The time series of meteorological conditions, particle number size distribution, PM2.5, OC and EC concentrations during the campaign are presented in Fig. 3. Overall, the campaign was under the influence of the prevailing northerly wind with an average wind speed and temperature (± 1 standard deviation) of 1.73 ± 0.95 m s-1 and 14.8 ± 5.1 ∘C, respectively. The average PM2.5 concentration was 48 ± 26 µg m-3. A few colder periods were observed, during which the wind speed increased and the temperature decreased. In general, the low wind speed favored the accumulation of PM2.5. During the campaign OC concentrations ranged from 0.5 to 47.0 µg m-3 with an average of 9.0 ± 6.0 µg m-3, while EC concentrations ranged from 0.2 to 23.0 µg m-3 with an average of 3.4 ± 3.0 µg m-3. OC1, the most volatile group among OC1 to OC4 in OC / EC analysis, accounted for one-third of the total carbon mass (Fig. 4).

On 17 February and 12 and 17 March 2014, the daily averaged PM2.5 concentrations exceeded 95 µg m-3, and they were nearly twice the daily averaged values of the other days (Fig. 3, shaded area in grey). Results of 72 h back trajectories (Stein et al., 2015; Rolph, 2016) showed that air masses arriving at the site on or before these 3 days mostly originated from the continental or oceanic area close to Eastern China (Fig. S3). The SMPS data also showed a mode near 100 nm with a high particle number concentration (Fig. 3).

The temporal variation of the number concentration of MV particles having an initial diameter of 80 nm or above tracked reasonably well with the accumulation of PM2.5 as the particles aged and became more internally mixed (Figs. 3 and S4). Furthermore, the number concentration of MV particles showed a size dependence in the 80–300 nm particles. There were days, e.g., from 24 February to 10 March 2014, when the number concentration of 300 nm MV particles did not track well with PM2.5 mass concentration. The mode of the total particle number size distribution was below 100 nm and the number concentrations of 300 nm particles were low (Fig. 3). PM2.5 mass concentration tracked better the number concentrations of 80 to 150 nm MV particles (Figs. S4a to S4c) than those of 200 and 300 nm MV particles (Figs. S4d and S4e).

The average number and volume fractions of CV, HV, MV and LV in VTDMA measurements at 300 ∘C are summarized in Table 2. VM is internally mixed with MV and HV residuals, and hence does not have a separate contribution to number concentrations. Overall, HV and MV particles, an indicator for aged aerosols with internally mixed non-volatile and volatile material, accounted for 57 to 71 % of the total particle number concentration. Non-volatile material (LV, MV and HV residuals) accounted for 15 to 26 % of the total volume of selected particles before heating. While the CV and HV fractions were larger in the finest particles selected (D0= 40 nm), MV and LV were more abundant in larger particles (D0 > 80 nm). As in Rose et al. (2006), fresh emissions like soot adsorb or absorb volatile material during atmospheric processing. The smaller particles grew faster than the larger ones because of their higher ratio of surface area to volume. When they were heated in the VTDMA at 300 ∘C, the smaller particles decreased more substantially in size, as reflected in the higher CV and HV fractions and lower MV and LV fractions. The higher abundance of MV and LV in the larger size particles could also be explained by the aged particles arriving at the sampling site. Since the sampling site is located on top of a mountain with an altitude of 150 m, the particles were likely aged upon arrival. Non-volatile particles in the ultrafine modes from fresh emissions can be aged with both non-volatile and volatile material, and became larger in size. Nevertheless, the detection limit of the downstream DMA and CPC in the VTDMA system is 10 nm. The particles with diameters below the detection limit leads to an overestimation of CV and an underestimation of the non-volatile residuals for the finest particles selected (with an initial diameter of 40 nm).

Figure 5 shows diurnal variation in the fraction of CV, HV residual, MV residual, LV residual and VM in the total volume of particles of dry initial diameters of 40, 150 and 300 nm. For 40 nm particles, a clear maximum and minimum of the fraction of CV, VM and HV residuals are observed at 08:00 and 13:00, respectively. The diurnal variation of the HV and MV particles in the 40 nm particles was clearer in terms of number fraction (Fig. S5). Furthermore, the trend of CV was opposite to those of VM, HV and MV. The increase in CV in the 40 nm particles and to a lesser extent in LV in the 150 and 300 nm particles in the morning is consistent with traffic patterns. Fresh emissions of volatile and non-volatile material, likely OC and EC, were externally mixed and contributed to CV and LV, respectively. As time progresses in a day, the highly volatile species (CV) that were freshly emitted in the morning may evaporate and react to form less volatile particles and become VM instead of CV (Robinson et al., 2007). Alternatively, these CV particles could also coagulate with smaller particles to form VM-containing particles. Less fresh emissions with more CV particles turning into VM on MV and HV particles could explain the trend that the number and volume fractions of CV decreased while those of MV and HV increased (Figs. 5 and S5).

We also used the diurnal variations in the volume fraction remaining (VFR), again defined as the volume ratio of the residual to its host particle (not to the total volume of all particles), to examine the size changes of the non-volatile residuals of HV and MV particles. The VFR of HV did not exhibit any obvious diurnal variations, but the VFR of MV peaked near 18:00. The VFR of the 40 nm MV particles increased after 14:00, while those of the 150 and 300 nm MV particles increased after 15:00. Since the VFR of HV and MV were relatively constant during the day, the increase in the VM fraction after the morning rush hours could be attributed to the increase in the number concentrations of the HV and MV particles instead of changes in the amount of VM on the MV or HV residuals.

The diurnal variations for particles larger than 80 nm were much less obvious than those for 40 nm particles in this study and in others (Rose et al., 2011; Cheng et al., 2012; Zhang et al., 2016). In winter, the atmosphere is more stable, resulting in a poorer dilution of aged particles with the less polluted aerosols from higher up (Rose et al., 2006). When the aged pollutants were trapped near the ground, the effect of aging of fresh emissions weakened. Therefore, although a daily maximum and a daily minimum were still observed for particles larger than 80 nm, the variation was mostly within 15 %.

The diurnal variations in the mass fractions of OC and EC in PM2.5 provided further insights to the observations above (Fig. 6). The OC and EC data on 12 and 17 March were excluded since they were more than 2 standard deviations higher than those on other days. Subtle morning peaks between 06:00 and 10:00 were observed for the volume fraction of LV residuals (Fig. 5). A similar peak was observed for the mass fraction of EC in PM2.5 in the morning (Fig. 6). This suggests that the LV particles may be related to the EC from vehicle emissions in the morning. This EC was relatively less aged and externally mixed with the other volatile material. In the late afternoon, the LV residuals showed another peak between 17:00 and 19:00, whereas the mass fraction of EC in PM2.5 exhibited a minimum at 15:00, after which it increased continuously. The continuous increase in EC at night is likely related to the increase in heavy-duty diesel traffic (Zhang et al., 2015), which was restricted during daytime (Bradsher, 2007).

Although OC1 contributed to about half of the total OC mass, the diurnal variation in the mass fraction of OC in PM2.5 was driven by the total mass of OC2, OC3 and OC4 (OC2-4), which reached a minimum between 05:00 and 09:00 and increased until 19:00. OC can be attributed to both primary and secondary sources. The increased mass fraction of OC in PM2.5 and the OC-to-EC ratio in the afternoon suggest that the sources of OC were less related to traffic but more to the aging and formation of secondary organic aerosols (Turpin et al., 1990; Chow et al., 1996). These OC2, OC3 and OC4 may be highly oxygenated species or oligomers that are less volatile than primary or less oxygenated organics (Kalberer et al., 2004; Huffman et al., 2009).

It is interesting to note that the volume fraction of the LV residuals and the VFR of MV particles at different sizes showed a dip in the afternoon (Fig. 5, third column from the left). The VFR of 40 nm MV particles showed a dip at 14:00, while those in 150 and 300 nm particles showed a dip at 15:00. The volume fraction of LV residuals in 150 and 300 nm particles reached a minimum at 13:00 and 15:00, respectively. Because EC decreased between 12:00 and 15:00, the increase in the volume fraction of LV residuals in 150 nm particles since 13:00 and the VFR of 40 nm MV particles since 14:00 may be related to the increased presence of aged organics as well as the EC particles that aged via coagulation and condensation.

We calculated 72 h back trajectories of air masses arriving at the sampling site (23∘00 N, 113∘25′′ E) at 4 h intervals (at 00:00, 04:00, 08:00, 12:00, 16:00 and 20:00 local time, UTC +8) using the PC version of the HYSPLIT4 (Hybrid Single Particle Lagrangian Integrated Trajectory, version 4) model (Stein et al., 2015; Rolph, 2016). Archived meteorological data from the Global Data Assimilation System (GDAS) 1∘ were employed and the receptor height was set at 500 m above ground level (a.g.l.). The 191 back trajectories calculated were grouped into six clusters based on their spatial distribution (Fig. 7).

Overall, the sampling site was mostly affected by northwesterly and northeasterly air masses. Clusters 1 and 3 are coastal and continental air masses, respectively, although both originated from the northeast. Clusters 4, 5 and 6 represent continental air masses originating from the northwest. Cluster 2 is a group of maritime air masses originating from the East China Sea northeast or east of Guangzhou. While the air masses in cluster 6 were transported at relatively high speeds and altitudes (over 3000 m a.g.l.), the air masses in all the other clusters were transported at an altitude below 1500 m a.g.l. for over 40 h before arriving at the site. As the air masses in cluster 6 only occurred for less than 3 days and since the corresponding VTDMA and OC / EC data were sometimes unavailable, cluster 6 will be excluded from the following discussion.

The average PM2.5, OC and EC concentrations associated with the air masses from the northeast of Guangzhou (clusters 1, 2 and 3) were higher than those from the northwest (clusters 4 and 5, Table 3). Days associated with the coastal and maritime air masses were more polluted than days associated with continental air masses for several reasons. First, South China as a region is often affected by the high-pressure system moving eastward or southward from the continent out to sea in winter. When the maritime or coastal air streams entered from the southeast of the sampling site at Panyu, the atmosphere at the sampling site became more stable with low local wind speeds (e.g., the polluted days on 17 February and 12, 16 and 17 March, Figs. 3 and S3). The local pollutants accumulated and the city was also affected by pollutants from the southeastern areas of the site (e.g., Shenzhen, Nansha and Dongguan). Second, land–sea breeze circulation was observed when the sampling site was under the influence of maritime air masses from 18 to 20 March. During the day, southeasterly wind prevailed and the wind speed was higher. In the evening, the southeasterly wind was gradually replaced by a southwesterly or northwesterly wind and the wind speed decreased (Fig. 3). The cycle started again in the morning when the westerly wind was gradually replaced by southeasterly wind. Such land–sea breeze effects can result in an effective redistribution and accumulation of air pollutants within the PRD region (Lo et al., 2006).

Furthermore, PM2.5 in the northeastern parts of China can exceed 200 µg m-3 due to both enhanced emissions from coal combustion for heating and poor dispersion during wintertime (Gu et al., 2014). Under the influence of the prevailing northerly or northeasterly wind in China, these pollutants were often transported to southern China and the East China Sea (Chen et al., 2012). The pollutants might also have accumulated when the maritime air masses spent about 2 days across Taiwan and the coast of southern China. In contrast, continental air masses in cluster 5 moved slightly faster, and were often associated with the cold front period during which the local wind speed and pressure increased but the temperature decreased (Fig. 3). As the cold air masses passed through the city, dispersion and clearance of pollutants were promoted, resulting in lower PM2.5 concentrations (Tan et al., 2013a). Therefore, unlike in other coastal cities like Hong Kong (Lee et al., 2013), in Panyu the maritime air masses could lead to more severe pollution than the continental ones in winter.

The five clusters were further analyzed to study the influence of air mass history on aerosol volatility. The number fractions of CV, HV, MV and LV of the six selected diameters in VTDMA measurements are regrouped based on the clusters as shown in Fig. 8. The total number fractions of the non-volatile residuals (sum of HV, MV and LV) were similar in all clusters. The maritime air masses (cluster 2) had a slightly higher fraction of LV particles, while the continental air masses originating from the northwest of the site (clusters 4 and 5) had a higher fraction of HV particles. Although the air masses in clusters 1 and 5 originated from further away and traveled at relatively higher speeds than those in clusters 2, 3 and 4, all the clusters involved transport at low altitudes (below 1500 m) for over 40 h, likely due to the generally lower mixing heights in winter. Therefore, it is plausible that the aerosol particles in these air masses were all well aged upon arrival. Similar results were observed in Beijing by Wehner et al. (2009). This could be another reason for the lack of size dependence of the number, volume fractions and diurnal variation for the particles larger than 80 nm. When the transported air masses mixed with the local pollutants, the size dependence of the number fractions of different volatility groups as well as the aging of local emissions was further reduced.

We also examined the volatility shrink factor (VSF) distributions of 40, 110 and 300 nm particles upon heating at 300 ∘C (Fig. 9). Log-normal fittings with a three-peak solution were applied to the distributions. The average VSF modes of the peaks were located at 0.38 ± 0.021 (peak 1), 0.60 ± 0.066 (peak 2) and 0.95 ± 0.007 (peak 3), respectively. The standard deviation of the corresponding normal distribution (σ) of peak 3 was the smallest among the three peaks (σ < 0.1). For the same particle size, the VSF distributions in the VSF range between 0.3 and 0.8 in cluster 5 were relatively more uni-modal than those of other clusters (Fig. 9b and c). This suggests that the composition in cluster 5 was more homogeneous. Cluster 1 also consisted of long-range transported air masses, but they likely passed through areas that are more polluted and mixed with different types of pollutants.

Two new particle formation (NPF) events were observed in the campaign on 20 February and 13 March 2014 (Fig. 3). Since VTDMA data were not available during the NPF event on 13 March 2014, we only focus on the NPF event on 20 February 2014 that happened after a cold front under a low PM2.5 concentration. On 20 February 2014, a sub-20 nm particle mode was first observed at 12:00. This particle mode grew continuously until it reached 120 nm at 02:00 on 21 February 2014. In the VTDMA data, a sharp increase in the number concentration of HV particles having an initial diameter of 40 nm was observed at 17:00 on 20 February 2014 (Fig. 10). This event is likely related to the growth of the newly formed particles when they mixed with the volatile material accumulated via condensation or adsorption. The volatile material that extensively condensed on the pre-existing particles could be sulfate, ammonium and organics. They were found to be the major species contributing to particle growth in the NPF events at different locations (Zhang et al., 2004; Smith et al., 2008; Zhang et al., 2011; Yue et al., 2016). Zhang et al. (2004) observed that sulfate was always the first and the fastest species to increase in concentration during an NPF event. They also suggested that photochemically formed secondary organics contributed significantly to the growth of the ultrafine particles. Recently, Yue et al. (2016) reported that sulfate, ammonium and organics were the main contributors to particle growth in the NPF events in Taoyuan of the PRD region. As these particles aged further, they grew larger, as reflected in the increase in number concentrations of larger MV particles and the increase in PM2.5 mass (Fig. 10). Similar results were also observed in the study in Beijing by Wehner et al. (2009). Furthermore, the growth of the newly formed particles can also be observed from the number size distributions of the HV, MV and LV particles at different times on 20 and 21 February 2014 (Fig. 11). The mode of the HV particles increased from 40 nm at 17:00 to 80 nm at 21:00 on 20 February 2014. The mode stayed at 80 nm, while the corresponding number concentration decreased at 02:00 on 21 February 2014. In contrast, the number concentration and diameter mode of the MV particles grew continuously. The HV and MV particle concentrations and diameter modes underwent much smaller changes on the non-event day of 28 February 2014 (Fig. 11).

The closure analysis of EC or the sum of EC, OC2, OC3, and OC4 and the total mass of LV and MV residuals was conducted (Fig. 12). Good correlations (R2 > 0.9) for both EC and the sum of EC, OC2, OC3, and OC4 with the total mass of LV and MV residuals were obtained. Nonetheless, the slope for the total mass of LV and MV residuals to the mass of EC (2.94) is more than 2 times of that for the total mass of LV and MV residuals to the sum of EC, OC2, OC3, and OC4 (1.22), indicating that EC alone cannot account for the total mass of LV and MV residuals. Including non-volatile OC (sum of OC2 to OC4) gave a better mass closure with the total of LV and MV residuals. This further supports our initial postulation that the non-volatile residuals that remained intact upon heating at 300 ∘C in the VTDMA may contain a significant amount of non-volatile OC. However, the total mass of EC, OC2, OC3, and OC4 did not explain all the mass of LV and MV residuals. A possible explanation could be that the vaporizing temperatures of some OC1 are close to the upper limit (310 ∘C); hence, they were not completely vaporized in the heated tube and remained in the non-volatile residuals. The presence of other refractory material and the assumption made about the density of LV and MV are two other possible explanations.

Other possible errors for the closure could be related to the different heating environments in the VTDMA and the OC / EC analyzer. In the OC / EC analyzer, OC was measured when the samples were heated in the presence of a non-oxidative carrier gas (He). In the VTDMA, aerosols were heated in air that contained O2. Therefore, some “OC2-4” that evaporated at 475 ∘C or above in the OC / EC analyzer may have been oxidized at 300 ∘C in the VTDMA. Charring of organic matter could also occur (Philippin et al., 2004). Further study is needed to quantify the effect of oxygen on the oxidation of OC in the VTDMA. The extrapolated log-normal fitting of the size distribution of non-volatile particles can also cause errors if the mode diameter of the fitting is beyond the VTDMA's range of measurements. While the VTDMA measured the size distribution of particles between 10 nm and 400 nm in diameter, the OC / EC analyzer took into account particles up to 2.5 µm in diameter. Yu et al. (2010) reported three EC and OC modes between 400 nm and 10 µm in ambient aerosols in Guangzhou: 400 nm, 900 nm and 5 µm. The 400 nm mode accounted for 44 to 49 % of the measured EC, but only 17 to 20 % of the measured OC.