The sampling periods were from 25 April to 1 June 2011 (spring), from 1 to 29 September 2011 (summer), from 28 October to 15 December 2011 (autumn), and from 19 January to 1 March 2012 (winter). HR-ToF-AMS measurements were conducted at the HKUST AQRS (22∘20′ N, 114∘16′ E), which sits on the hillside of Clear Water Bay on the east coast of Hong Kong, which has a relatively low population density. There is little local emission around the site. Two minor local sources include Clear Water Bay Road and a small student canteen. The Clear Water Bay Road is located outside of the campus and has mild traffic. The student canteen within 200 m of the site, which operates only from 18:00 LT (local time) until midnight. The HR-ToF-AMS operating procedure, data analysis and species determination have been discussed in Lee et al. (2013a), Li et al. (2015) and Huang et al. (2015). Briefly, the AMS was operated alternatively between the V + PToF combined mode and the W mode for 5 min each. A collection efficiency of 0.5 was employed for measurements at this site, where the particles have overwhelmingly dominant sulfate content (Aiken et al., 2009; Li et al., 2013). Concentrations of methanesulfonic acid (MSA) and organosulfates (OSs) were estimated by combining the V-mode data for total concentrations and the W-mode data for high-resolution mass spectral analysis for specific ions (Huang et al., 2015). Mass spectra of organic sulfur compounds were obtained from standards in laboratory experiments to support the W-mode data analysis (Huang et al., 2015). We further calculated the particle liquid water content (LWC) by applying E-AIM II (Clegg et al., 1998) to explore the effects of aqueous processing on PM1 composition. Gaseous species (CO, CO2, SO2, NO, NO2, and O3) were measured with standard gas analyzers (Teledyne API). Volatile organic compounds (VOCs) were measured by gas chromatography (Synspec GC955). Meteorological parameters were measured by an automatic weather station mounted on a tower right next to the supersite. Particle hygroscopicity and size distribution measurements have previously been taken with a HR-ToF-AMS at this site (Cheung et al., 2015; Man et al., 2015; Meng et al., 2014; Yeung et al., 2014), and direct reference to the resulting publications will be made where necessary.

Table 1 summarizes the meteorological conditions, PM1 concentrations, and the estimated LWC in the 10 high-PM episodes. All of the episodes involved air masses that originated over East Asia from the north, northeast or northwest of Hong Kong and swept over part of the PRD region before reaching the site. Seven of the 10 episodes (E1–E7) were characterized by medium-range trajectories and the other three (E8–E10) by long-range trajectories. The individual trajectories are shown in Fig. S4. E1 and E2 had much lower solar irradiance (IR) but higher LWC than the other episodes, which distinguished them from the other medium-range transport episodes. Thus, E1 and E2 were categorized as episodes of the LWC type (medium-range transport with high LWC and low IR) and the other medium-range transport episodes as episodes of the IR type (medium-range transport with high IR and low LWC). Li et al. (2013) referred to E1 and E2 as foggy episodes, while Li et al. (2013) and Lee et al. (2013a) referred to E3 as a hazy episode. The long-range transport episodes might be less associated with the local site-specific conditions and were categorized as episodes of the LRT type.

High concentrations of PM can have a number of causes, including enhanced primary emissions (Ji et al., 2014), concentrating effects due to a decrease in the height of the planetary boundary layer (Petäjä et al., 2016), regional transport (Huang et al., 2009), and active secondary formation (Hayes et al., 2013). Local primary emissions were not very significant at this site, as can be seen from the low contribution of POA (less than 6 %) throughout the whole campaign. As an indicator for primary PM, elemental carbon (EC) concentrations in PM2.5 filter sampling at this site from March 2011 to February 2012 were also found to be low throughout the year (0.86 ± 0.53 µg m-3) (X. H. H. Huang et al., 2014). Boundary layer dynamics on the high-PM days can be a factor affecting PM concentration, but the effects were likely minimal as the highest concentration was usually observed during the day at higher mixing heights (Fig. S5). Therefore, regional transport and active secondary formation would be the most probable causes for the episodic events of high PM concentrations at this suburban site. More detailed meteorological conditions with chemical characteristics in each episode can be found in Fig. S5.

As is apparent from Table 1, the occurrence of different types of episodes exhibits a seasonal trend. LWC episodes occurred only in spring and LRT episodes only in winter, while IR episodes took place in spring, summer and autumn. This result is consistent with previous results (Huang et al., 2009) in that the frequency of high-PM days in Hong Kong had a strong seasonal variation. In winter, the overwhelming northerly wind brings pollutants via long-range transport (Fang et al., 1999). In spring, foggy days with high PM levels are common due to the moisture-laden air masses coming in from the sea and aqueous-phase processing of particulate species (Li et al., 2013). In summer and autumn, however, hazy days are mainly due to high photochemical activities in this subtropical area, resulting in the formation of secondary aerosols (Hu et al., 2008; Zhou et al., 2014).

Figure 1 shows the chemical constituents of NR-PM1 in the three types of episodes. It is apparent that sulfate dominated in all types of episodes. In Hong Kong, sulfate is largely regarded as a major regional pollutant with little spatial variability, as in the rest of the PRD (Hagler et al., 2006; Louie et al., 2005). Nitrate contributed less than 4 % in LWC episodes and IR episodes but more than 7 % in LRT episodes. As LRT episodes occurred in wintertime, the higher nitrate concentration was likely driven by gas–particle partitioning of ammonium nitrate to the particle at low temperatures (Seinfeld and Pandis, 2006). Using the PMF-resolved SVOOA and LVOOA as proxies for less-oxidized and more-oxidized SOA respectively (Zhang et al., 2011), more details of OA can be revealed. SVOOA had higher contributions in IR episodes, while LVOOA contributed roughly twice as much as SVOOA did in LRT episodes, because the air mass was already quite aged when it reached the site. LVOOA and SVOOA made similar contributions in LWC episodes.

Figure 2 shows the diurnal variations of the NR-PM1 species, PMF-resolved organic factors, and Ox (O3+ NO2) in these three types of episodes. SVOOA and LVOOA as well as Ox increased during the day in IR episodes, with a time lag between SVOOA and LVOOA. A similar time lag was also observed between SVOOA and LVOOA in the Yangtze River Delta (YRD), another fast-developing region of China (Huang et al., 2012). These delays may be the result of conversions from SVOOA to LVOOA in the afternoon. We explore such a possibility in Sect. 3.6. SVOOA and LVOOA both exhibit flat diurnal patterns in LWC episodes and LRT episodes.

Elemental analysis of OA (ratios of H : C, O : C, N : C, S : C and OM : OC) from the high-resolution mass spectra provides useful information to assess OA evolution. Recently, Canagaratna et al. (2015) used an updated (“Improved-Ambient”) method to estimate O : C and H : C ratios, and reported 27 % higher O : C ratios and 11 % higher H : C ratios than those estimated using the original (“Aiken-Ambient”) method. Recalculating the elemental ratios for the September dataset using the updated method shows little difference from those obtained by simply applying the respective factors of 1.27 and 1.11 to the O : C and H : C ratios (Fig. S6). Hence, the O : C and H : C ratios in this study were corrected by factors of 1.27 and 1.11, respectively, with Aiken-Ambient values reported in our previous studies. In the Van Krevelen diagram (Heald et al., 2010; Ng et al., 2011) shown in Fig. 3a, data points for LWC episodes (blue) fall into a lower O : C region than do the data points for IR (red) and LRT episodes (green). Although aqueous-phase processing might generate highly oxygenated organic compounds (Li et al., 2016; Mazzoleni et al., 2010; Zhao et al., 2013), sampling by the AMS was only limited to interstitial particles and a portion of very small fog droplets after drying. This can also lead to O : C ratios lower than those in IR periods when most of the photochemically oxidized OA were effectively sampled. Even though data points for IR episodes and LRT episodes have similar slopes and intercepts in the Van Krevelen diagram, data points for IR episodes had a much wider spread. These trends are also reflected in the diurnal patterns of carbon oxidation state (OS‾c≈2×O:C-H:C) (Kroll et al., 2011) in Fig. 3b. The OS‾c diurnal pattern in LRT episodes was relatively flat, suggesting that oxidized organics were mostly transported to the site with minor in situ oxidation. The OS‾c in IR episodes gradually increased from 09:00 until 15:00 LT. Similar trends were observed for Ox, LVOOA and, to a lesser extent, SVOOA. With all these combined, we believe that the local photochemical processing of OA was more likely at play in IR episodes than the long-range transport of processed aerosols was.

Figure 4 shows the peak fitting results of the type-averaged size distributions of organics and sulfate mass. The mass-mode diameters (Dva) for both the small and large modes of organics and sulfate did not differ considerably across the episode types (differing by less than 5 %). Within each type of episode, sulfate had a smaller fraction of small particle mode than organics did, indicating that sulfate was relatively aged, while organics received contributions from local fresh emissions. LWC episodes received the largest contribution from small-mode sulfate because of some local influences, whereas LRT episodes received the smallest contribution with relatively few local activities. The oxidation mechanisms, however, might be different. Aqueous-phase oxidation may dominate in LWC episodes, while photochemical oxidation may dominate in IR episodes.

Various studies have analyzed the particle mixing state using single-particle instruments such as aerosol time-of-flight mass spectrometers (Healy et al., 2013, 2014; Yang et al., 2012) and single-particle aerosol mass spectrometers (Wang et al., 2015). Particle mixing state can also be inferred from particle size information obtained with the AMS. If the organics and sulfate are internally mixed (i.e., they exist in the same particle), their diameters should be strongly correlated with each other and their size should grow at a similar rate. The observed strong correlation and slope of unity (correlated in time and size) suggest that these species are likely internally mixed, although we cannot completely exclude the possibility of external mixing. On the other hand, if the mode diameters of sulfate and organics did not change coherently and exhibit a strong correlation with a slope close to unity, these particles were more likely externally mixed. Bahreini et al. (2003) have used such correlations in size to indicate the mixing states of species. In our study, the large-mode diameters of organics and sulfate were strongly correlated (Pearson's R value equals 0.7) with a slope close to unity in LWC episodes (Fig. 5), suggesting that organics and sulfate were likely internally mixed in the large particles. However, these conditions of correlation and slope are necessary but not sufficient evidence for internal mixing. This internal mixing may occur during the process of local aqueous oxidation. In IR episodes, during which local photochemical oxidation may have a more obvious influence, larger particles do not mix well internally (poor correlation between the large-mode diameters of organics and sulfate with Pearson's R (Rpr) = 0.2). As discussed in Sect. 3.2, in IR episodes, organics showed a clear noontime peak associated with local photochemical activities, while sulfate was still mainly a regional pollutant. As a result, large particles of organics and sulfates were very likely to have been externally mixed during IR episodes. A good correlation (Rpr = 0.7) with the slope deviating substantially from unity (slope = 0.5) was observed in LRT episodes. As long-range transport was the dominant process causing high PM levels during LRT episodes, organics and sulfate would have been brought to the site together, so their large-mode diameters tend to be strongly correlated. However, they may have different origins and may also have undergone different aging processes during the course of long-range transport, and thus their mode diameters would be different. The correlations between the small-mode diameters of organics and sulfate were notably weaker, with Rpr = 0.5 in LWC episodes, Rpr = 0.2 in IR episodes, and Rpr = 0.2 in LRT episodes, suggesting that freshly formed small particles mixed externally.

As discussed earlier, there may be some local atmospheric processing of aerosols in LWC and IR episodes but not in LRT episodes. Therefore, we further explored the mechanisms underlying the atmospheric processing of LWC and IR episodes based on the size variations before and during episodic events. Figure 6 shows the particle mass-mode diameters and areas (concentrations) in the LWC and IR types of episodes. We obtained the percentage changes in mode diameters by comparing the smallest diameter before the episode and the largest diameter during the episode for each episode. These percentages in each episode were then averaged to obtain the percentage changes for each episode type. The results show that the changes in mode diameter were small in the LWC episodes: -2.5 % for small-mode organics, +8.1 % for large-mode organics, +1.6 % for small-mode sulfate, and -3 % for large-mode sulfate. In contrast, the changes in mode diameters were much more drastic in the IR episodes: +51.3 % for small-mode organics, +40.5 % for large-mode organics, +45.4 % for small-mode sulfate, and +35.9 % for large-mode sulfate. Furthermore, particle size usually increased more rapidly before the IR episodes (shaded in blue in Fig. 6) than during the episodes (shaded in orange). With fewer pre-existing particles before the episodes, particle growth – likely via condensation and reactive uptake of semi-volatile components – was more rapid than during the episodes. The number concentration is discussed in detail in the Supplement.