter in the treatment incorporated in CMAQv5.2.1 (Murphy et al., 2017).While its value (6.6) has been validated in comparisons with observations conducted in the United States, this value may not be suitable for Asia, where POA emissions are relatively high because there are fewer emission controls.Suitable values for the pcVOC/ POA ratio need to be examined based on the current conditions in Asia (Cai et al., 2019;Morino et al., 2018;Liu et al., 2017).In addition, it may not be appropriate to apply a single pcVOC/POA ratio value to emissions from all sources, including biomass burning (Murphy et al., 2017).However, in CMAQv5.2.1,only a single pcVOC/POA ratio can be used, because the model accepts only one emission input involving all emission sources.The subse-quent CMAQ version (v.5.3) has realized the incorporation of multiple emission inputs and their respective pcVOC/POA ratios.

5 Metals
Figure 9 shows the observed and simulated monthly mean concentrations of Ti, Mn, Fe, Al, Si, K, and Ca in the fine fraction, and the monthly mean source sensitivities simulated by CMAQv5.2.1, averaged over all hours in each month at four PX-375 stations.The corresponding values at each PX-375 station are shown in Fig. S3 in the Supplementary Materials.Figure 10 shows the monthly variations in the correlation coefficients calculated from the hourly metal concentrations at all the four PX-375 stations, derived from the CMAQv5.2.1 base and sensitivity simulations.The corresponding values at each PX-375 station are shown in Fig. S4 in the Supplementary Materials.
The model performance differed depending on the metal.While the concentrations of Mn, Fe, Al, and Si were all underestimated, K and Ca concentrations were overestimated.The influence of transboundary transport was apparent for all metals.The correlation coefficients of the base simulation were distinctly higher than those of the sensitivity simulation without transboundary transport.Transboundary transport contributed significantly to the simulated metal concentrations and improved the correlations.Itahashi et al. (2018a) described the importance of Mn and Fe in SO 4 2-formation; therefore, PM speciation profiles that include Mn and Fe, as proposed by Fu et al. (2013), have been used for countries other than Japan.Although Fu et al. (2013) established profiles for 24 emission sectors, these sector profiles need to be merged into the four sectors that are available in the HTAP emissions used in this study.Differen ces in the profiles for different countries, regions within countries, and particular sectors with regions, were not con sidered.In this study, the rough treatment of the metal profiles in the PM may have reduced the model per formance for metals.To address this problem, a detailed data -base of metal emissions is required for Japan and other countries (Kajino et al., 2020).
The similarity of the physical processes embedded in CMAQv5.2.1 and CMAQv5.0.2 resulted in similar simulated metal concentrations.However, the metal concentrations simulated by CMAQv5.2.1 for May were much higher than those simulated by CMAQv5.0.2.CMAQ v5.2.1 implements the windblown dust emission parameterization of Foroutan et al. (2017), whereas dust emissions are not considered by CMAQv5.0.2 used in this study.Differences in the simulated metal concentrations were caused by dust emissions, as indicated by the simulated sensitivity toward them (Fig. 9), and by better correlations in the base simulation than the sensitivity simulation without dust emissions (Fig. 10).
Figure 11 compares the observed hourly Fe and Ca con centrations with the simulated concentrations derived from the CMAQv5.2.1 and CMAQv5.0.2 base and sensitivity simulations, using Fukuoka in May 2017 as an exam ple.A heavy dust storm in east Asia at the beginning  of May 2017 (Zhang et al., 2018) caused distinct peaks in Fe and Ca concentrations on May 6-9.Although these peaks were not fully reproduced, the Ca concentrations from the base simulation were clearly higher than those from the sensitivity simulation without dust emissions.For Fe, the differences between the base and sensitivity simulations were marginal, although visible.The parameterization embedded in CMAQv5.2.1 satisfactorily reproduced some of the Fe and Ca peaks that originated from dust emissions.The peaks in Ca concentrations were overestimated after the dust storm.They were exclusively affected by transboundary transport, as indicated by the significantly low concentrations obtained in the sensitivity simulation without transboundary transport.Improved metal emissions and speciation profiles in other countries are required to improve the model performance.

CONCLUSIONS
This study evaluated the advantages of the continuous monitoring of PM 2.5 composition conducted by MOEJ for regional air quality simulations.Continuous monitoring contributed greatly in two respects: in identifying drivers of hourly variations in PM 2.5 component concentrations, and in estimating the influences of emission sources.Regarding hourly variations, the observed and simulated concentrations clearly revealed increased evening OBC concentrations due to the relatively stable atmosphere, and increased early morning fNO3 concentrations due to low temperature and high humidity, as well as remaining gaseous HNO 3 .These factors cannot be distinguished when discrete daily concentrations are used.Continuous monitoring therefore helps in elucidating the validity of the mechanisms leading to simulated high PM 2.5 component concentrations and in improving model performance.However, it is also important to recognize the influences of uncertainties in the monitoring, particularly for components such as cSO4, OBC, and WSOC.
In terms of estimating the influences of the emission sources, we focused on dust emissions, volcanic emissions, and transboundary transport.These isolated or distant emission sources increase PM 2.5 component concentrations, but only when the airmass originates at the location of the distant source.Therefore, their influences are clearly reflected in hourly variations in concentrations.This study illustrates that regional air quality simulations can reproduce hourly variations caused by targeted emission sources.In other words, hourly variations in observed concentrations of PM 2.5 components are likely to be affected by transport from these emission sources.Using regional air quality simulations, this study reveals that continuous monitoring data is useful for identifying key processes and emission sources that increase PM 2.5 component concentrations.Such data is likely to have various other purposes, including receptor modeling.Continuous monitoring can contribute to the development of effective strategies for suppressing ambient PM 2.5 .It is therefore valuable to maintain and expand the continuous monitoring of PM 2.5 compositions.

Fig. 1 .
Fig. 1.Target domains (d02 nested inside d01) of the simulations performed in this study.The monitoring station sites are indicated by red and blue dots in d02.
Choi et al. (2020) implied that the existing emission

Fig. 4 .
Fig. 4. Observed and simulated hourly fine-fraction optically measured black carbon (OBC) at Nagoya, Japan, in (a) November 2017 and (b) February 2018.Simulated values are based on the three models used for the base and sensitivity simulations.Models used: CMAQ v. 5.2.1, CMAQ v. 5.0.2, and CAMx.w/o: without.

Fig. 5 .
Fig. 5. Observed and simulated hourly fine-fraction SO 4 2-(fSO4) concentrations at Osaka, Japan, in (a) May 2017 and (b) July 2017.Simulated values are based on the three models used in the base and sensitivity simulations.Models used: CMAQ v. 5.2.1, CMAQ v. 5.0.2, and CAMx.w/o: without.

Fig. 6 .
Fig. 6.Observed and simulated hourly fine-fraction NO 3 -(f NO3) concentrations in February 2018 at (a) Tokyo and (b) Fukuoka.Simulated values are based on the three models used in the base and sensitivity simulations.Models used: CMAQ v. 5.2.1, CMAQ v. 5.0.2, and CAMx.w/o: without.

Fig. 7 .
Fig. 7. Observed and simulated hourly concentrations of (a) fine-fraction and (b) coarse-fraction NO 3 -(f NO3 and cNO3, respectively), based on the three models used in the base and sensitivity simulations, for Tokyo in June 2017.(c) Cumulative hourly concentrations of cNO3, f NO3, gaseous HNO 3 (simulated by CMAQ v. 5.2.1), temperature, and relative humidity, in the base and sensitivity simulations.Models used: CMAQ v. 5.2.1, CMAQ v. 5.0.2, and CAMx.w/o: without.

Fig. 9 .
Fig. 9. Observed and simulated monthly mean concentrations and source sensitivities of Ti, Mn, Fe, Al, Si, K, and Ca in the fine fraction.Values were averaged over all hours in each month for the four stations.Measurements were performed using PX-375.Concentrations are denoted by lines and markers, sensitivities by cumulative bars.Models used: CMAQ v. 5.2.1 and CMAQ v. 5.0.2.

Fig. 10 .
Fig. 10.Monthly variations in the correlation coefficients based on hourly metal concentrations (Ti, Mn, Fe, Al, Si, K, and Ca) in the fine fraction at all four PX-375 stations derived from the base and sensitivity CMAQ v. 5.2.1 simulations.

Fig. 11 .
Fig. 11.Observed and simulated hourly concentrations of (a) Fe and (b) Ca, based on CMAQ v. 5.2.1 and CMAQ v. 5.0.2 base and sensitivity simulations, for Fukuoka in May 2017.w/o: without.

Fig. S1 .
Fig. S1.Air quality parameters at the ten ACSA-14 stations, giving the observed and simulated monthly mean concentrations and source sensitivities.The values were averaged over all hours in each month.Concentrations are denoted by lines and markers; sensitivities by cumulative bars.Models used: CMAQ v. 5.2.1 and 5.0.2, and CAMx.

Fig. S3 .
Fig. S3.Air quality parameters at the four PX-375 stations, giving the observed and simulated monthly mean concentrations and source sensitivities.The values were averaged over all hours in each month.Concentrations are denoted by lines and markers; sensitivities by cumulative bars.Models used: CMAQ v. 5.2.1 and 5.0.2.

Table 1 .
Air quality model performances for the entire target period (April 1, 2017-March 31, 2018), based on annual mean observed (Obs.) and simulated (Sim.)concentrations in Japan.