New insights into PM 2 . 5 chemical composition and sources in two major cities in China during extreme haze events using aerosol mass spectrometry

During winter 2013–2014 aerosol mass spectrometer (AMS) measurements were conducted for the first time with a novel PM2.5 (particulate matter with aerodynamic diameter ≤ 2.5 μm) lens in two major cities of China: Xi’an and Beijing. We denote the periods with visibility below 2 km as extreme haze and refer to the rest as reference periods. During the measurements in Xi’an an extreme haze covered the city for about a week and the total non-refractory (NR)-PM2.5 mass fraction reached peak concentrations of over 1000 μg m. During the measurements in Beijing two extreme haze events occurred, but the temporal extent and the total concentrations reached during these events were lower than in Xi’an. Average PM2.5 concentrations of 537± 146 and 243± 47 μg m (including NR species and equivalent black carbon, eBC) were recorded during the extreme haze events in Xi’an and Beijing, respectively. During the reference periods the measured average concentrations were 140± 99 μg m in Xi’an and 75± 61 μg m in Beijing. The relative composition of the NR-PM2.5 evolved substantially during the extreme haze periods, with increased contributions of the inorganic components (mostly sulfate and nitrate). Our results suggest that the high relative humidity present during the extreme haze events had a strong effect on the increase of sulfate mass (via aqueous phase oxidation of sulfur dioxide). Another relevant characteristic of the extreme haze is the size of the measured particles. During the extreme haze events, the AMS showed much larger particles, with a volume weighted mode at about 800 to 1000 nm, in contrast to about 400 nm during reference periods. These large particle sizes made the use of the PM2.5 inlet crucial, especially during the severe haze events, where 39± 5 % of the mass would have been lost in the conventional PM1 (particulate matter with aerodynamic diameter ≤ 1 μm) inlet. A novel positive matrix factorization procedure was developed to apportion the sources of organic aerosols (OA) based on their mass spectra using the multilinear engine (ME-2) controlled via the source finder (SoFi). The procedure allows for an effective exploration of the solution space, a more objective selection of the best solution and an estimation of the rotational uncertainties. Our results clearly show an increase of the oxygenated organic aerosol (OOA) mass during extreme haze events. The contribution of OOA to the Published by Copernicus Publications on behalf of the European Geosciences Union. 3208 M. Elser et al.: OA chemical composition and sources during haze events in China total OA increased from the reference to the extreme haze periods from 16.2± 1.1 to 31.3± 1.5 % in Xi’an and from 15.7± 0.7 to 25.0± 1.2 % in Beijing. By contrast, during the reference periods the total OA mass was dominated by domestic emissions of primary aerosols from biomass burning in Xi’an (42.2± 1.5 % of OA) and coal combustion in Beijing (55.2± 1.6 % of OA). These two sources are also mostly responsible for extremely high polycyclic aromatic hydrocarbon (PAH) concentrations measured with the AMS (campaign average of 2.1± 2.0 μg m and frequent peak concentrations above 10 μg m). To the best of our knowledge, this is the first data set where the simultaneous extraction of these two primary sources could be achieved in China by conducting on-line AMS measurements at two areas with contrasted emission patterns.

During the measurements in Xi'an an extreme haze covered the city for about a week and the total non-refractory (NR)-PM 2.5 mass fraction reached peak concentrations of over 1000 µg m −3 . During the measurements in Beijing two extreme haze events occurred, but the temporal extent and the total concentrations reached during these events were lower than in Xi'an. Average PM 2.5 concentrations of 537 ± 146 µg m −3 10 and 243 ± 47 µg m −3 (including NR species and equivalent black carbon, eBC) were recorded during the extreme haze events in Xi'an and Beijing, respectively. During the reference periods the measured average concentrations were 140 ± 99 µg m −3 in Xi'an and 75 ± 61 µg m −3 in Beijing. The relative composition of the NR-PM 2.5 evolved substantially during the extreme haze periods, with increased contributions of the inorganic 15 components (mostly sulfate and nitrate). Our results suggest that the high relative humidity present during the extreme haze events had a strong effect on the increase of sulfate mass (via aqueous phase oxidation of sulfur dioxide). Another relevant characteristic of the extreme haze is the size of the measured particles. During the extreme haze events, the AMS showed much larger particles, with a volume weighted mode 20 at about 800 to 1000 nm, in contrast to about 400 nm during reference periods. These large particle sizes made the use of the PM 2.5 inlet crucial, especially during the severe haze events, where 39±5 % of the mass would have been lost in the conventional PM 1 (particulate matter with aerodynamic diameter ≤ 1 µm) inlet. A novel positive matrix factorization procedure was developed to apportion the sources of organic aerosols Introduction  (Xu et al., 2006;Zhang and Tao, 2009;Huang et al., 2014;Wei et al., 2015). In China, severe pollution events often occur during wintertime, when stagnant meteorological conditions confine the gas-and particle-phase pollutants at the ground level. The particles can either be directly emitted as primary aerosols (e.g. particles emitted 5 from combustion sources) or formed in the atmosphere by condensation of oxidation products of sulfur dioxide, nitrogen oxides and volatile organic compounds (secondary aerosol).
The first step for developing air pollution control strategies requires the identification of the major sources and processes producing airborne particles. Most previous 10 aerosol studies in the areas of Xi'an and Beijing, two major Chinese cities, are based on filter measurements (Cao et al., 2012;Wang et al., 2013;P. Wang et al., 2015;Huang et al., 2014;Ho et al., 2015;M. Gao et al., 2015;Xu et al., 2015;Yang et al., 2015). Carbonaceous materials, water-soluble ions (e.g. sulfate, SO 2− 4 , nitrate, NO − 3 , and ammonium, NH + 4 ) and mineral dust have been found to be major constituents 15 of fine particles in both cities during wintertime. During haze days, elevated concentrations of secondary ion species contribute considerably to the decrease in visibility (J. J. Gao et al., 2015;Zhang et al., 2015a). High relative humidity resulting in enhanced water uptake by the hygroscopic aerosol particles and formation of secondary aerosol have been suggested as an important factor during haze events in China (Sun cooking emissions. Sun et al. (2013b) found coal combustion particles to dominate the organic aerosol (OA) in Beijing in wintertime (on average 33 % of the OA) and enhanced contribution of this factor during polluted periods. Lower contributions of coal combustion aerosol were found in measurements performed in January 2013 (Zhang et al., 2014;Sun et al., 2014), with coal combustion explaining 15 and 19 % of the 10 total OA, respectively. Among all three studies, the average contribution of traffic to the OA varied between 11 and 18 %, while cooking emissions explained between 12 and 20 % of the OA. However, all these studies failed to resolve a factor related to biomass burning, which is known to be a major particle source in winter. In addition, each study reported two to three oxygenated OA (OOA) components resulting from 15 secondary processes. Secondary organic aerosol (SOA) was found to dominate the OA mass concentrations in January 2013 (54 % of OA in Zhang et al., 2014 and55 % in Sun et al., 2014), with increased relative contribution during more polluted days. Similar real-time measurements in other Chinese cities, including Xi'an, are scarce, preventing an accurate assessment of the spatial variation of the aerosol composition 20 and sources in China during haze events. Despite the widespread use of PM 2.5 as an air quality standard, previous online aerosol mass spectrometry measurements have only been able to measure the submicron fraction. In this work we present the first online high-resolution time-of-flight aerosol mass spectrometer (HR-ToF-AMS) measurements of the non-refractory (NR)-Introduction Xi'an, with over 8 million inhabitants in 2013, is the largest city in western China. Besides the local anthropogenic emissions, the region is often affected by the transport of dust particles from the Gobi desert and by the accumulation of pollutants when 10 stagnant air masses are confined in the Guanzhoung basin. The sampling site was located within the High-Tech area south-west from the urban core, surrounded by many office buildings, some factories and construction sites. Nearby streets were sporadically watered during high pollution periods to minimize road dust resuspension.
Beijing, the capital of China, with over 20 million inhabitants in 2013 is one of the Introduction  Weingartner et al. (2003). A PM 2.5 cyclone was located in front of the main inlet of the Aethalometers. The particles were transmitted from the cyclone to the Aethalometer through ∼ 3 m of copper tube (12 mm outer diameter) at a flowrate of ∼ 4 L min −1 . Source apportionment was performed on the organic AMS data using PMF as implemented by the multilinear engine (ME-2; Paatero, 1997) and controlled via the interface SoFi coded in Igor Wavemetrics (Source Finder;Canonaco et al., 2013). PMF is a bilinear unmixing receptor model which enables describing the variability 10 of a multivariate database as the linear combination of static factor profiles and their corresponding time series. This is achieved by solving Eq. (1), where X is the measurement matrix (consisting of i rows and j columns), G contains the factor time series, F the factor profiles and E the model residuals. The model uses a least squares approach to iteratively minimize the object function Q (Eq. 2), defined as the sum of the squared 15 residuals (e i j ) weighted by their respective uncertainties (σ i j ).

Source apportionment techniques
In our case, the model input consists of a data and error matrix of OA mass spectra, where the rows represent the time series and the columns contain the ions fitted in 20 high resolution (HR) for the V mode data. Considering only the mass from the HR fits (up to m/z 115), 10 ± 8 % of the OA mass was excluded. The initial error values were calculated by the HR-AMS data analysis software previously described (PIKA) and a minimum error corresponding to the measurement of a single ion was enforced 30136 Introduction  (Ulbrich et al., 2009). Further, as suggested by Paatero and Hopke (2003), variables with signal-to-noise ratio (SNR) lower than 0.2 were removed and variables with SNR between 0.2 and 2 were down-weighted by increasing their error by a factor of 3. Finally, all variables directly related to m/z 44 in the organic fragmentation table (i.e. m/z's 16, 17, 18 and 28) (Allan et al., 2004) were excluded for the PMF analysis to appropriately 5 weight the variability of m/z 44 in the algorithm and were reinserted post-analysis. After the aforementioned corrections were applied, the final input matrix contained 270 ions and 50909 points in time (with steps of 60 s). PMF was solved using the multi-linear engine (ME-2, Paatero, 1999), which in contrast to unconstrained PMF analyses enables complete exploration of the rotational ambiguity (i.e. different combinations of the matrices G and F can give solutions with the same mathematical quality) of the solution space. For computational efficiency, in this study this was achieved by directing the solution towards environmentally meaningful rotations using the a value approach. This method constrains one or more output factor profiles to fall within a predetermined range governed by the combination of an 15 input profile and a range-defining scalar (a value). For example, in the case in which a factor profile (f j ) is constrained with a certain a value (a), the following condition needs to be fulfilled: The number of factors in PMF is determined by the user and the solutions of the model 20 are not mathematically unique, due to rotational ambiguity. Therefore, it is very important to use criteria such as chemical fingerprint of the factor profiles, diurnal cycles and correlations between the time series of factors and external measurements to support factor identification and interpretation (Ulbrich et al., 2009;Canonaco et al., 2013). 25 As mentioned in Sect. 53.8 ± 1.3 % relative contribution to OA mass), but a very low mass fraction from this source in Xi'an (10.5 ± 0.4 µg m −3 on average and only 9.2 ± 0.3 % relative contribution to OA mass). Using these results, we estimated the ratio eBC/CCOA to be 0.037 ± 0.006 in Beijing. This was accomplished by fitting eBC as a linear combination of the three identified combustion sources: traffic (hydrocarbon-like OA, HOA), 20 biomass burning (biomass burning OA, BBOA) and coal combustion (CCOA). Although the eBC measurements in Beijing were conducted at 2.8 km south from our sampling site, the reconstruction of the eBC concentration based on OA primary fractions from ME-2 shows a very good agreement with the measured eBC (see Fig. S2) and the obtained eBC/CCOA ratio is in good agreement with previous values reported in litera- 25 ture (Zhang et al., 2008). Using the ratio eBC/CCOA obtained for Beijing, we estimate that coal combustion contributed on average only 2.2 ± 1.4 % to the measured eBC in Xi'an. Moreover, also for Xi'an the reconstruction of eBC by means of the combustion OA sources is very successful (as shown in Fig. S2). Therefore, we conclude that the  Sandradewi et al. (2008) to separate eBC wb and eBC tr can be reasonably applied to our data from Xi'an, but not in the case of Beijing. For the eBC source apportionment in Xi'an, Angstrom exponents of 0.9 and 1.7 were considered for traffic and wood burning, respectively, following the suggestions in Zotter et al. (2015) presenting a re-evaluation of the method developed in Sandradewi 5 et al. (2008). The eBC wb to BBOA ratio was found to be 0.14, which is in good agreement with previous reported values ( Gilardoni et al., 2011;Zotter et al., 2014). The ratio eBC tr to HOA was 0.79, which is lower than the ratios reported in previous European studies (El Haddad et al., 2013 and references therein) but is in good agreement with results derived from measurements in China (Huang et al., 2012;Zhou et al., 2014).

10
This difference in the eBC tr to HOA ratio at the two locations is most probably related to the higher percentage of gasoline vehicles in China compared to Europe.

PAH quantification
PAH concentrations were quantitatively determined from the high resolution AMS data. All details about the method used can be found in Bruns et al. (2015)

Interpretation and optimization of OA source apportionment
A key consideration for PMF analysis is the number of factors selected by the user. As currently no methodical and completely objective approach exists for choosing the 5 right number of factors, this selection is generally evaluated through comparisons of the time series of the factor and external tracers as well as the analyses of factor mass spectra and diurnal patterns. In this work we present a detailed source apportionment that has been optimized to minimize the user subjectivity on the solution and better estimate the uncertainties of the final solution.

10
In a first step, we examined a range of solutions with two to eight factors from unconstrained runs (see Fig. S3). The solution that best represented the dataset is the fivefactor solution, which yields factors interpreted as hydrocarbon-like OA (HOA), biomass burning OA (BBOA), coal combustion OA (CCOA), cooking OA (COA) and oxygenated OA (OOA). The HOA profile is distinguished by the presence of alkyl fragment signa- , which are known fragments from anhydrous sugars present in biomass smoke (Alfarra et al., 2007). The key feature of the CCOA is the presence of unsaturated hydrocarbons, with higher explained vari-20 ability of these unsaturated fragments at higher m/z. The COA profile is very similar to the HOA spectra but has higher contributions of the oxygenated ions at m/z 55 (C 3 H 3 O + ) and m/z 57 (C 3 H 5 O + ). Finally, the OOA profile is characterized by a very high m/z 44 (CO + 2 ). COA is not resolved in solutions with a lower number of factors. Meanwhile, when a six-factor solution is considered, OOA splits into two factors with 25 very similar profiles and whose time series reflect the change in the instrument tuning (Fig. S3). Further increasing the number of factors does not improve the interpretation ACPD 15,2015 New insights into PM 2.5 chemical composition and sources of the data, as the new factor time series and spectral profiles are highly correlated with those extracted from lower order solutions and cannot be explicitly associated to distinct sources or processes. Although the unconstrained five-factor solution appears to be a reasonable representation of the data, the mass spectral profiles indicate mixing between the sources.

5
This is specifically the case between HOA and BBOA, where the HOA profile contains a higher than expected contribution of C 2 H 4 O + 2 (m/z 60), and between COA and OOA, with a rather high contribution of CO + 2 (m/z 44) in the COA profile. Precisely, in the unconstrained solution the m/z 60 in HOA is 0.009 ± 0.001 % (standard deviation from average over 10 seed runs), compared to 0.002±0.002 % obtained from the average of 10 multiple ambient datasets (Ng et al., 2011). Likewise, the m/z 44 in the unconstrained COA profile is 0.069 ± 0.001 %, compared to 0.013 ± 0.004 % obtained as an average of previously reported COA spectra (He et al., 2010;Crippa et al., 2013;Wolf, 2014).
To decrease the influence of BBOA on the apportionment of HOA, we constrained HOA using the profile from Crippa et al. (2013), which is characterized by a minor con-15 tribution of m/z 60. Note that while other approaches were explored throughout the entire analysis, including the use of other HOA profiles or increase of the factor number, the BBOA-HOA separation couldn't be significantly improved. Although constraining the HOA improves the HOA-BBOA separation, it compromises the apportionment of cooking emissions, with a higher background mass and unexpectedly high concen-20 tration overnight in the diurnal trend. To avoid the mixing of COA with other sources, the COA profile of Crippa et al. (2013) was constrained. In the following we discuss the sensitivity of the results to the a values used to constrain the HOA and COA factor profiles.
Considering a values between 0 and 1 with a step of 0.1 for both HOA and COA 25 yields 121 possible combinations of a values. A set of three criteria was established to assess the solutions that represent environmentally better the OA fractions.
(1) Minimization of m/z 60 in HOA. A threshold for the maximal fractional contribution of m/z 60 in HOA was set to 0.006 based on profiles derived from multiple ambient ACPD 15,2015 New insights into PM 2.5 chemical composition and sources  , 2011). The fractional contribution of m/z 60 to the normalized HOA profiles varied between 0.0016 and 0.0092 % over the full a value space. This criterion eliminated all solutions with an a value for HOA of 1, as shown in Fig. S4.
(2) Optimization of COA diurnals. Unambiguous chemical markers for cooking emis-5 sions are not yet clearly established, hindering their use for the optimization of the apportionment of this source. A valuable characteristic for the identification of COA is the analysis of its diurnal trends: near the emissions source (e.g. in an urban area) COA typically has a distinctive diurnal with maxima at lunch and dinner times. In order to categorize the solutions, a novel approach using cluster analysis was utilized.
The normalized COA diurnals of all studied a value combinations were grouped using k-means cluster analysis. This technique aims at grouping the observations into k clusters, by minimizing the first term (T1) from the cost function (CF) shown in Eq. (4). This term represents the sum of the Euclidian distances between each observation (x i ) and its respective cluster center (µ zi ). The results from the cluster analysis are 15 shown in Fig. 1, for two-, three-, and four-cluster solutions. For each solution, the first panel shows all diurnals pertaining to the different clusters, the second plot shows the diurnal pattern of the cluster center and the third plot shows the clusters' attribution in the a value space. An issue encountered in cluster analysis is the determination of the number of clusters (k) that best describes the data. Increasing k decreases T1, while 20 adding complexity to the solution. A common approach to select the optimal number of clusters is to explicitly penalize the higher order solutions for complexity by using the Bayesian information criteria (BIC). This penalty is introduced with the second term (T2) in Eq. (4), given by the product of the number of clusters (k) and the logarithm of the dimensionality of the cluster (D= 24 h in our case):  Fig. 1, the diurnals of the purple cluster exhibit a higher background concentration over the full day, which are difficult to reconcile with the expected COA emission trends. The red and blue clusters have both lower background values; however the blue cluster has some peaks over the night hours that aren't expected from COA emissions. Moreover, the solutions in the red cluster are more similar to previous reported COA 5 spectra (He et al., 2010;Crippa et al., 2013;Wolf, 2014), as they have a lower contribution of m/z 44 compared to the solutions in the other two clusters (see Fig. S5). Specifically, the average relative contribution of m/z 44 in the COA spectra from literature previously mentioned is 0.013 ± 0.004 %, which is in good agreement with the relative contribution of 0.013 ± 0.002 % found for the red cluster. As the spectrum for 10 the blue and purple clusters have higher contributions of m/z 44 (0.026 ± 0.008 % and 0.025 ± 0.019 %, respectively), only the solutions belonging to the red cluster are considered as good solutions. A disadvantage of the k-means algorithm is that the solution space might have several local minima and therefore the result could strongly depend on the initialization. Hence, 100 random initializations of the algorithm were 15 performed and only the a value combinations that fell into the red cluster more than 95 % of the time were retained as good solutions. Combining these results with the criterion previously applied on the HOA profile, we obtained the range of accepted a values combinations shown in Fig. 2b.
(3) Factor-tracer correlation. The following correlations between the identified pri-20 mary sources and the external tracers were considered: In all cases, low concentration points (below the 5th percentile, P05) were discarded. 25 Note that the separation between eBC tr and eBC wb was only possible with the data 30143 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | collected in Xi'an (see Sect. 2.3.2). Moreover, as the eBC wb does not follow the BBOA time series during the haze event (see discussion in Sect. 4.2), only data from the reference period was considered for this analysis. The linear relation between PAH and BBOA, CCOA and HOA will be discussed in detail in the source apportionment result section (Sect. 4.3).

5
For each of these parameters, the normalized difference, S, between the measured and calculated marker concentrations was retrieved for all accepted a value combinations using Eq. (8). The standard deviations of S, which are considered as an estimate of the variability between the factor and its corresponding marker, were combined in quadrature as shown in Eq. (9): where σ ALL is the object function that needs to be minimized for the optimization of the selected solutions and is represented with a color scale in Fig. 2c. The standard deviations of the individual parameters (σ PAH , σ eBC tr and σ eBC wb ) within the accepted a value 15 space are shown in the Supplement (Fig. S6). As seen from Fig. 2c, the solution obtained using an a value of 0.9 for the HOA profile and 0.6 for the COA profile, has the minimum σ ALL (σ min = 0.94). In order to establish the stability of the solution at a certain a value combination with respect to the measurement uncertainty, we examined the variability of σ min for the best solution, by reinitializing 50 times the ME-2 algorithm 20 with different input matrices. For each repetition, the elements of the OA input matrix were varied within twice their uncertainties (OA(i , j ) ± 2OA error (i , j )). All of the 50 solutions satisfied the two criteria previously described (i.e. minimization of m/z 60 in HOA and optimization of COA diurnal) and σ ALL presented 7.5 % variability among the 50 iterations. Considering all solutions inside the 95 % confidence interval (i.e. twice its variability, σ ALL < σ min + 15.0 %) to represent the data equally well compared to the best solution, all a value combinations within the red region in Fig. 2c  All results presented hereon are averaged over all these possible a values combinations, and their standard deviation is considered as our best estimation of ME-2 errors. Note that these errors are very likely lower estimates of the model uncertainties, as the solution space could not be fully explored. The error bars in Fig. 3 represent the variability of each m/z fraction (standard deviation) across all good solutions in the 5 a value space. As this retained range of solution is a direct consequence of our input error estimate, we assessed the sensitivity of our results to the input errors by running the algorithm by varying the OA input matrix within smaller limits (OA ±1OA error ). This lead to similar results as the method described above, with the only difference that two additional a value combinations (marked with the dashed line in the left corner of Compared to the unconstrained PMF (average over 10 seeds), the optimized solution (average over all good a value combinations) has more genuine factor profiles ( Fig. 3), with decreased contributions of m/z 60 in the HOA spectra (from 0.009 ± 0.001 % to 0.003 ± 0.001 %) and of m/z 44 in the COA spectra (from 0.069 ± 0.001 % to 15 0.013 ± 0.002 %). Moreover, σ ALL decreases considerably from 3.3 ± 0.1 in the unconstrained solution to 1.0±0.1 in the optimized solution. In terms of the model mathematical performance, there is only a moderate increase in the residuals in the optimized solution compared to the unconstrained run. Specifically, Q normalized by its expected value (Q/Q exp) (Paatero and Hopke, 2009) increases from 7.5 ± 0.1 in the uncon-20 strained solution to 8.5 ± 0.4 in the optimized solution. The correlations between the OA factors from the optimized solution and its corresponding tracers are presented in Fig. S7 and the correlation parameters (R 2 and slope) are reported in Table S1. These analyses were conducted separately for the four periods of interest and very good correlations are found in most of the cases. background relate to extreme haze events, which are defined by a visibility below 2 km (Zhang et al., 2015b). We recognize that the reduction of the visibility is partially due to the increase of the aerosol water content as a result of the increase in the RH. However, during the extreme haze periods a significant increase in the total aerosol burden is observed, with total PM 2.5 mass reaching peak concentrations above 1000 µg m −3 in 10 Xi'an. Regarding the chemical composition, Fig. 4a shows an increase in the inorganic aerosol species during the extreme haze periods, while organic aerosols dominate the particle mass in the reference periods (i.e. visibility above 2 km). In the top panel of Fig. 4a, the ratio PM 1 to PM 2.5 mass is reported. This ratio was obtained from the integration of the collected PToF data. The size distributions of each species were in-15 tegrated over the full measured size range (up to 6000 nm) to determine the total mass measured with the PM 2.5 lens and until 800 nm as an estimation of the mass that the commonly used PM 1 lens would have detected (the 50 % cut-off diameter of the PM 1 inlet is at about 800 nm vacuum aerodynamic diameter, Liu et al., 2007). As mentioned in Sect. 2.2.1, the actual upper cut-off of the PM 2.5 inlet has been determined to be 20 above 2.5 µm (Williams et al., 2013). This comparison between PM fractions might suffer from the slow evaporation of some particles in PToF mode, which would lead to a higher apparent d va and a calculated higher than true mass loss in the PM 1 lens. On the other hand, it is possible that super-micron particles are more prone to particle bounce . Despite the uncertainties related to this calculation, the im-  As different emission sources can be present in the two measurement locations and some characteristics of the aerosols are expected to be distinct during the extreme haze periods, results are presented below for four different time frames: (1) Xi'an extreme haze (17 December to 26 December 2013), (2) Xi'an reference (13 December 2013 to 6 January 2014, excluding extreme haze), (3) Beijing extreme haze (15 5 January to 17 January 2014, with a small gap of some hours) and (4) Beijing reference (9 January to 26 January 2014, excluding extreme haze).
The median diurnal trends of the AMS species and eBC are shown in the top panel of Fig. 5 (see the 25th and 75th percentiles in Fig. S8). The extreme haze events in Beijing occurred twice over night and therefore the diurnals are incomplete and hard to 10 interpret. The diurnal trends are rather flat during the extreme haze in Xi'an, and exhibit more variation (with maximum concentrations at night) for the reference periods in Xi'an and Beijing. This variation is strongly influenced by the evolution of the planetary boundary layer height (which governs the vertical dilution of pollutants) and by the diurnal cycle of the emissions. During the reference periods, the increased solar radiation 15 induces the development of the mixing layer during daytime, and therefore the dilution of the pollutants. At night, the pollutant concentrations increase as a result of additional emissions in an increasingly shallower planetary boundary layer. During extreme haze periods, less solar radiation reaches the Earth's surface (see Fig. S9) and therefore dilution is reduced and particle concentrations remain elevated throughout the day.

20
Another important characteristic of the extreme haze events is the size of the measured particles. On average larger particles were detected during the extreme haze periods (size distribution modes at about 800 nm in Xi'an and between 800 and 1000 nm in Beijing) than during the reference periods (distribution modes at about 400 nm for both cities, Fig. 5). 25 As mentioned above, the mass of all aerosol components increased considerably during the extreme haze periods. The measured mean concentrations (and standard deviations as errors) were 537 ± 146 and 243 ± 47 µg m −3 during the extreme haze periods in Xi'an and Beijing, and 140 ± 99 and 75 ± 61 µg m −3 for the reference periods Introduction in Xi'an and Beijing, respectively. As shown by the relative contributions in the bottom panel of Fig. 5, the increase in mass during the extreme haze events is particularly high for the inorganic species (NO 3 , SO 4 , NH 4 and Cl) and therefore, the resulting ratio of inorganic (inorg) to organic (org) species is much higher during the extreme haze periods. Specifically, inorg/org ratios of 1.3 and 1.4 were obtained for the extreme haze conditions in Xi'an and Beijing, while the ratio dropped to 0.6 and 0.8 for the reference periods in Xi'an and Beijing, respectively. The mass concentration and relative contribution of eBC is higher in Xi'an than in Beijing, probably due to a higher contribution of older cars in Xi'an.

10
In this section the final results of the OA source apportionment are presented. All results are averages of all a value combinations that were accepted in the methodology described in Sect. 3. The absolute concentrations and relative contributions of the OA sources over time are shown in Fig. 6a together with the time series of external tracers. The absolute concentrations of the sources have rather small variability across all good 15 solutions (see Fig. S10). The mean relative contributions of the identified OA factors are shown in Fig. 6b for the four periods of interest. Lastly, the daily patterns of the absolute concentrations of the identified OA factors and the external tracers are reported in Fig. 7. Similar to the inorganic species and total OA, the diurnals of the OA factors are partially driven by the PBL dynamics, with increased dilution during daytime and 20 accumulation of the particulate mass overnight. Nevertheless, some factor-dependent differences are evident. The 25th and 75th percentiles of these diurnals and the standard deviation among all considered a value combinations are reported in Fig. S11. Potential Source Contribution Function (PSCF) analysis was performed to explore the geographical origin of the air masses during the measurements and to identify source 25 regions and other transport-related pollution events (see Fig. S12). OOA. A pronounced increase in the OOA mass concentration is observed during the extreme haze periods (blue background in Fig. 6a and afternoon. These increases are most probably related to regional production of OOA due to enhanced photochemical activity. These results are in agreement with the PSCF results, where shorter backward air mass trajectories during the extreme haze period in Xi'an indicate that regional emissions (within around 1000 kilometers) might play a dominant role during the extreme haze. OOA concentrations are higher 10 with northeast winds during the haze period in Xi'an (which might indicate a source region as there isn't a characteristic diurnal variation for the wind direction) while for the reference periods in Xi'an and Beijing the OOA shows rather homogeneous spatial distributions. COA. The COA average relative contribution to total OA is generally low for the 15 extreme haze periods (3.6 ± 0.5 % in Xi'an, 5.8 ± 1.0 % in Beijing) and around 10 % (9.3 ± 1.6 % in Xi'an, 11.5 ± 1.9 % in Beijing) for the reference periods. For all four periods, COA shows a very distinct diurnal trend with very strong peaks at lunch (between noon and 13:00 local time, LT) and dinner (19:00 to 20:00 LT) times. A small increase in the COA concentrations is also observed in the morning (06:00 to 07:00 LT), coin-20 ciding with breakfast time. The fragment ion C 6 H 10 O + at m/z 98 has been suggested among others as a marker ion for the COA factor (Sun et al., 2011;Crippa et al., 2013). Nevertheless, the correlation between these two components is very poor, mostly during the extreme haze period in Xi'an (R 2 = 0.21, see Table S1). This low correlation is mostly due to increased concentrations of C CCOA. Coal emissions are high in Beijing, dominating OA burden with contributions greater than 45 % of the OA mass (46.8 ± 1.2 and 55.2 ± 1.6 % for extreme haze and reference periods, respectively). In comparison, in Xi'an CCOA is of lower importance (5.7 ± 0.1 and 14.0 ± 0.6 % for extreme haze and reference periods, respectively). The CCOA mass slightly increases during the haze periods (more clearly seen in the case 5 of Xi'an), probably due to the accumulation of primary emissions during the stagnant conditions. CCOA concentrations decrease substantially during day time, due to dilution of the emissions in a deeper PBL. CCOA concentrations peak in the morning (at around 09:00 LT) and at night (starting to rise at 18:00 LT), probably due to domestic heating activities. Moreover, the CCOA is characterized by a strong peak in concentrations at around 03:00 to 04:00 LT, especially during the extreme haze period in Beijing. This peak, which is also present in the corresponding BBOA diurnal, might result from the late night burning emissions in a shallower boundary layer or from the advection of evening emissions from the surrounding areas. The PSCF results indicate that the high concentrations of CCOA (and BBOA) measured at the sampling site in Beijing might 15 be related to air masses coming from southwest of the sampling site (from the Hebei region).
BBOA. Unlike CCOA, BBOA is much more important in Xi'an, comprising about 40 % of the OA mass in the two considered periods (43.4±1.1 % and 42.2±1.5 % for extreme haze and reference periods, respectively). In Beijing instead, BBOA represents less 20 than 15 % of the total OA (13.8 ± 0.8 % and 8.9 ± 0.3 % for extreme haze and reference periods, respectively). Accordingly, while combustion emissions from domestic heating and cooking predominate the organic aerosol mass at both locations, our results highlight the clear difference in the type of fuel used for burning, with a higher fraction of coal burned in Beijing vs. a higher fraction of biomass burned in Xi'an. BBOA primary 25 emissions appear to accumulate under the stagnant conditions during severe haze events. In particular, in the last days of extreme haze in Xi'an, very high concentrations of BBOA are perceived without a significant increase in eBC wb . Nonetheless, the temporal evolution of BBOA correlates with the ion C 2 H 4 O + 2 at m/z 60 (overall R 2 = 0.97), 30150 ACPD 15,2015 New insights into PM 2.5 chemical composition and sources confirming the assignment to BBOA. This specific episode might be related to special conditions with lower amounts of eBC wb (e.g. from smoldering conditions) or absorbing wood burning organic carbon (e.g. from smoldering conditions or aged emissions -which would result in an overestimation of eBC tr ). The characteristics in the diurnal trends of BBOA are similar to those found in CCOA. The dilution of the particles 5 in a deeper PBL during day time results in a decrease in the BBOA concentration at around 16:00 LT, while peaks related to residential heating appear in the morning (between 09:00 to 10:00 LT) and at night (starting to rise at 18:00 LT). As already mentioned there is a strong peak at around 03:00 to 04:00 LT in the BBOA signal, which is probably related to the late night biomass burning emissions in a shallower PBL. In Xi'an, the PSCF results show that high concentrations of BBOA (and also HOA and CCOA) are observed when the air parcels are transported to the sampling site from northwest, indicating a possible major local pollution area northwest of the sampling site. In Beijing BBOA seems to be transported together with CCOA from the southwestern province of Hebei. 15 HOA. Despite the larger vehicle fleet in Beijing, higher concentrations of HOA are noticeable in Xi'an, possibly owing to a higher contribution of older cars. Accordingly, HOA is the third contributing source in Xi'an, explaining about 15 % of the OA mass (16.0 ± 1.6 and 18.3 ± 1.9 % for extreme haze and reference periods, respectively). By contrast, in Beijing, HOA is the least important source together with COA, explaining 20 only around 8 % of the OA mass (8.6 ± 1.3 and 8.7 ± 1.2 % for extreme haze and reference periods, respectively). An increase in HOA levels can be noticed during the haze periods, related to the accumulation of primary emissions under stagnant conditions. The HOA diurnals show peaks during morning and evening rush-hours (07:00 to 08:00 and 20:00 LT, respectively), as is typically the case for traffic-related pollutants. 25 Additional peaks are observed in the HOA during the night hours (between 23:00 and 06:00 LT). These peaks might be related to truck activity, which is strongly enhanced during the night hours as in both cities truck activity is banned during the day. While during the extreme haze event in Xi'an the PSCF results indicate an HOA source re- gion northwest from the sampling site, homogeneous distributions of the HOA factor are found for the reference periods in both Xi'an and Beijing, indicating a homogeneous distribution of this source.

Effect of relative humidity on aerosol composition
As previously mentioned, periods identified as extreme haze in this study are charac-5 terized by high RH (see Fig. S9). We examine in Fig. 8 the impact of RH on aerosol concentration and composition following the approach proposed by Sun et al. (2013).
As we have identified different emission patters in the two cities and the RH was only few times above 60 % in Beijing, the analysis is only performed for the case of Xi'an. In Fig. 8a in CCOA and COA with higher RH. These effects can not therefore be unequivocally attributed to the change in RH. More importantly, although the OOA mass concentration increases from about 10 to 60 µg m −3 when RH varies from 50 to 90 %, when normalized to its potential precursors, OOA does not show significant variability with RH. This suggests that unlike sulfate, whose production is highly enhanced in the aqueous 5 phase at high RH, OOA production rates seem to be independent of RH. The strong increase of the normalized sulfate at high RH suggests that aqueous phase oxidation of SO 2 could be an important process during extreme haze events. To investigate the oxidation degree of sulfur at different RH, the sulfur oxidation ratio (F SO 4 , Sun et al., 2006) was calculated according to Eq. (10) (where n is the molar 10 concentration) and is reported in Fig. 9 as a function of RH (note that this plot contains the full campaign data).
As seen in Fig. 9, F SO 4 has a clear exponential trend with RH. At RH below 50 % F SO 4 is rather constant and low (about 0.045 on average), while for higher RH the oxidation 15 ratio rapidly increases reaching 0.62 on average for the last RH bin (90-100 %). This extremely high oxidation degree of sulfur under high RH is an indication that aqueous phase production of sulfate might play a very important role during extreme haze events in China, in good agreement with the results reported by Sun et al. (2013a) for wintertime in Beijing.

PAH sources
To identify all sources emitting PAHs, PMF was performed using the OA matrix as an input, with an additional column containing the total PAH mass concentration calculated from the AMS. PAH errors were calculated assuming a Poisson distribution and the goodness of the combination of the two datasets (OA and PAH) was evaluated examin-Introduction are distributed around zero. However a small increase in their weighted residuals (Fig. S13b) is observed over night. The average PAH attribution was 28.9 ± 0.4 % to BBOA, 57.0 ± 0.7 % to CCOA and 14.1 ± 0.4 % to HOA (errors denote the standard deviation from 10 seed runs). The same combined input matrix was afterwards tested in the ME-2 approach, with the HOA profile constrained with an a value of 0.9, the COA 5 profile constrained with an a value of 0.6, and the PAHs unconstrained in all factors. Also in this case the PAHs were attributed to these three combustion sources with similar results (28.6 ± 0.4 % to BBOA, 62.0 ± 0.1 % to CCOA and 9.4 ± 0.3 % to HOA, with errors being the standard deviation among 10 seed runs). Hence the measured PAHs in our dataset can be fully attributed to biomass burning, 10 coal burning and traffic emissions. Using a linear regression of BBOA, CCOA and HOA to fit the measured PAHs (see Eq. 5 in Sect. 3) very similar attributions of the mass are found (27.6 ± 0.7 % attributed to BBOA, 66.4 ± 0.4 % to CCOA and 6.0 ± 0.5 % to HOA). The result of this fit (averaged over all good a value combinations) is shown in Fig. 10a together with the total mass of the measured PAH. As it can be seen from this 15 time series, the linear regression can reconstruct the measured PAH very precisely (R 2 = 0.94) and peaks of over 10 µg m −3 of PAHs can be attributed to the combined biomass burning, coal combustion and traffic emissions. Figure 10b presents the relative contributions of the three combustion sources to the measured PAHs for the different measurement periods. During the extreme haze event 20 in Xi'an, 63.8 ± 1.1 % of PAH are attributed to biomass burning, 25.3 ± 0.4 % to coal combustion, and the rest (10.9 ± 0.9 %) to traffic emissions. For the reference period the contribution of coal increases to about 55.9 ± 0.9 %, the biomass burning influence decreases to around 36.4±1.4 % and the traffic remains a minor contributor, explaining about 7.7 ± 0.8 % of the PAHs mass. In Beijing, coal emissions completely dominate ing that coal burning emission is an asymmetric source of PAHs, compared to other combustion emissions.

Discussion and conclusions
This work presents a thorough analysis of extreme haze events (visibility below 2 km) which occurred in Xi'an and Beijing during winter 2013-2014. Online aerosol mass 5 spectrometer analyses provided a detailed characterization of the chemical composition and size distribution of the aerosol components during the different measurement periods.
The extreme haze events were produced by a combination of primary emissions of particulate matter, generation of secondary aerosol, and stagnant meteorological con-10 ditions which confined the pollutants in the basin. Under such conditions, the mass concentrations of all aerosol components strongly increased, with resulting average PM 2.5 mass concentrations of 537 ± 146 µg m −3 in Xi'an and 243 ± 47 µg m −3 in Beijing (in contrast to 140 ± 99 µg m −3 and 75 ± 61 µg m −3 average NR-PM 2.5 mass measured during the reference periods in Xi'an and Beijing, respectively). Among all aerosol com- 15 ponents, sulfate and nitrate show the strongest enhancements during the extreme haze periods. Moreover, source apportionment of the organic aerosol (OA) fraction shows that also the formation of oxygenated organic aerosols (OOA) is strongly enhanced during the haze events. The high relative humidity characteristic of the periods with extreme haze was shown to favor the heterogeneous oxidation of SO 2 on deliquesced 20 aerosols and can therefore drive the drastic increase in sulfate concentrations. In contrast, aqueous phase processing appears not to significantly affect the formation of OOA and the other inorganic species. Another distinct feature of the aerosols during extreme haze events is their larger size compared to particles during lower pollution periods (the distribution mode of all NR-25 aerosol compounds shifts from around 400 nm during the reference periods to about 800 to 1000 nm during extreme haze events in both cities). The growth of the particles is associated with high secondary aerosol fractions and condensation of semi-volatile compounds on preexisting particles. Given the large mean aerosol diameters found during the extreme haze periods, the use of a PM 2.5 inlet for the AMS was a crucial point for the meaningfulness of our results, as 39 ± 5 % of the mass would have been neglected if a standard PM 1 inlet had been deployed.