A Teflon-coated cyclone (URG, model 2000-30ED) with a cutoff diameter of 2.5 µm was installed in the sampling inlet of the SP-AMS. A detailed description of SP-AMS has been reported by Onasch et al. (2012); thus, only a brief description of its working principle is given in this section. SP-AMS is based on the design of a high-resolution time-of-flight aerosol mass spectrometer (HR-ToF-MS; DeCarlo et al., 2006). HR-ToF-AMS can quantify NR-PM by flash vaporization of aerosol particles at 600 ∘C on a resistively heated tungsten surface. SP-AMS has an additional Nd-YAG intra-cavity infrared (IR) laser module at the wavelength of 1064 nm. rBC absorbs strongly at this wavelength and consequently heats up to ∼4000 ∘C for rBC vaporization (Onasch et al., 2012). The term rBC is an operationally defined term that applies to BC particles detected by SP-AMS. The resulting gas-phase molecules are ionized using the electron impact (EI) ionization method at 70 eV and are analyzed by a ToF-MS. Non-refractory and refractory species that are internally mixed with soot particles, such as organics, inorganics, or trace elements, can be detected via laser vaporizer measurements (Carbone et al., 2015; Corbin et al., 2018; Onasch et al., 2012). In the present study, the SP-AMS was operated at 1 min alternating intervals between the IR laser-on (i.e., dual vaporizers) and laser-off modes (i.e., tungsten vaporizer only), with a mass spectral resolving power of approximately 2000 at m/z 28. The vacuum aerodynamic diameter (dva) of ambient particles is determined by the measured particle time-of-flight (PToF) from the chopper wheel to the tungsten vaporizer (Jayne et al., 2000).

Calibrations of the SP-AMS for quantifying NR-PM in the laser-off mode were performed by generating dried monodispersed (300 nm) ammonium nitrate (NH4NO3) particles. The NH4NO3 solution was atomized by a constant output atomizer (TSI, Model 3076). The aqueous droplets were subsequently dried by passing them through a diffusion dryer and were size-selected using a differential mobility analyzer (DMA, TSI Inc., Model 3081). The ionization efficiency (IENO3) and the mass-based ionization efficiency (mIENO3) were determined based on three sets of calibration data throughout the sampling period. Similarly, dried, monodisperse (300 nm) BC particles generated by atomizing the standard Regal black (Regal 400R pigment black, Cabot Corp., recommended by Onasch et al., 2012) were used to determine the mass-based ionization efficiency of rBC (mIErBC) and its ionization efficiency relative to nitrate (RIErBC = mIErBC / mIENO3) for the quantification of rBC in the laser-on mode.

The calibration and field data were processed using the Igor-based AMS data analysis software SQUIRREL v. 1.16I for unit mass resolution (UMR) data and PIKA, v. 1.57I for high-resolution peak fitting (Sueper, 2015) with the corrected air fragment column of the standard fragmentation table based on particle-free ambient air data (Allan et al., 2004; DeCarlo et al., 2006). The sum of the carbon ion clusters C1+–C9+ measured in the laser-on mode was used for quantifying rBC mass in both calibration and ambient data. The average C1+/C3+ ratio (0.65) obtained from Regal black was used to correct the non-refractory organic interference in C1+ for the quantification of total rBC mass loadings. The campaign averages of RIErBC and RIENH4 were 0.15 (±0.04) and 4.24 (±0.04), respectively. The default RIE values of nitrate (1.1), sulfate (1.2), and organics (1.4) were used for the respective mass concentration quantification (Jimenez, 2003). NR-PM was quantified based on the laser-off mode measurements, and the collection efficiency (CE) was determined using the composition-dependent approach (CDCE, Fig. S2a) (Middlebrook et al., 2012). Note that the sample stream was dried to less than 30 % RH using a Nafion dryer (PD-200T-12 MPS, Perma Pure) throughout the sampling period in order to reduce CE uncertainties due to particle bouncing on the surface of the tungsten vaporizer. The SP-AMS measurements were compared with the sulfate and organic matter (OM) concentrations measured by the MARGA and the OC/EC analyzer, respectively, showing strong temporal (r values of 0.77 and 0.93, respectively) and quantitative (slopes of 0.81 and 0.88, respectively) agreements between these measurements (Fig. S3a, b). A CE of 0.6 was used for rBC quantification due to an incomplete overlap between the laser vaporizer and the particle beam (Willis et al., 2014). The corrected rBC concentrations were also comparable to those measured by the Aethalometer (r value of 0.96 and a slope of 0.83) and the EC measured by the OC/EC analyzer (r value of 0.84 and slope of 1.1) (Fig. S3c, d). Potassium (K+), sodium (Na+), vanadium (V+), nickel (Ni+), and rubidium (Rb+) ions were detected (Table S2). The signals of these trace metal ions were not calibrated; thus, their raw signals were used to investigate their temporal variations.

Positive matrix factorization (PMF) is widely used to identify sources of OA measured by different versions of aerosol mass spectrometers (Ulbrich et al., 2009; Zhang et al., 2011). Each factor contains signals at m/z values, which have common source characteristics and chemical properties that can be used to trace the origin and processing history of that factor. The PMF evaluation tool (PET, v3.00D; Ulbrich et al., 2009) was used to identify the potential sources and characteristics of OA based on the organic fragments measured in the laser-off mode. In the process of applying PMF, the ions with a signal-to-noise ratio (SNR) of 0.2< SNR <2.0 were downweighted by a factor of 2. CO2+-related ions (O+, HO+, H2O+, and CO+) were downweighted, and bad ions (SNR <0.2) were removed from the analysis. The results were obtained for 8 to 10 factors with an F peak varying from -1 to 1 and increasing by a step size of 0.2. Elemental ratios (O/C, H/C, and N/C) of each factor were determined using the improved-ambient method, as described by Canagaratna et al. (2015b). A five-factor solution was selected as final result from the OA laser-off PMF analysis.

PMF analysis was further applied to determine how rBC was associated with different OA factors. For this purpose, the PMF inputs were generated using (1) organic fragments measured by the laser-on mode (Fig. S4a–c) and (2) both organic and Cn+ fragments measured by the laser-on mode (Fig. S4d–f). The Cn+ in the mass spectra of each PMF factor was used to quantify the fraction of rBC associated with the identified OA components. Note that the interference of non-refractory organic signals on refractory Cn+ fragments was subtracted based on the method described in Wang et al. (2018); hence, the C1+ fragment was not corrected based on the C3+ fragment in the PMF analysis. Lastly, five metal ions (K+, Na+, Ni+, V+, and Rb+) were included in the PMF model (Fig. S4g–i), as the majority of their signals were higher than their respective limit of detection (Table S2). The metal ion signals were corrected to nitrate equivalent mass concentrations by assuming that their RIE values were equal to 1, and the K+ signals were downweighted by a factor 2. A similar approach was applied by Carbone et al. (2019), who included rBC fragments, calcium (42Ca+), and zinc (68Zn+) isotopic ions to distinguish between lubricating oil and fuel OA types emitted at the exhaust of diesel engines. Note that the correction approach for obtaining refractory Cn+ signals can only be applied when both the mass spectra and time series of each OA component identified in the laser-off and laser-on PMF analysis are similar. The scatter plots (Fig. S5) and correlation coefficients (r; Table S3) highlight the consistency of the mass spectra (0.91<r<0.99), as well as time series (0.78<r<0.98), of the five OA types identified in this work. The time-dependent collection efficiency (or CDCE) of rBC was determined based on the BC measured by the Aethalometer (Fig. S2b). Additional PMF analyses, with the CDCE applied on the input matrix for both the OA and refractory components, were also conducted (see details in the Supplement, S1). The PMF results were compared with those obtained without the application of the CDCE on the input matrix, as shown in Tables 1 and S4.

The origins of air masses were investigated using 5 d back trajectories generated every hour (648 back trajectories in total) at a height of 64 m with the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model coupled with meteorological data from the Global Data Assimilation System (GDAS, 1∘). This model was developed by the National Oceanic and Atmospheric Administration (NOAA; Draxler and Rolph, 2003). A cluster analysis was performed on all of the hourly back trajectories, and the three clusters identified are shown in Fig. S1c. Details on the back-trajectory clustering can be found in Baker (2010) and Borge et al. (2007).

The potential sources of pollutants were investigated using ZeFir nonparametric wind regression (NWR) and potential source contribution function (PSCF) package developed by Petit et al. (2017) using Igor Pro. These analytical methods require high time resolution atmospheric measurements as inputs, and they combine them with local wind measurements for NWR and with air mass back trajectories for PSCF. The NWR is commonly employed to identify nearby sources by coupling atmospheric species concentrations with co-located wind speed and direction (Henry et al., 2009). This method consists of weighting the average concentrations with each wind speed and direction pair and using two kernel smoothing functions. In this study, the NWR graphs were generated for an angle resolution of 1∘ and a radial resolution of 0.1 m s-1, and their respective smoothing parameters (or kernel parameters) were empirically set at 5 and 2.5. For regional source investigation, the PSCF has been favored, as larger geographical scales can be studied (Polissar et al., 2001). The PSCF principle consists of redistributing high pollutant concentrations based on air mass trajectories' residence times. Thus, individual species concentrations are redistributed along the trajectories into geographical emission parcels. In our study, the PSCF calculations were performed considering only pollutant concentrations above their respective 75th percentile (graphic parameters: cell size of 0.1∘ and smoothing of 2).

HOA was the main contributor to the POA fraction and was moderately correlated with the combustion tracers including rBC (r value of 0.79), NOx (r value of 0.83), and CO (r value of 0.81). The mass spectrum of HOA is similar to those from engine exhausts (Crippa et al., 2013; Ng et al., 2011; Zhang et al., 2005) and is dominated by non-oxygenated fragments (i.e., CxHy) with average H/C and O/C ratios of 2.12 and 0.07, respectively. HOA accounted for up to ∼20 % of total OA during the morning (∼09:00 LT) which matched the enhancement of the traffic volume during rush hours. In addition, the concentrations of HOA started increasing at the beginning of the evening rush hour (∼17:00 LT) and reached a maximum at ∼21:00 LT. HOA and combustion tracers (rBC, NOx, and CO) exhibited a hot spot in the NWR plots (Figs. 2; S7e, f) under relatively low wind speed conditions. The above observations suggest that HOA concentrations were largely influenced by nearby on-road engine exhausts and the contraction of the boundary layer in the evening.

High concentrations of rBC, NOx, and HOA were observed during the middle of the night between 25 and 28 May, indicating the presence of large combustion sources during nighttime that might occasionally impact the air quality at our sampling site (Fig. 1a). This observation was also reflected in the diurnal plots (Figs. 3, S8), as the mean mass concentrations of HOA, rBC, NOx, and CO during the middle of the night were much higher than their corresponding median values. Figure S8a shows that the highest N/C ratio was observed during the morning traffic peak hours, but it remained relatively low for the rest of the period. This observation suggests that the emission characteristics of these unknown combustion sources at nighttime could be different from those associated with traffic emissions during the daytime. Although it is difficult to confirm the origin of the nighttime combustion emissions without further evidence, the observed events were coupled with low wind speed conditions, suggesting that they were likely due to local emissions from the city. For example, significant emissions due to flaring from petrochemical industries, which largely depends on plant operation, during relatively stagnant atmospheric conditions could lead to elevated concentrations of combustion-related species.

O-HOA showed distinctive time series and oxygenation level compared with HOA (r of 0.17 and O/C of 0.07). The mass spectral profile (Fig. 3) illustrates that O-HOA could be an oxygenated fraction of combustion particles that was co-emitted with HOA and/or produced rapidly via the oxidation of HOA near emission sources. However, similar to the case of rBC, the elevated concentrations of O-HOA observed from the southwest direction with a strong wind speed (∼18 m s-1) suggests that O-HOA was partly related to the emissions transported from the industrial zone (Fig. 2). The temporal variability of O-HOA depicts a rather singular behavior, showing a higher average concentration before 24 May (2.2 µg m-3). During rest of the campaign, the principal source of O-HOA was mainly local anthropogenic activities (Fig. S7a–d) and remained at a rather low concentration, with an average of 0.9 µg m-3. The relatively constant diurnal cycle of O-HOA compared with that of HOA further underlines multiple sources for this OA component.

COA is another major POA component identified in this study. COA concentrations remained low in the morning and exhibited peaks around lunchtime (13:00 LT) and dinnertime (21:00 LT), as seen in Fig. 3. This type of COA diurnal pattern has been commonly observed in urban studies (Allan et al., 2010; Sun et al., 2011; Fröhlich et al., 2015). Furthermore, the COA mass spectrum shows a higher m/z 55/57 ratio than HOA, as has been reported in previous studies (Allan et al., 2010; Mohr et al., 2012), and is strongly correlated with the COA mass spectrum reported in France (Crippa et al., 2013) and in the megacities of China (Hu et al., 2013; Elser et al., 2016). An organic fragment tracer, C6H10O+, tracked the time profile of COA (r2 of 0.90), further confirming the existence of COA components (Zhang et al., 2011). A local COA hot spot shown in the NWR plots (Fig. 2) suggests that the campus canteens and a residential building nearby are likely the major contributors to the observed COA mass loadings. Note that COA has been shown to contribute up to 50 % of the total OA in other urban locations (e.g., downtown area with substantial emissions from restaurants) depending on the site characteristics (Allan et al., 2010; Huang et al., 2010; Sun et al., 2012; Crippa et al., 2013) and cooking styles. Similar to other combustion-related emissions, high concentrations of the COA component were also observed during the middle of the night between 25 and 28 May (Fig. 1). This observation implies that combustion-related emissions might contain some COA-like particles, contributing to the reported mass concentrations of COA in general. After removing the data from 25 and 28 May, the diurnal profile shows that lower mass concentrations of COA were observed overnight (Fig. S8d).

The mass spectra of LO-OOA was dominated by the oxygenated fragments of C2H3O+ and CO2+, whereas MO-OOA was mainly characterized by CO2+ fragments (Fig. 3). The different degrees of oxygenation between LO-OOA and MO-OOA and their secondary nature are shown in the van Krevelen diagram (Fig. S9) (Heald et al., 2010; Ng et al., 2011). MO-OOA (O/C ratio of 0.94 and N/C ratio of 0.01) likely represented a more aged fraction of SOA. The MO-OOA concentrations were insensitive to the local wind pattern (i.e., more homogeneous distribution in the NWR plot, Fig. 2) and increased gradually in the afternoon, possibly due to the combined effects of photochemistry and the mixing of air masses transported from regions without significant local industrial influences. Similar to many other studies, the source identification of the MO-OOA component is not straightforward, as their mass spectral features can be obtained by the oxidation of different types of POA and SOA components (Donahue et al., 2009; Jimenez et al., 2009; Zhang et al., 2007b). Additional information is essential to further interpret their potential sources and/or formation mechanisms (see discussion in Sect. 3.5).

In contrast, the moderate correlation observed between LO-OOA and SO42- (r value of 0.73) and the similarity between their diurnal patterns illustrate that LO-OOA were freshly formed SOA materials (O/C ratio of 0.41). The latter might be formed via the local photooxidation of POA and/or SOA precursors under the influence of anthropogenic emissions transported from the southwest by the sea breeze (Fig. 2). Previous laboratory studies have evidenced the importance of acidic seeds and RH on SOA formation (Liggio and Li, 2006; Wong et al., 2015). The acidic nature of sulfate particles and RH encountered on site might facilitate the partitioning of volatile organic compounds (VOCs, from both biogenic and anthropogenic sources) into the particle-phase, leading to the production of SOA and, perhaps, organosulfur compounds, as discussed in Sect. 3.2. Furthermore, Kasthuriarachchi et al. (2020) reported that the concentrations of organonitrate and nitrogen-containing fragments (i.e., CxHyN+ and CxHyNOz+) slightly increased during daytime (e.g., Fig. S8a) with LO-OOA, contributing to 30 % of the observed CxHyNOz+ fragments (Fig. S10). This suggests that the photooxidation of VOCs under high-NOx conditions could be a potential pathway toward LO-OOA formation given that the NOx concentrations reached up to 160 ppb (Fig. 1a; mean of 16.5 ppb) during the campaign.

Based on the PMF results that include trace metal ions, sodium (Na+) was mainly associated with HOA, O-HOA, and LO-OOA (Fig. 5a); this may have been due to different types of fossil fuel combustion emissions (e.g., local traffic, shipping, and various industrial activities), as discussed in Sect. 3.4. The Na+ signals also correlated strongly with rBC (r value of 0.80, Fig. 5b) and moderately with HOA (r value of 0.65). The absence of a correlation between rBC and Na+ in the laser-off mode measurement suggests a large degree of internal mixing between Na and rBC. The presence of sodium compounds in fuel, including remaining catalyst used for biodiesel esterification, drying agents, corrosion inhibitors, and fuel additives, has previously been observed as a cause of fuel injector fouling (Barker et al., 2013; Coordinating Research Council, 2013); hence, the co-emissions of sodium with rBC and HOA from on-road engines was highly possible. Recent experiments conducted with SP-AMS measurements have also reported non-negligible amounts of Na in soot particles emitted by marine, locomotive, and vehicle engines (Dallmann et al., 2014; Omidvarborna et al., 2016; Saarikoski et al., 2017; Corbin et al., 2018; Carbone et al., 2019). Na+ is widely used to identify the influence of marine sources in source apportionment analysis; hence, the potential contribution of sea spray aerosols to Na+ signals cannot be neglected, especially for factors (i.e., O-HOA and LO-OOA) connected to sea breeze transport. However, it is important to emphasize that Na+ and Cl- exhibit rather poor temporal correlations (r value less than 0.30) for both laser-off and laser-on data regardless of the influence of sea breeze. Biomass burning can be a possible source of Na+ (Hsu et al., 2011), although no major fresh biomass burning emissions were observed in this study. The MO-OOA factor is suspected to be more influenced by aged regional biomass burning emissions (see further discussion in Sect. 3.5.2), but Na+ was not strongly associated with this factor.

Vanadium (V) and nickel (Ni) are usually residual trace metals in fuel, and their ratio in emissions can vary with the fuel type or combustion process observed (Moldanová et al., 2009; Yakubov et al., 2016). This ratio can be used to trace emissions from ships or heavy oil combustion industries (Ault et al., 2010; Liu et al., 2017). In particular, Fig. 5c shows that the V+/Ni+ ratio exhibited a distinctive diurnal variation, with higher values from 12:00 to 18:00 LT, which coincided with the sea breeze pattern. Two respective maxima are observed at 13:00 and 17:00 LT, and the hourly averaged V+/Ni+ ratio is between 5 and 7 (75th percentiles >12; Fig. 5c). Assuming a similar RIE for V+ and Ni+ measured by the SP-AMS, our observation is consistent with other studies that have reported V+/Ni+ ratios ranging from 4 to 7 for ship emissions and heavy fuel combustion by engines (Agrawal et al., 2009; Corbin et al., 2018; Viana et al., 2008). In addition to a moderate correlation between LO-OOA and V+ (r value of 0.47), ∼70 % of the V+ signals were associated with the LO-OOA component (Fig. 5a). Figure 5d clearly shows that the V+/Ni+ ratios increased with the concentrations of SO42- and LO-OOA under southwesterly sea breeze conditions. These observations confirm that a part of the SO42- and LO-OOA encountered on site was influenced by ship emissions and/or heavy oil combustion from the industrial zone.

Although a large local source of biomass burning in Singapore is uncommon, it is important to investigate the potential influences of biomass burning events, including agricultural burning and forest fires, from nearby countries (Fig. S12a). Budisulistiorini et al. (2018) reported high mass loadings of biomass burning OA (BBOA) during an Indonesian wildfire in 2015 which had strong impacts on the air quality in Southeast Asia. BBOA is not identified in the PMF analysis in this field study. However, the average fraction of C2H4O2+ to total organic aerosol (fC2H4O2+) of 0.8 % (Fig. S13a) with ∼86 % of the data giving fC2H4O2+ values above 0.3 % (i.e., a background fC2H4O2+ values for non-biomass burning; Cubison et al., 2011) in the laser-off mode measurement suggests potential influences from aged biomass burning emissions that cannot be easily separated by the conventional PMF analysis of organic fragments alone. Note that MO-OOA and COA present slightly higher fC2H4O2+ values of 0.5 % and 0.8 %, respectively (Fig. 6a). However, the origin of MO-OOA is highly uncertain compared with the other OA factors identified in this study due to the fact that atmospheric aging can diminish the distinctive mass spectral features of POA in general, e.g., decreasing in fC2H4O2+ for BBOA (Cubison et al., 2011) and converging in the f43–f44 framework for OOA production (Ng et al., 2010). Therefore, the following discussion will focus on evaluating the potential connections between the MO-OOA component and aged regional emissions.

Figures 6c and S14b show that high Rb+ and K+ signals were associated with the more oxygenated (f44>0.7) fraction of OA, and moderate correlations between MO-OOA and the two metal were observed (Rb+ had an r value of 0.58 and K+ had an r value of 0.71; Fig. S13b). Furthermore, the results of the PMF analysis demonstrated that both K+ and Rb+ were mainly associated with MO-OOA (58 %–66 %, Fig. 5a) followed by the two combustion-related components (HOA and O-HOA). Although potassium and rubidium are not unique tracers for a specific combustion source, previous studies have shown that these two metals can be largely associated with biomass burning emissions (Artaxo et al., 1993; Lee et al., 2016; Achad et al., 2018). Note that rubidium has also been used as a coal combustion tracer in previous studies (Fine et al., 2004; Irei et al., 2014). Unlike the ambient OA component, the chemical identities of K+ and Rb+ are unlikely to be modified by the oxidative aging of aerosol particles. Therefore, a strong temporal correlation between Rb+ and K+ (r value of 0.85, Fig. 6b) further suggests that they were likely of similar origins in this study.

The regional origin of K+, Rb+, C2H4O2+, and MO-OOA was investigated using their PSCF. Their PSCF graphs (Figs. 6d, S15a–c) show several common origins with the high probability that the highest concentrations could be influenced by biomass burning events from Indonesia (Fig. S12a). Nevertheless, coal-fired power plants are located near the identified hot spots of Rb+ and K+ (Fig.  S12b), so that a regional transport of coal-fired power plant emissions alongside biomass burning plumes would be possible. Note that MO-OOA contributes to the highest fraction of nitrogen-containing organic fragments (CxHyNOz+ ∼32 % and CxHyN+ ∼46 %) that can be generated by biomass burning emissions (Mace et al., 2003; Laskin et al., 2009; Desyaterik et al., 2013; Mohr et al., 2013). It is important to point out that most of the previous studies usually describe MO-OOA (or LV-OOA in some earlier studies) as an aged SOA component without providing further details regarding its potential origin or emission characteristics. Our observations underline the possibility of better understanding the origin of the MO-OOA component via measurements of refractory metals, even when atmospheric oxidative processing has made the mass spectral features of aged OA materials less distinguishable.