Taiwan is an island located in the western Pacific, separated from mainland China on the west by the Taiwan Strait. To the north is the East China Sea and to the south sits the Philippines across the Bashi Strait. For a meteorological evaluation, we chose the following eight representative stations (all shown in Fig. 1) operated and maintained by the Taiwan Central Weather Bureau (CWB): Peng Jiayu (PJY), Taipei (TPE), Chupei (CP), Taichung (TC), Chiayi (CYm), Tainan (TNm), Kaohsiung (KH), and Hengchun (HCm) to evaluate the modeling performance of temperature, relative humidity, wind speed, and wind direction. The propeller wind direction anemometer (Komatsu Ltd.), quartz precision thermohygrograph (model no. 3-3122; Isuzu Seisakusho Co., Ltd.), and Pt electrical resistance thermometer (model no. 05103; R.M. Young Company) were used to monitor the wind speed and direction, relative humidity and air temperature, respectively. The measurement equipment was under routine calibration by the Taiwan CWB (https://www.cwb.gov.tw/Data/knowledge/announce/MIC.pdf, last access: 27 November 2020).

Since most residents live in the relatively flat western Taiwan region, the observations of air quality monitoring stations (see Fig. 1 for the locations) operated and maintained by the Taiwan Environmental Protection Administration (TEPA) at the Banqiao (BQ), Pingzhen (PZ), Miaoli (ML), Zhongming (ZM), Chiayi (CYa), Tainan (TNa), Zuoying (ZY), and Hengchun (HCa) stations were chosen for PM2.5 evaluation. The BAM-1020 particulate monitor (Met One Instruments Inc.) was used to monitor PM2.5. The automatic meteorological and air quality data are provided in hourly recordings to the public.

In this study, we also compared the modeling results with the PM2.5 composition analyzed by Lee et al. (2017) at BQ, ZM, and CYa for 13 January and 18 July 2017. They used the SASS PM2.5 samplers (Met One Instruments Inc.) to collect the 24 h (00:00 to 00:00 local time – LT) PM2.5 composition samples every 6 d. The quality assurance of the PM2.5 monitoring and analysis is referred to in Lee et al. (2017; see chapter 4).

This study applied the WRF v3.9.1 (Skamarock and Klemp, 2008) and CMAQ v5.2.1 (Byun and Schere, 2006) for scenario simulations. The WRF and CMAQ modeling used two-way and one-way nesting methods, respectively, in this study. The initial meteorological condition was from ds083.3 NCEP GDAS and FNL 0.25 Degree Global Tropospheric Analyses and Forecast Grids (https://rda.ucar.edu/datasets/ds083.3/, last access: 27 November 2020). The horizontal resolutions of the four domains, from outer to inner, were 81, 27, 9, and 3 km, respectively. The first domain covered East Asia and Southeast Asia, and the fourth domain contained only the Taiwan island. The vertical layers were 46, approximately 20 layers below 1.7 km, in which the boundary layer was well resolved. The model's top is set to 50 hPa. In order to obtain a better meteorological field, the WRF modeling applied 4D data assimilation with grid nudging for domains 1, 2, and 3, and with observation nudging for domain 4. The anthropogenic emissions for East Asia and the Taiwan island were obtained from MIX (see M. Li et al., 2017) and the Taiwan Emission Data System (TEDS) 10.0 (TEPA, 2017), which are based on the years 2010 and 2016, respectively. The MIX emissions of SO2, NOx, NMHC, NH3, CO, PM10, and PM2.5 covering Chinese mainland were adjusted with changes of -62 %, -17 %, 11 %, 1 %, -27 %, -38 %, and -35 %, respectively, according to the change in annual emissions between 2010 and 2017 (Zheng et al., 2018). This study assumes the emissions of 2017 in Taiwan are the same as those of 2016. The biogenic emissions were prepared by the Biogenic Emission Inventory System version 3.09 (BEIS3; Vukovich and Pierce, 2002) for the Taiwan island and Model of Emissions of Gases and Aerosols from Nature v2.1 (MEGAN; Guenther et al., 2012) for regions outside Taiwan. The biomass burning emissions imported the data of the Fire INventory from National Center for Atmospheric Research (NCAR; collectively FINN) v1.5 inventory (Wiedinmyer et al., 2011). The model configurations of physics and chemistry for this study are listed in Sect. S1 of the Supplement, and the emission maps of, for example, NO for four domains are referred to in Sect. S2. In addition, in order to clearly reveal the origin of the EAH arriving at Taiwan, the NOAA's Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT; Stein et al., 2015; https://www.ready.noaa.gov/HYSPLIT.php, last access: 27 November 2020) was applied in which the ensemble trajectory method and the reanalysis database were chosen.

This study used statistical indexes such as mean bias (MB), mean average gross error (MAGE), and index of agreement (IOA) to evaluate temperature, wind speed, and relative humidity and used wind normalized mean bias (WNMB) and wind normalized mean error (WNME) for wind direction in the fourth domain. For PM2.5 performance in the same domain, we applied the MB, mean fractional bias (MFB), and mean fractional error (MFE), R (correlation coefficient), and IOA indexes. All of the formulas for the above indexes are from Emery (2001) and TEPA (2016), as illustrated in Sect. S3.

As mentioned, the impact was considered as being the reduction in a specific source or roughly the contribution of that specific source for the BFM method, i.e., the difference between the base and zero-out scenarios, is applied in this study. For the impact of the three industrial regions on PM2.5 in Taiwan in January 2017, the monthly mean impact from BRIR (the difference between base and BRIR scenarios) was approximately 0.7–1.1 µg m-3, as illustrated in Fig. 2a. The relative impact was higher in northern Taiwan, with approximately 5 % of total PM2.5. The proportion of influence gradually decreased from north to south (Fig. 2b).

Comparing Fig. 2a and b with Fig. 2c and d, it is apparent that the monthly mean influence from YRDIR was higher than from BRIR. The reason is that YRDIR is closer to Taiwan than BRIR. The monthly mean impact from YRDIR was approximately 1.2–1.9 µg m-3, the highest in northern Taiwan, with a proportion of approximately 7.5 % of the total monthly average PM2.5 concentration. The spatial influence from BRIR was similar to YRDIR since these two industrial regions are both located off the north of Taiwan, i.e., upstream of Taiwan under the prevailing northeasterly wind. For the daily mean influence, the impact of YRDIR was also higher than BRIR, and the influencing period was almost the same for both regions because the EAH originated from YRDIR and BRIR arrived in Taiwan one after another under the prevailing northeasterly wind (Fig. 3a1–a3, b1–b3). In particular, the contributions from BRIR and YRDIR to northern Taiwan could reach daily averages of 8 and 11 µg m-3 on 9 January 2017. In January 2017, the proportion of influence was higher than on 8–9, 14–15, and 20–23 January. The influence of the EAH was closely related to the intrusion of Asian anticyclones (Chuang et al., 2008a, b).

The spatial distribution of the influence from PRDIR was totally different from BRIR and YRDIR, as shown in Fig. 2e and f. Interestingly, the impact from PRDIR was higher on the mountains than on the flat plain. For the stations on flat western Taiwan, there was slight influence on 8 to 12 January 2017 (Fig. 3c1–c3). It was found that there was a stationary front from the sea north of Taiwan that extended southwest to the Fujian and Guangdong provinces (F and G in Fig. 1) on 7 January (Fig. S4.6a). The front passed Taiwan on 8 January (Fig. S4.6b). Figure 3c1–c3 show that the influence on southern Taiwan was higher than that on northern Taiwan. Similar fronts passed Taiwan on 10 January (Fig. S4.6c) and 12 January (Fig. S4.6d). From Fig. 4, it can be seen that the air mass from PRDIR would transport pollutants upward, above the top of the boundary, and then move them eastwards (Fig. 4a1, b1). When the pollutants ran into the mountains in Taiwan, they were blocked and transported to the ground through vertical mixing (Fig. 4a2–a3, b2–b3). This transport mechanism is quite similar to the biomass burning aerosols from Southeast Asia to Taiwan (Yen et al., 2013; Chuang et al., 2016). The boundary layer mixing was enhanced by the passing of a cold surge and increased PM2.5 on the ground.

This study applied the process analysis technique in the CMAQ model, in which the terms of horizontal advection (HADV), vertical advection (ZADV), horizontal diffusion (HDIF), vertical diffusion (VDIF), emissions (EMIS), dry deposition (DDEP), cloud process and aqueous chemistry (CLDS), gas chemistry (CHEM), and aerosol chemistry (AERO) in the diffusion equation can be resolved (Byun and Schere, 2006). Each term contributes to the rate of change in the PM2.5 level at the following locations (see Fig. 1) chosen in this study: position 1, located between the Bohai Sea and East China Sea; position 2, located between the East China Sea and Taiwan; position 3, located in the middle of the Taiwan Strait; position 4, located in the northern South China Sea; BQ in northern Taiwan; ZM in central Taiwan; and CY in southwestern Taiwan. Although CY and ZM are closer to each other than BQ, CY was selected due to PM2.5 being sampled at this station, and it is representative among many stations in southern Taiwan. Those positions were chosen because they are on the path of the northeasterly wind. Through the value of each term in the process analysis, we can understand which term can produce or remove PM2.5 at these positions and, therefore, realize the physical and chemical processes during LRT. It should be noted that each term resolved by the process analysis is based on modeling results, and no evaluation of such processes was available.

Similar to Fig. 2, we deduced the differences in base and zero-out scenarios for the IPR analysis. This study considered the reduction as the approximate contribution by each industrial region. Therefore, the following discussion is satisfied when the chemical reaction between each industrial region and the surrounding area was ignored. The physical or chemical terms in Fig. 5a1 and a2 did not always appear synchronously, and their proportions in total were not equal. This implies that position 1 was influenced by both BRIR and other nearby sources. The increase in PM2.5 was caused mainly by the process of HADV, followed by ZADV and VDIF, and the removal process was mainly AERO. The removal process is likely caused by the evaporation of ammonium nitrate in the PM2.5 plume moving from high-latitude regions to low-latitude regions through increasing ambient temperature (Stelson and Seinfeld, 1982; Chuang et al., 2008b). In contrast, there was occasionally less PM2.5 from YRDIR (Fig. 5a3) and nearly none from PRDIR (Fig. 5a4). This is expected because the northeasterly wind prevails in winter, and the BRIR and YRDIR and PRDIR are located upstream and downstream of position 1, respectively. From Fig. 5b1–b4, among the three industrial regions, it is apparent that position 2 was influenced by both the BRIR and YRDIR as it was mainly produced through inconsistent HADV, VDIF, ZADV, and CLDS, removed through AERO and occasional HADV and DDEP processes, and almost unaffected by PRDIR. For position 3, PM2.5 was influenced mainly by YRDIR (Fig. 5c3) and occasionally by BRIR (Fig. 5c2) and was also influenced by PRDIR from 8 to 12 January (Fig. 5c4), with a positive contribution of CLDS, which could be attributed to the high relative humidity environment over the Taiwan Strait. The production from BRIR and YRDIR were mainly attributed to CLDS, and the removal process was mainly AERO and, second, DDEP. The positive and negative contributions of PM2.5 for position 4 were very similar to position 3 but slightly lower (Fig. 5d1–d4) because it is further from BRIR and YRDIR than position 3. Although position 4 is very near PRDIR, it was influenced more by YRDIR (Fig. 5d3) and other sources in eastern and northern China rather than three industrial regions since the prevailing wind was mainly the northeasterly wind in January. From the above, it was found that the PM2.5 plume was transported southwards from BRIR or YRDIR on a 3D path, i.e., horizontal and vertical advection, and via vertical diffusion over the Bohai Sea and the East China Sea. During the southward transport, AERO was always the major removal process, i.e., evaporation of volatile species. When the plume was transported to subtropical regions, cloud processes became the major production process of PM2.5. The reason for this was the condensation in the mix of a cold PM2.5 plume from high-latitude regions to warm air and sea at low-latitude regions.

The build-up of PM2.5 at BQ was mainly HADV with minor CLDS, and the removal processes were mainly ZADV with minor AERO (Fig. 5e1). This suggests that the PM2.5 plume was mainly transported horizontally when it was close to or when it reached northern Taiwan. Moreover, each industrial region contributed PM2.5 to BQ in very similar processes (Fig. 5e2–e4). In addition, certain PM2.5 was formed in northern Taiwan, probably due to the high relative humidity, which was probably induced by the clouds or fog produced by terrain uplifting. The removal process of PM2.5 at BQ was mainly ZADV, which can be explained by BQ being located in the Taipei basin, and the PM2.5 is transported up to leave the basin. Comparing Fig. 5f1 with Fig. 5f2–f3, it is obvious that the PM2.5 of ZM was produced by local pollution, i.e., the downward diffusion of VDIF, which probably came from northern Taiwan and was removed further through HADV to southern Taiwan under the prevailing northerly wind. In other words, the PM2.5 in upstream northern Taiwan was vertically advected and diffused southwards to central Taiwan and then horizontally advected to downwind areas. On the other hand, the influence from PRDIR was much less when the prevailing wind was the northeasterly monsoon (Fig. 5f4). However, when the cold surge passed Taiwan (8 and 10 January), the influence from PRDIR could not be ignored, which is illustrated in Figs. 2f, 4, and 5f4. On 8 to 10 January, the negative ZADV indicated the concentration was decreasing in the lower 20 averaged layers, where the daily process occur, but the concentration gradient was positive (∂PM2.5∂z>0; the concentration of PM2.5 from PRDIR was higher at a high altitude than that at a low altitude over Taiwan), which implies the vertical velocity had to be negative, i.e., a downward motion. Therefore, the boundary layer mixing of the aloft PM2.5 plume was enhanced by the passing of the cold surge (Yen et al., 2013; Chuang et al., 2016). For CY, located in southwestern Taiwan, VDIF and HADV mainly contributed to the gains of PM2.5, and the removal processes were mainly ZADV and AERO; however, occasionally, when the positive contributions to PM2.5 were ZADV and VDIF, the removal processes were HADV and AERO (Fig. 5f1). Comparing Fig. 5f2–f4 with Fig. 5g2–g4, it is obvious that the positive and negative contributions to PM2.5 for CY were very similar to those for ZM. The impact from BRIR and YRDIR was less than from local sources. When the cold surge passed Taiwan, PRDIR influenced PM2.5 at CY as well.

On most days in the winter period, northeasterly winds prevailed over East Asia. In this section, we chose 13 January 2017 to discuss the physical and chemical processes in detail because it is a classical moderate EAH episode in which PM2.5 sampling was implemented, and it will be discussed in Sect. 3.6. On this day, the Asian anticyclone transported pollutants from the Asian continent to Taiwan and caused high PM2.5 episodes. The 72 h backward trajectory ensemble, starting from BQ, ZM, and CY, obviously traced back to East Asia, where BRIR and YRDIR are located (Fig. S4.7a1–a3).

Although the impact of LRT on 13 January was less than on 8, 9, 20, or 22 January (Fig. 3), the physical and chemical processes during transport were similar for these days since the weather patterns were quite analogous to each other. Such LRT events occurred in a weather pattern, as illustrated in Fig. 6a. The Asian anticyclone was moving from East Asia to the western Pacific. The peripheral circulation of the Asian anticyclone was the strong northeasterly wind on the coastal areas and the sea. It was found the northeasterly wind formed a lee calm wind region in southern Taiwan, where PM2.5 accumulated (Fig. 7a–b). When the leading edge of the Asian anticyclone arrived, the wind speed increased and therefore enhanced the dispersion of PM2.5 in southern Taiwan (Fig. 7c–e). Subsequently, the LRT haze arrived (Fig. 7f) and split to the eastern and western side of Taiwan due to the blocking of mountains, with more hazing moving west (Fig. 7g–i).

Figure 8a1–a4 shows that the influence of BRIR was more than that of YRDIR and PRDIR at position 1 on 13 January, since BRIR is located upstream of position 1 under the northeasterly wind. The major production process was VDIF below 760 m (layer 14) and AERO with less CLDS above 760 m. This implies that the transport path from BRIR to position 1 could be horizontal between BRIR and position 1 and then vertical at the location of position 1. The removal process was AERO below 760 m and VDIF above. This suggests that the ascent and subsidence of air parcels might enhance the formation and removal of aerosols below and above 760 m, respectively. It is possible that the ascent motion of the air parcel near the warm surface moved to a cold environment at a higher altitude, up to 760 m. This may cause condensation and trigger heterogeneous reactions of aerosols. In contrast, the descent motion of the air parcel above 760 m may cause the evaporation of aerosols due to a warmer environment at lower altitudes than aloft. Although position 1 was slightly influenced by YRDIR, the contribution of different processes from YRDIR on position 1 was less and inconsistent (Fig. 8a3). The contribution of different processes from PRDIR to position 1 was also inconsistent and even less (Fig. 8a4). From Fig. 8b1–b4, it was found that position 2 was mainly influenced by YRDIR on 13 January. The major processes below layer 9 (∼ 310 m) that contributed to the increase in PM2.5 were HADV, VDIF, and ZADV, and the removal processes were DDEP and AERO (Fig. 8b3). Position 2 was slightly influenced by BRIR, with the major production processes being VDIF and ZADV, and the removal process was AERO (Fig. 8b2). On 13 January, position 3 and position 4 were less influenced by all industrial regions (Fig. 8c2–c4, d2–d4). This implied that position 3 was possibly influenced by the nearby Fujian province on the northern and western side of the Taiwan Strait. On 13 January, position 4 was also less influenced by the three industrial regions, probably due to BRIR and YRDIR being distant and PRDIR being located downstream of position 4. Comparing Fig. 8e1 with Fig. 8e2–e4, it was found the BQ was influenced more by YRDIR. The major contribution processes at BQ below 200 m (layer 7) were HADV, followed by AERO, and above 200 m they were either VDIF, ZADV, or CLDS or mixture of them. The plume moved horizontally, close to BQ, and formed a certain amount of PM2.5 when reaching BQ. The major removal process was ZADV followed by VDIF below 200 m but HADV and AERO above. BQ was less influenced by BRIR due to the long distance, deviation in the wind direction and PRDIR, since BQ is located upstream of PRDIR. In this event, ZM and CY were less influenced not only by BRIR and PRDIR but also by YRDIR (Fig. 8f1–g4). This explains the haze plume that passed BQ and was then transported for a limited distance in front of southern ZM and CY on 13 January.

The severe EAH episodes always accompany the arrival of strong anticyclones (Fig. 6b). This study chose 9 January to discuss because of its largest impact on January 2017. The 72 h backward trajectory ensemble starting from BQ, ZM, and CY on 9 January is similar to that on 13 January (Fig. S4.7b1–b3). The PM2.5 event occurring in western Taiwan on 9 January was similar to that on 13 January, and both were LRT of EAH. However, there were still differences between these two events. First, the impact of the three industrial regions on PM2.5 in western Taiwan was much higher on 9 January than 13 January. Second, for the haze from BRIR and YRDIR, the positive and negative contribution processes on BQ were mainly HADV and AERO and ZADV and VDIF below 200 m (layer 7; Fig. 8e3), and there were fewer different processes at different layers above 200 m on 13 January. On 9 January, the major processes leading to the increase in PM2.5 at BQ were mainly HADV below 380 m (layer 10), AERO between 120 and 900 m (layer 5 to 15), and ZADV and CLDS between 650 and 1500 m (layer 13 to 19), as illustrated in Fig. 9e2–e3. The removal process was mainly ZADV below 460 m (layer 11), HADV between 550 and 900 m (layer 12 to 15), and HADV and AERO between 1000 and 1300 m (layer 16 to 18). Third, the stronger event occurring on 9 January had a more obvious impact on ZM and CY than that of 13 January. The higher production of HADV without AERO near the surface on 9 January explains the massive accumulation of EAH over the Asian continent and the rapid movement of the anticyclone. The strong and fast plume passing BQ led to insufficient time for the formation of PM2.5 at BQ, but it could transport the EAH further to southern ZM and CY. In contrast, the higher production of AERO near the surface occurring on 13 January explains how the slow-moving EAH had time to react with the local pollutants, e.g., HNO3 in the Asian plume reacted with local NH3 to form NH4NO3, which has been discussed in Chen et al. (2014).

Figure 10a and b reveal that the impact of BRIR on PM2.5 in Taiwan was negligible in July compared to January. The monthly contribution was less than 0.01 µg m-3 or less than 0.04 % of total PM2.5 in western Taiwan. The influence from YRDIR and PRDIR on Taiwan was equally small with BRIR (Fig. 10c–f). As illustrated in Fig. S4.8, the daily contribution from the three industrial regions to western Taiwan was similar for all cities. The contribution from BRIR was only less than 0.1 µg m-3 from 25 to 28 July (Fig. S4.8a1–a7), and from YRDIR it was approximately 0.1–0.3 µg m-3 from 27 to 29 July (Fig. S4.8b1–b7) and detectable on 28 July but increased to 0.2–0.5 µg m-3 on 30 to 31 July (Fig. S4.8c1–c7). Owing to the small impact from the three industrial regions on western Taiwan, the physical and chemical processes were small for all days in July 2017 except for the last few days in that month, as illustrated in Fig. S4.9. The weather map revealed that there was a thermal low near Taiwan at the end of July (Fig. S4.10). In short, during the period of a prevailing southwesterly to southeasterly wind, the influence of BRIR, YRDIR, or PRDIR could be ignored unless there was a special weather system, such as the aforementioned thermal low, which could transport less PM2.5 from distant sources. We can consider that the Asian continent has almost no impact on Taiwan in July. In other words, the origin of PM2.5 in Taiwan in July is local pollution and the background atmosphere.

Take, as an example, 18 July 2017 on which the PM2.5 sampling was implemented. It was found that position 1 was influenced more by YRDIR than BRIR among three industrial regions (Fig. S4.11a1–a4). The positive and negative contribution processes were inconsistent below 80 m (layer 4). However, from 120 to 460 m (layer 5 to 11), the major processes in the build-up of PM2.5 were AERO and ZADV, and the removal process was mainly HADV. Figure S4.11 shows that the influence of the three industrial regions on position 2, position 3, position 4, BQ, ZM, or CY was almost negligible. Furthermore, the 72 h backward trajectory ensemble starting from BQ, ZM, and CY on this day was traced back to the clean southwestern Pacific, which implied that the airflow was controlled by the Pacific high (Fig. S4.7c1–c3). This suggested that the PM2.5 was mainly from local pollution and background atmosphere on 18 July. On the other hand, on 30 July, the 72 h backward trajectory ensemble starting from the end at BQ, ZM, and CY went through a cyclone near Taiwan and then on to the South China Sea and Philippines (Fig. S4.7d1–d3). As mentioned earlier, the thermal low over the Taiwan Strait (Fig. S4.10) caused an unstable wind field and transported pollutants from the southeastern coastal areas of the Asian continent to the northern South China Sea, the Taiwan strait, and Taiwan (Fig. S4.12). In July 2017, there was hardly any PM2.5 transported from the three industrial regions to those specific locations, except from PRDIR to position 4, as illustrated in Fig. S4.13d4.

Lee et al. (2017) conducted PM2.5 sampling at BQ, ZM, and CY every 6 d in 2017. Only the sampling days are suitable for analysis in this study. The sampling from 13 January was compared with the simulated PM2.5 compositions, as indicated in Fig. 11. Previous studies (Wang et al., 2016) suggested that it took approximately 28 h for the PM2.5 haze to be transported from the Yangtze River estuary to the northern tip of Taiwan island. Therefore, the simulated PM2.5 compositions at position 1 and position 2 on 12 January were also illustrated. As illustrated in Fig. 11, on both 12 and 13 January the major simulated compositions were sulfate and organic carbon (OC) for position 1–4. However, the proportion of nitrate in PM2.5 at position 1 on 12 January was slightly higher than that at position 2 but much higher than that at position 3 and position 4. This can be explained by the nitrate evaporating from the aerosol phase to the gas phase for the PM2.5 plume transported from high- to low-latitude regions (Chuang et al., 2008b). The simulated proportions of Na+ and Cl- in PM2.5 at position 3 and position 4 were higher than those at position 1 and position 2. The higher sea salt due to a stronger wind speed is expected because the Taiwan Strait was a wind tunnel between the Central Mountain Range in Taiwan and the Wuyi Mountains in the Fujian province (Lin et al., 2012). In addition, the simulated proportions of nitrate in PM2.5 at BQ, ZM, and CY were higher than those over position 1–4, which should be caused by the local pollution. The comparison between the simulation and observation indicated that the performance of the simulation was not bad. The simulated proportions of nitrate and ammonium in PM2.5 were slightly lower than the observations. While the simulated proportions of K+, Ca2+, Mg2+, and Na+ were slightly overestimated. This suggested that the emissions of biomass burning and windblown dust over Taiwan island and the influence of sea salt still have room for improvement.

We also compared the simulated PM2.5 compositions with observations on 18 July 2017 (Fig. S4.14). As mentioned earlier, position 1 was influenced by upstream YRDIR, and the simulated proportion of nitrate in PM2.5 at position 1 was higher than further upstream position 2, position 3, and position 4. The simulated proportion of nitrate in PM2.5 at position 3 and position 4 was higher than position 2, which implies that position 3 and position 4 were influenced more by PRDIR than position 2. For BQ, ZM, and CY, the proportion of simulated OC in PM2.5 was slightly overestimated compared with observations, but nitrate, sulfate, and others were underestimated. Since BQ, ZM, and CY were less influenced by PRDIR on 18 July, the overestimation of OC and the underestimation of nitrate should be related to the bias of the local emissions inventory. In addition to the local emissions inventory, the underestimation of sulfate could possibly be related to the underestimation of emissions from uncertain sources, e.g., ships around Taiwan or local point sources, since the local line and area sources of SO2 are both low. Moreover, the uncertainty of emissions in the Southeast Asia Peninsula and Philippines is also another issue that needs to be improved.