The use of the EESI-TOF allows continuous and highly time-resolved measurements of organic aerosol composition on a near-molecular level (i.e. chemical formulae of molecular ions) with negligible thermal decomposition or ionization-induced fragmentation. A detailed description of the EESI-TOF and its operating principles has been provided elsewhere (Lopez-Hilfiker et al., 2019). Briefly, the sampled aerosols first pass through a multi-channel charcoal denuder to strip gaseous components. Breakthrough of gas-phase species can cause high background signals, as previously observed in Beijing (Tong et al., 2021). In that study, a small denuder (diameter of 4 mm and length of 30 to 40 mm) was used, whereas in the present study we used a larger denuder (i.e. with 69 channels, outer diameter 8.5 mm, length 60 mm). The denuder was exchanged every 48 h and regenerated by baking at ∼200 ∘C for 12 h. The denuder was positioned ∼20 cm upstream of the electrospray and mass spectrometer inlet to avoid decreased transmission of larger particles (Tong et al., 2021). After passing through the denuder, the sampled particles intersect with a plume of electrospray (ES) droplets generated by a commercial electrospray probe and delivered through a fused silica capillary with precut tips with an inner diameter of 50 µm (BGB Analytik AG). The ES solution used in this study consisted of a 1:1 water / acetonitrile (v/v) mixture doped with 100 ppm NaI. The ES solution was charged by applying a high potential difference (∼2.5–2.7 kV) at the ES capillary tip. The flow of ES solution through the silica capillary is controlled by a high-accuracy microfluidic flow controller (Fluigent GmbH). Upon intersection with ES droplets, the soluble fraction of analyte aerosol is extracted into the droplets. The analyte-laden droplets then pass through a heated stainless-steel capillary (∼250 ∘C), wherein the electrospray solvent evaporates, and analyte ions are generated. Due to the short residence time (∼1 ms) in the capillary, heat transfer to the particles is limited, and negligible thermal decomposition is observed. Finally, the ions are analysed by a high-resolution long-time-of-flight mass spectrometer (LTOF-MS, Tofwerk AG, Switzerland) configured for positive ion detection. In the configuration of the mass spectrometer and ionization scheme used in this study, one can detect a wide range of molecules present in the organic aerosols, including sugars, alcohols, acids, and organonitrates. The detected molecular classes include nearly all the compounds present in the secondary organic aerosol (SOA), with notable exceptions of organosulfates, which are typically detected as negative ions, and non-oxygenated species such as alkanes and alkenes. The mass resolution (M/ΔM) achieved by the mass analyser in this study was ∼8000. This resolution is enough to separate isobars (compounds with same nominal mass) and determine the molecular formula. Due to the lack of direct structural information, the molecular assignments given here are, however, tentative and may include contributions from multiple isomers. The ions were observed predominantly as adducts with Na+, i.e. [M]Na+, but for simplicity are denoted herein using the neutral formula (M). For example, an ion observed as [C6H10O5]Na+ is reported as C6H10O5.

The EESI-TOF sampled continuously at a flow rate of ∼1 L min-1, alternating between direct ambient sampling (5 min) and sampling through a particle filter (3 min) for measurement of the instrument background. This measurement–filter cycle is shorter than those used in previous ambient measurements to avoid clogging of the EESI capillary and to increase the spray stability. The EESI-TOF data analysis was performed using Tofware version 2.5.7 (Tofwerk AG, Switzerland). The original data were acquired at a time resolution of 1 Hz, and high-resolution (HR) peak fitting was applied to data averaged to 10 s for the mass-to-charge ratio (m/z) range 145–350. In total, 1030 ion formulae were fitted. The ambient aerosol composition (Mdiff) was calculated by subtracting background spectra obtained during particle filter sampling (Mfilter) from the mass spectra obtained during direct ambient sampling (Mtotal). The background spectra (Mfilter) were calculated by interpolating the average between adjacent background spectra (i.e. particle filter measurements before and after an ambient sampling period). This methodology is similar to the one reported in previous EESI-TOF studies (Qi et al., 2019; Stefenelli et al., 2019; Tong et al., 2021). After obtaining the ambient aerosol composition, the ambient aerosol spectra (Mdiff) were averaged to 10 min for further processing. The error matrix (σdiff) corresponding to Mdiff values was calculated from Poisson ion counting statistics (Allan et al., 2003) from ambient sampling σtotal(i,j) and filter sampling periods σfilter(i,j), added in quadrature as follows: 1σdiffi,j=σtotal2i,j+σfilter2i,j. As a final step, data were filtered to remove ions whose total signal and/or variability was either dominated by the background or too noisy for meaningful interpretation. Ions that met both of the following criteria were accepted for further analysis:

The ratio of signal to uncertainty, i.e. Mdiff/σdiff, was considered, where σdiff represents the precision-based uncertainties as calculated by using Eq. (1). Ions with a median Mdiff/σdiff<0.2 were removed from further analysis (Paatero and Hopke, 2003).

The HR-AMS (Aerodyne Research Inc., Billerica, MA, USA) was equipped with a PM1 aerodynamic lens and measured the composition of NR-PM1. A detailed description is given elsewhere (DeCarlo et al., 2006; Canagaratna et al., 2007). Briefly, ambient air is sampled continuously through a critical orifice into a PM1 aerodynamic lens, which focuses the particles into a narrow beam and accelerates them to a velocity that is inversely related to their vacuum aerodynamic diameter (Williams et al., 2013). The particles then impact a resistively heated surface (∼600 ∘C) and flash vaporize. The resulting gas is ionized by electron impact ionization (EI, 70 eV) and detected by a time-of-flight mass spectrometer. The detected ion rate measured at a specific mass-to-charge ratio (m/z) is then converted to mass concentration in µg m-3 (Jimenez et al., 2003).

In this study, two ionization efficiency (IE) calibrations of the instrument were performed (one before the commencement of the campaign and the other at the end) using 300 nm NH4NO3 particles. More details on the instrument operation during this campaign have been reported elsewhere (Singh et al., 2021, 2019). The instrument was operated in V mode at a time resolution of 2 min. Every 30 s, it switched between mass spectra (MS) and particle time-of-flight (PToF) mode, completing two cycles within each integration period. The mass spectra of an ensemble of particles are measured in MS mode, whereas in the PToF mode, the particle beam is modulated by a chopper spinning at 130 Hz, resulting in the size-resolved mass spectra. Unit mass resolution (UMR) data were analysed using the SQUIRREL data analysis toolkit (version 1.59) programmed in the Igor Pro 6.37 software environment (WaveMetrics, Inc., Portland, OR, USA). High-resolution peak fitting analysis was conducted using PIKA (version 1.19) (DeCarlo et al., 2006) for m/z 12 to 120. The collection efficiency was estimated using the composition-dependent algorithm of Middlebrook et al. (2012) implemented in SQUIRREL. The Pieber correction was applied according to the method recommended by Pieber et al. (2016).

Source apportionment was performed separately on the AMS OA and EESI-TOF datasets using the positive matrix factorization (PMF) algorithm (Paatero and Tapper, 1994) implemented within the multilinear engine (ME-2; Paatero, 1999). In this study, the Source Finder Professional (SoFi Pro 6.8, Datalystica Ltd.) interface was used for model configuration and post-analysis (Canonaco et al., 2013). PMF is a bilinear model that represents the sample matrix X with dimensions of m×n, representing m measurements of n variables as a product of two matrices G (dimensions of m×p) and F (dimensions of p×n). The number of columns in the modelled matrix G and rows in the modelled matrix F are equal to the number of factors p, i.e. individual sources chosen to describe the dataset. The PMF model operates under non-negativity constraints; i.e. negative values are not permitted in G or F. The PMF model is expressed as 2X=G×F+E. Here G represents the time-dependent factor concentrations (i.e. time series) and F represents the chemical composition (i.e. mass spectrum) of the resolved factors. Model residuals are contained in E.

The PMF model solves Eq. (2) using a least-squares algorithm that iteratively minimizes the objective function Q, defined as 3Q=∑i=1n∑j=1meijσij2. In Eq. (3), eij represents elements of the residual matrix and σij represents the measurement uncertainties corresponding to the input point xij, where i and j are the indices representing measurement time and variable (or integer m/z), respectively. The theoretical value of Q, denoted Qexp, can be estimated as 4Qexp=mn-pm+n. PMF is subject to rotational ambiguity, meaning that different combinations of G and F matrices exist that can yield the same or similar Q values. Some of these combinations may represent environmentally unreasonable representations of the dataset. To direct the model towards interpretable rotations, a priori information can be introduced by constraining selected factor time series or mass spectra using an a-value approach (Canonaco et al., 2013; Crippa et al., 2014). In this method, one or more factor profiles and/or time series are constrained to resemble reference profiles and/or time series, with the scalar a (0≤a≤1) determining the tightness of constraint. As an example, if constraints are applied to mass spectra, the a value determines the extent to which a factor mass spectrum in the final solution (fj,solution) is allowed to deviate from the anchor mass spectrum (fj) provided to the model as the initial starting point. 5fj,solution=fj±a⋅fj As an example, if an a value of 0.1 is used, all the variables in the resulting mass spectrum can vary between ± 10 % of the input constraining the mass spectrum. Note that post-PMF normalization of factor profiles may cause the final values to slightly exceed the limits defined by Eq. (5).

The AMS OA matrix XAMS consisted of organic ion time series derived from high-resolution (HR) peak fitting for m/z 12 to 120 and the integrated signal across integer m/z (unit mass resolution or UMR) for m/z 121 to 300. A total of 507 variables were used in the sample matrix XAMS, 332 of which had chemical formulae assigned to them through HR fitting. The remaining 175 variables were UMR species. For the UMR data, we excluded m/z 149 due to interference from phthalic acid emitted by the servo housing, as well as m/z 183, 184, and 186 due to the interference from the tungsten filaments. Uncertainties were calculated according to the method by Allan et al. (2003), which accounts for counting statistics of the individual ions as well as the uncertainty in the detector response to individual ions. Variables with a signal-to-noise ratio (SNR) <0.2 were downweighted by a factor of 10, whereas those with an SNR <2 were downweighted by a factor of 2 (Paatero and Hopke, 2003). Further, ions calculated from the CO2+ signal (i.e. O+, OH+, H2O+, and CO+) were removed from XAMS prior to PMF analysis to avoid overweighting CO2+ intensity (Ulbrich et al., 2009) and were recalculated from CO2+ during post-analysis. Note that this remove-and-reinsert strategy is preferable to downweighting of CO2+-dependent ions as it avoids the potential for small biases induced by the combination of AMS minimum errors and dynamic downweighting in “robust mode” operation of the PMF.

As a first step, we ran the PMF in unconstrained mode with the number of factors ranging from 3 to 8. Each solution was inspected based on its Q/Qexp value and physical interpretation of individual factors. Large decreases in the Q/Qexp values were observed when the number of factors increased from 3 to 5, while small incremental changes were observed when the number of factors increased beyond 5. Further, solutions with more than 5 factors yielded only additional biomass-burning-related factors, the differences between which could not be physically interpreted. Hence we chose a 5-factor solution as the best representation of the data. The unconstrained PMF resulted in an HOA factor with a high degree of oxygenation, i.e. O:C ratio ∼ 0.15, which is a factor of ∼3 higher than the HOA factor obtained at the same site in a recent study (Lalchandani et al., 2021). To obtain a cleaner HOA profile, we took the HOA factor profile from an unconstrained 8-factor solution and used it in SoFi to constrain the HOA factor in the final 5-factor solution. We explored the PMF solutions with higher numbers of factors, but the O:C ratio of the HOA profile did not show a significant decrease for solutions with more than 8 factors.

The factors obtained from the AMS source apportionment were identified based on their correlations with external measurements, mass spectral features, diurnal trends, and relationship to anthropogenic activities as well as meteorological and environmental conditions (e.g. temperature, expected trends in human activities). The interpretation of the final 5-factor solution is discussed in Sect. 3.1.

A total of 641 ion formulae from m/z 140–350 were used in the final PMF input matrix XEESI of the EESI-TOF data. The initial PMF model was run without constraints for 6 to 15 factors, and each solution was checked for the interpretability of the results. The 6-factor solution yielded a factor identified as primary biomass burning (characterized by ∼90 % of factor profile signal from C6H10O5, which is likely dominated by levoglucosan, a biomass burning tracer), and 5 other factors related to primary cooking emissions, aged biomass burning, and 3 SOA factors (described in Sect. 3.2). Although the main spectral and temporal features of these factors were not consistent with primary biomass burning, they nonetheless contained significant signals from C6H10O5 (comprising 10 %–15 % of the factor profiles), consistent with mathematical mixing of biomass burning into these factors. Increasing the number of factors from 6 to 10 decreased the contribution of C6H10O5 in the aged biomass burning factor to ∼12 %, consistent with similar factors observed in previous studies (Qi et al., 2019; Tong et al., 2021). For the non-biomass-burning factors, the contribution of C6H10O5 to the factor profiles decreased to <2.5 %, while key spectral and temporal features were retained. As the number of factors increased, the newly added factor profiles all had high (>20 %) contributions from C6H10O5, which is characteristic of primary biomass burning. However, these new factors could not be physically interpreted and were therefore considered to result from mathematical splitting. Increasing the number of factors to 11–15 yielded only further splitting of the primary biomass burning profiles and no longer affected the C6H10O5 contributions to the non-primary biomass burning factors.

This preliminary analysis suggested that the variability in the dataset is optimally represented by 6 factors. However, because the unconstrained 6-factor solution did not provide unmixed factors (as described above), we constructed an unmixed 6-factor solution by constraining profiles for primary cooking, aged biomass burning, and the 3 SOA factors. The reference profiles used in SoFi for these 5 factors were taken from the unconstrained 10-factor solution. The remaining five profiles (from the unconstrained 10-factor solution) were combined on a mass-weighted basis to form a single primary biomass burning profile. This 6-factor solution is referred to as the “base case” hereafter.

The statistical stability of and uncertainties in the base case were accessed by a combined bootstrap analysis–randomized a-value selection (i.e. sensitivity test of the tightness of constraint). Bootstrapping was implemented by random resampling of the rows of the original data matrix and corresponding entries of the error matrix, such that in each bootstrap iteration some rows were sampled multiple times while others were not sampled at all, thus creating new matrices in each iteration of the bootstrap analysis that were of the same dimensions as the original input matrices (Davison and Hinkley, 1997; Paatero et al., 2014). Simultaneously the a values of the 5 constrained factors (primary cooking-related, aged biomass burning, and 3 unique SOA factors) were randomly selected from within predefined limits chosen to maximize exploration of the solution space while maintaining computational efficiency. The bootstrap–a-value analysis was conducted in two stages: (1) an exploratory analysis on a small number of runs that was used to determine the a-value limits and (2) the final analysis on 1000 bootstrap runs with a-value randomization occurring within these limits.

The a-value limits for the combined bootstrap–a-value randomization analysis were selected after an exploratory analysis of 250 bootstrap runs in which the a values of every constrained factor were allowed to vary over the full range (0 to 1), with a step size of 0.1. The 250 individual solutions were analysed and classified as “good” or “mixed” following the method of Stefenelli et al. (2019), which consists of the following steps: (1) calculation of the Spearman correlation coefficients between the time series of each factor from the base case and a bootstrap solution, yielding a correlation matrix for each bootstrap run with the correlation values between bootstrap factors and corresponding base case factors on the matrix diagonal; (2) requirement that the correlation coefficient on the matrix diagonal was higher than those on the intersecting row and column by a statistically significant margin (based on a preselected significance level p from a t test). Solutions satisfying this requirement were classified as good solutions, whereas those failing this test were classified as mixed solutions. From visual analysis of ∼50 randomly selected solutions, we selected p=0.3 as the appropriate confidence level. We then assessed the acceptance probability as a function of a value, selecting the a-value upper boundary to be the value above which 75 % of solutions were classified as mixed. The ranges of a values selected for cooking-related OA and 4 SOA factors are given in Supplementary Table S1.

The a-value limits obtained above were utilized in a final combined bootstrap–a-value randomization analysis, consisting of 1000 runs. Solutions resulting from this 1000-run bootstrap were separated into good and mixed solutions using the same acceptance/rejection criteria as used in the exploratory bootstrap. The final bootstrap analysis resulted in 835 good solutions out of 1000 that were kept for further analysis. The solution presented in Sect. 3 is the average of these 835 solutions.

The mass spectrum of primary biomass burning is dominated by C6H10O5, likely associated with anhydrosugars such as levoglucosan, mannosan, and galactosan. C6H10O5 constitutes 81.1 % of the total mass spectral signal in this factor (Fig. 2a). The second-highest contribution to the mass spectrum comes from the ion C8H12O6 (2.1 % of the total signal), which could possibly be a derivative of syringol, a prominent compound found in wood-burning smoke (Yee et al., 2013). These features are qualitatively similar to primary biomass burning mass spectra observed by EESI-TOFs in previous studies (Qi et al., 2019; Stefenelli et al., 2019; Tong et al., 2021). The next three highest contributing ions with 0.47 %, 0.44 %, and 0.33 % of total signal are C11H14O4, C6H12O5, and C6H10O4, respectively. C11H14O4 could be tentatively assigned to syringyl ethanone, whereas C6H10O4 may be associated with methylglutaric acid. Both of these compounds have previously been found in biomass burning smoke (Bertrand et al., 2018; Qi et al., 2019).

The time series of the primary biomass burning factor observed by the EESI-TOF (Fig. 2b) correlates strongly with the summed time series of the two BBOA factors from AMS (r=0.85). The diurnal trend of this factor shows a distinct peak during the evening rush hours between 18:00–22:00 LT and thereafter a steady decline throughout the night with an early morning rise starting between 05:00–06:00 LT and peaking at 08:00 LT before decreasing to low values during midday hours (12:00–16:00 LT) (Fig. 2c). The diurnal trend is also qualitatively similar to the AMS HOA and BBOA factors, in that the time of early morning rise coincides with an increase in anthropogenic activities. Another observation is that during the daytime (12:00–16:00 LT) the primary biomass burning declines to less than 4 % of its average concentrations from the previous night. Three effects might drive these very low daytime concentrations of primary biomass burning: first, the decline in source intensities; second, the strong dilution effects from boundary layer expansion; and third, increased evaporation of semi-volatile constituents due to higher temperature and the aforementioned dilution. The primary biomass burning factor constitutes on average around 70 % of the total EESI-TOF OA signal, whereas the AMS BBOA factors contribute on average 44.7 % to the total organic mass measured by the AMS. This is due to the higher sensitivity of the EESI-TOF towards levoglucosan as compared to most other classes of compounds, as consistently observed in laboratory and field studies (Lopez-Hilfiker et al., 2019; Stefenelli et al., 2019; Tong et al., 2021). (Note that the EESI-TOF insensitivity to HOA cannot explain this discrepancy, as AMS BBOA contributes only 50.9 % of the non-HOA organic mass, i.e. BBOA/(OA-HOA).)

A notable feature in the mass spectrum of this factor is that ∼9.2 % of the total signal in this factor comes from ions with H:C ratios >1.7 and O:C <0.25. The contribution of such ions to biogenic SOA is 5.0 %, with all other factors falling below 2.1 %. The carbon number distribution of these ions is shifted towards higher carbon numbers, consistent with saturated and non-saturated fatty acids (see Fig. 3). Such molecules are prominent constituents of cooking oils (Orsavova et al., 2015). The high relative contributions from such species are consistent with previous cooking-related factors resolved in EESI-TOF studies (Tong et al., 2021; Qi et al., 2019).

The time series of cooking-related OA explains a large fraction of the signal from ions consistent with fatty acids, e.g. 32.1 %, 38.8 %, 34.6 %, and 33.9 % of C16H32O2, C16H30O2, C18H34O2, and C18H36O2, respectively. The other five factors combined explain only 15 %–20 % of the variation in these fatty-acid-like compounds, while the rest remains unexplained; ∼5 %–10 % is explained by residuals and ∼90 %–95 % by noise. Figure S3 shows the fractional contribution of all EESI-TOF factors to the diurnal trends of two selected fatty-acid-like compounds, tentatively attributed to oleic acid (C18H34O2) and stearic acid (C18H36O2). As one can clearly see, the cooking-related OA factor is the dominant contributor to these species, regardless of time of day or ion concentration. This observation of the high contribution of cooking-related factor to observed diurnal patterns of fatty-acid-like compounds is also consistent with previously defined cooking-related OA factors from the EESI-TOF (Tong et al., 2021; Qi et al., 2019).

The diurnal trend of cooking-related OA shows qualitatively similar features to primary biomass burning, in that it decreases during daytime hours and peaks during the late evening (Fig. 2c). Cooking-related OA increases by a factor of ∼6 from 16:00–19:00 LT, remaining roughly stable till 00:00 LT, followed by a decline till 06:00 LT, and then remaining approximately stable until 10:00 LT. During the day, a small peak is observed during lunchtime (13:00–15:00 LT), which may indicate active sources in the vicinity of the measurement site. The average day-to-night ratio is approximately a factor of 10 higher than that of primary biomass burning, which further supports the possibility of active sources during the day.

We note that no cooking-related factor was identified in the AMS PMF results. The unconstrained PMF analysis yielded an HOA factor with high levels of oxygenated fragments, especially the oxygenated fragments at m/z 55 and 57, which might indicate mixing of cooking-related factor into HOA. However, it is also possible that these oxygenated fragments were contributed by some other sources (e.g. BBOA or LO-OOA); hence a definite conclusion on the mixing of the cooking-related factor into unconstrained HOA could not be drawn. Neither constraining a COA profile from the literature in the AMS PMF nor increasing the number of factors up to 15 yielded cooking-related factors. Possible reasons for this may be the similarity of the cooking-related OA spectrum with the HOA and BBOA spectra in the AMS, high relative concentrations of the other primary factors, and strong effects of boundary layer dynamics on the diurnal patterns of all factors (leading to collinearity among unrelated factors), all of which combine to make it difficult to separate a relatively minor cooking-related factor without the specific tracer ions provided by the EESI-TOF.

The mass spectrum of this factor has ∼63.0 % of its total signal contributed by compounds with six to nine carbons (C6-C9). A large fraction of this comes from molecules with an H:C ratio <1.5 (27.9 % of the total signal from C6-9HyOz ions and 7.1 % from C6-9HyOzN1-2 ions) (Fig. 3). As a comparison, biogenic SOA, aged biomass burning, and mixed urban SOA have an 11.7 %, 27.2 %, and 18.3 % contribution from C6-9HyOz ions and 4.0 %, 2.9 %, and 3.7 % contribution from C6-9HyOzN1-2 ions with H:C ratios <1.5, respectively. The low H:C ratios are associated with aromatic systems, as the precursor gases are highly unsaturated and contain fewer hydrogen atoms than more saturated straight-chain or ring-containing compounds. In a recent study on the source apportionment of VOCs at the same site, it was found that aromatic C6-9Hy VOCs constitute 45.4 % of total VOC loading and are emitted into the atmosphere predominantly from anthropogenic activities, of which traffic constituted the highest fraction during the daytime (Wang et al., 2020). Oxidation of these aromatic VOCs is most likely the dominant process leading to the formation of this factor.

In order to substantiate the claim that the major ions contributing to the ambient aromatic SOA factor are indeed formed by oxidation of aromatic VOCs, we conducted a chamber experiment in the Paul Scherrer Institute (PSI) smog chamber using a mixture of aromatic compounds consisting of benzene, toluene, ethylbenzene, and trimethylbenzene (Kumar et al., 2022). These compounds are well-established constituents of vehicular emissions (Cao et al., 2016; Yao et al., 2015) especially for gasoline vehicles during the cold start phase (Platt et al., 2017). OH radicals were produced in the chamber and reacted with the VOCs, resulting in the formation of SOA, whose chemical composition was subsequently compared with the aromatic SOA and other SOA factors obtained in this study.

Figure S2 shows the mass spectrum of the ambient aromatic SOA factor colour-coded by ions that were found in chamber SOA formed from oxidation of aromatics. Approximately 32.0 % of the EESI-TOF signal contained in the aromatic SOA factor overlapped with ions identified in the chamber experiment. This signal fraction is considerably higher than for other SOA factors, i.e. biogenic SOA, aged biomass burning, and mixed SOA had signal contributions from chamber SOA ions of 16.0 %, 26.0 %, and 20.0 %, respectively.

Despite the strong dilution of the boundary layer during the day, there is little variation in the aromatic SOA concentration during the day. This points to a strong local daytime source, such that the production of aromatic SOA is fast enough to offset the strong dilution effects of the expanding boundary layer as discussed further in Sect. 3.5.

The mass spectrum of biogenic SOA is shown in Fig. 2a. It contains high contributions from ions such as C9H16O5, C8H14O6, and C9H14O4, which have been previously identified in EESI-TOF factors representing biogenic oxidation products (Qi et al., 2020; Stefenelli et al., 2019). Compounds with H:C >1.5 constitute nearly 50 % of the signal of this factor, with 2.9 %, 5.0 %, 6.8 %, and 2 % of signal resulting from C7H10-14O4-8, C8H12-16O4-8, C9H14-18O4-8, and C10H16Oz, respectively. These compounds have been previously attributed to monoterpene oxidation products in Zurich, Switzerland (Stefenelli et al., 2019). There are, however, some notable differences from the Zurich study. Specifically, the present study shows a much smaller contribution from the C10H16Ozcompounds (6.2 % in Zurich vs. 2 % in this study), while CxHyOzN1 compounds comprise 26.7% of the factor profile (vs. ∼13% in Zurich). These differences might arise because of two different reasons. One is the probable contribution of not only monoterpene oxidation products (which dominate in Zurich) but also isoprene oxidation products, e.g. C5H10Ox and C5H9NOx (Chen et al., 2020), to the biogenic factor retrieved in this study. This is because of Delhi's location near the tropics, which results in a large contribution of isoprene to the total biogenic VOCs. Model estimates predict isoprene emission fluxes to be a factor of ∼20 higher than α-pinene in tropical regions of India, whereas in Europe the emission fluxes of isoprene and monoterpenes are similar (Guenther et al., 2012; Sindelarova et al., 2014). Zurich and Delhi also differ in terms of atmospheric conditions, in particular the much higher NOx levels in Delhi as compared to Zurich, consistent with the higher CxHyOzN1 fraction. Biogenic VOCs such as monoterpenes and isoprene are susceptible to oxidation by NO3 radicals, which can result in large amounts of biogenic SOA production. In Delhi, however, due to large concentrations of NO (∼200–300 ppbv) during the nighttime, the production of NO3 radicals is suppressed and the diurnal cycle of NO3 is actually inverted with the majority of available NO3 radicals actually present during the daytime (Haslett et al., 2022).

The diurnal trend of this factor resembles that of the EESI-PMF primary factors in that the concentration is highest overnight and a strong decrease is seen during daytime hours. A possible explanation of this behaviour could be more regional sources of the biogenic VOCs scattered over a large area. This means that the biogenic SOA factor most likely has only a small daytime source in the vicinity of the site.

The fractional contribution of levoglucosan (C6H10O5) to the aged biomass burning factor mass spectrum is ∼10 %, which is a factor of 8 lower than for the primary biomass burning factor. The lower levoglucosan content in the aged biomass burning mass spectrum as compared to the primary biomass burning mass spectrum is consistent with observations of similar factors in Zurich during winter (Qi et al., 2019). Additionally, chamber studies have shown that the levoglucosan concentration decreases in aged biomass burning particles (Bertrand et al., 2018), while the concentrations of secondary species increase, consistent with observations in this study.

The mass spectrum of this factor also has ∼2 % contributions each from two key ions, C6H8O6 and C7H8O7. These are most likely oxidation products of phenols and methoxy-phenols, which are abundant secondary compounds formed during the ageing of biomass burning emissions (Yee et al., 2013) and are important precursors of biomass burning SOA. The aged biomass burning factor contains dominant signals from various other small molecules with H:C ratios less than 1.3, which is consistent with the oxidation of small aromatic compounds emitted during biomass burning such as phenolic compounds. These compounds were observed to be major contributors of gas-phase solid-fuel combustion factors in a recent VOC source apportionment study conducted at the same site (Wang et al., 2020).

The diurnal trend of aged biomass burning is similar to the one of primary factors, characterized by increased concentrations during evening hours and a decline during daytime hours. The amplitude of the evening time peak however differs between aged biomass burning and primary biomass burning (Fig. S1). While the primary biomass burning increases by a factor of ∼50–60 between 17:00–22:00 LT, the increase in oxidized biomass burning is within a factor of ∼5 during the same time. Between 22:00–07:00 LT, the concentration of this factor steadily decays by a factor of 2.5. A distinct peak is also observed between 07:00–09:00 LT, which coincides with an increase in solar radiation, indicating that ageing of emissions takes place during early morning hours. Note that it is likely that a majority of local emissions are not oxidized during nighttime in Delhi due to very high levels of NO (∼200–300 ppb), which may scavenge both O3 and NO3 radicals during nighttime and inhibit nocturnal degradation of VOCs (Haslett et al., 2022).

The diurnal pattern of the LO-OOA factor from AMS correlates well with the aged biomass burning factor from the EESI-TOF and suggests that oxidation of biomass burning emissions may be the dominant contributor to the LO-OOA factor observed in the AMS.

The remaining factor is most likely a mixed SOA factor, which has influences from both anthropogenic and biogenic sources. The highest-intensity ions in this factor are C5H10O4, C9H14O5, C6H10O5, C9H16O5, C9H18O5, and C6H12O4. C5H10O4 is probably a product of isoprene oxidation, whereas the dominance of C9 compounds suggests contributions from oxidation products of C9 species. The C9 species have varied sources in urban areas which include evaporative losses of fuels (e.g. gasoline), solvent use, and unburnt exhaust emissions (Mehra et al., 2020; Zhang et al., 2013), and hence this factor is likely influenced by different sources linked to the Delhi urban area. This factor also has ∼2.5 % of mass spectral signal contributed by levoglucosan (C6H10O5). Either this could be due to non-perfect unmixing by PMF, or it could indicate the contributions from biomass burning with other above-mentioned sources in this factor. This factor is therefore named mixed urban SOA.

The diurnal pattern of this factor shows roughly stable concentrations from 00:00–06:00 LT, which is similar to the aromatic SOA factor. All other factors show a decrease during these hours. It increases by a factor of ∼3 between 11:00–14:00 LT and then steadily decays before being enhanced again by a factor of ∼15 between 18:00–21:00 LT. The daytime increase suggests photochemical production of this factor. The diurnal trend of this factor is stable during late night hours and does not show a marked early morning rush hour peak, indicating little or no influence of morning rush hour emissions.