Variations of Chemical Composition and Source Apportionment of PM 2 . 5 during Winter Haze Episodes in Beijing

PM2.5 samples were collected in Beijing between February 24 and March 12 of 2014, and analyzed to examine chemical compositions and origins of the PM2.5 at pollution levels of clean (PM2.5 < 75 μg m), light-medium (75–150 μg m), heavy (150–250 μg m) and severe (> 250 μg m). The mean PM2.5 concentration was 137.7 ± 124.8 μg m during the observation period, accounting for 66% of PM10. As all aerosol species concentrations increased with the pollution level, the contributions of secondary inorganic aerosols (SIA) to PM2.5 continuously increased while the contributions of OC and EC decreased, indicating a substantial contribution from secondary formation to the elevation of PM2.5 pollution. The acidity of PM2.5, the ratio of anion microequivalent concentration to cation, increased from 0.96 to 1.08 as pollution levels increased. Using a PMF model, secondary inorganic aerosols, industrial emissions, soil dust, traffic emissions, and coal combustion and biomass burning were identified as contributors to the PM2.5, and on average accounted 46%, 20%, 10%, 6% and 18% of the PM2.5, respectively, in the observation period. Industrial emissions were the dominant PM2.5 source during the clean period (60%). Except for traffic emission, sources of PM2.5 at the light-medium level were consistent, accounting for 17%–29%. Secondary inorganic aerosols were the largest origin of PM2.5 at heavy and severe pollution levels, accounting for 40% and 78%, respectively. In addition, the 48 h transport distances of air masses decreased from 2000 km (clean) to 300 km (severe level) and the proportion of air masses from south pollution areas in the total air masses at each pollution level increased from 0% to 97%, indicating that the stability of near surface air and the northerly transport of pollutants from the south at local and regional scales played a the key role in the PM2.5 elevation.


INTRODUCTION
Due to rapid industrialization and urbanization in recent decades, China has become one of the most significant source regions of anthropogenic atmospheric pollutants (Guo et al., 2014;Sun et al., 2015a;Lin et al., 2016).Beijing, the capital of China, a megacity with approximately 22 million inhabitants (Beijing Statistical Yearbook, 2015), has frequently suffered from extreme haze, which is characterized by high fine particle (i.e., PM 2.5 ) loading (Sun et al., 2006;Li et al., 2015).These fine particles have adverse effects on human health, can scatter and absorb the incident light and lead to atmospheric opacity, and horizontal visibility degradation (Han et al., 2015;Jing et al., 2015;Tao et al., 2015;Wu et al., 2016).
Previous studies on haze pollution in Beijing have been carried out, and primarily focused on three aspects: (1) formation mechanism and evaluation of haze episodes combined with synoptic weather, regional transport and heterogeneous reactions (Liu et al., 2013;Guo et al., 2014;Sun et al., 2015a;Yang et al., 2015;Zheng et al., 2015;Ma et al., 2017); (2) chemical characteristics and source apportionment of haze particles with online measurement, filter sampling and model simulating (Sun et al., 2013b;Huang et al., 2014;Wang et al., 2014;Gao et al., 2015;Chen et al., 2016;Elser et al., 2016;Zhang et al., 2016;Zhao et al., 2016); and (3) aerosol optical properties, hygroscopic properties, size distribution of haze particles (Zhang et al., 2012;Jing et al., 2014;Tao et al., 2015;Wang et al., 2015;Zhang et al., 2015a;Fajardo et al., 2016;Liu et al., 2016;Wu et al., 2016Wu et al., , 2017)).These studies were based on observations in haze episodes and have shown that aerosol properties and aerosol sources in the pollution and even at different altitudes.Recent studies found that the chemical compositions varied with the development of haze pollution, that secondary inorganic ions became more abundant with increases in PM 2.5 (Liu et al., 2015;Zhang et al., 2015a;Liu et al., 2016;Wang et al., 2016;Ma et al., 2017).However, chemical characteristics and sources of PM 2.5 have not been well examined at the different levels of haze pollution.
In this study, we applied the chemical mass balance, the ISORROPIA II thermodynamic equilibrium model, the positive matrix factorization (PMF) and the trajectory clustering to both characterize chemical species and explore potential origins of PM 2.5 at different haze pollution levels, as well as provide aids for controlling emission and reducing particle pollution.

Sampling Site and Collection
Beijing is located in the northeast part of the North China Plain and surrounded by mountains in three directions except in the south.Due to this distinctive topography, Beijing has frequently suffered from air pollution and extreme haze during the past decades.To investigate the formation mechanisms and potential sources of winter haze in Beijing, in-situ observations were performed during the period between 24 February to 12 March 2014 at the tower of the Institute of Atmospheric Physics, Chinese Academy Sciences (39.97°N, 116.37°E, 49 m above sea level, Fig. 1), a typical urban site located in the north of Beijing (Wu et al., 2017).
A R&P Partisol ® Model 2025 dichotomous sequential PM air sampler (Thermo, USA) and a MiniVol TAS PM sampler (Airmetrics, USA) were set on the roof of a twostory experiment building, approximately 8 m above the ground.The R&P sampler, operating at a nominal flow rate of 16.7 L min -1 , was used to simultaneously collect the fine (PM 2.5 ) and coarse particles (PM 2.5-10 ) on two individual Teflon filters with 47 mm in the diameter (Whatman PTFE).The MiniVol sampler was used to collect the PM 2.5 on a 47 mm quartz filter (Whatman QM-A), and operated at a flow rate of 5 L min -1 .The PM samples were synchronously collected twice per day, once during the daytime (from 7:00 to 19:00 at the local time) and once during nighttime (from 19:00 to 7:00 of the next day).

Chemical Components Analysis
One half of each PTFE filter was subjected to extraction and digestion with concentrated 4 mL of 60% HNO 3 and 2 mL of 48% HF, and 33 target metals were detected through inductively coupled plasma mass spectrometry (ICP-MS; Elan 6100; Perkin Elmer, USA).The 33 elements analyzed in PM 2.5 and PM 2.5-10 were Na, K, Ca, Fe, Al, Mg, Zn, Ti, Pb, Mn, Ba, Cu, Sr, Cr, Ni, Rb, V, Sn, As, Sb, Zr, Nb, Ga, Se, Mo, Co, Cd, Tl, Ge, Cs, U, Y, and Ag.Their recoveries were mostly within 10% of the certified or reference values (Hsu et al., 2010).The second half of each PTFE filter was subjected to extraction with 20 mL of Milli-Q water (18.2Ω) for 1 h, and ionic species concentration were determined through ion chromatography (Dionex ICS-90 for cations and ICS-1500 for anions), which was equipped with a conductivity detector (ASRSeULTRA).The water-soluble ionic species were K + , Na + , Mg 2+ , Ca 2+ , NH 4 + , SO 4 2-, NO 3 -, and Cl -, were analyzed with a precision better than 5% (Hsu et al., 2014).More information can be found in Lin et al. (2016).
In addition, QM-A filters were used to analyze organic carbon (OC) and elemental carbon (EC) in PM 2.5 .A punch of 0.526 cm 2 disc punch from each quartz filter was heated in a stepwise manner by a DRI-2100A carbonaceous aerosol analyzer in a pure helium atmosphere at 140, 280, 480 and 580°C, which converted particulate carbon to fractions OC1, OC2, OC3, and OC4, respectively.The filter was then continuously heated in 2% O2/98% He atmosphere at 580°C (EC1), 740°C (EC2) and 840°C (EC3), until the particulate carbon was converted to CO 2 .After the CO 2 was catalyzed by MnO 2 , it was reduced to CH 4 and then directly measured.According to the IMPROVE protocol (Chow et al., 2007), OC and EC are defined as: OC= OC1 + OC2 + OC3 + OC4 + OP; EC = EC1 + EC2 + EC3 -OP. ( where OP is the optical pyrolyzed OC.Detailed descriptions can be found in Zhang et al. (2013).
To examine chemical compositions and origins of PM 2.5 , we adopted the chemical mass balance method to reconstruct PM 2.5 mass concentration by considering SIA (NH 4 + , SO 4 2and NO 3 -), organic matter (OM), EC, fine soil (FS), water content (Water), trace element oxide (TEO), Na salt, K salt and PO4.A more detailed formation is shown in Table 1.
Because the concentrations of TEO, PO4, Na salt and K salt were low, and the contributions to PM 2.5 were small, we combined TEO, PO4, Na salt and K salt together as "others" in this study.The thermodynamic equilibrium model ISORROPIA II (reserve mode) (Fountoukis and Nenes, 2007) was used to calculate water content.As expected, the reconstructed PM 2.5 mass concentration was roughly consistent with that obtained from gravimetric measurement (R 2 = 0.996, Fig. S2)

PMF Model
The positive matrix factorization (PMF) model developed by the US Environmental Protection Agency, was a widely used receptor model, because of its high efficiency and convenience without the use of pollution discharge conditions.PMF version 5.0 has a key feature that allows specific constraints to be added into the factor profiles or factor contributions to reduce reduce results' uncertainties.Thus, the PMF5.0 was employed to identify the sources of fine particles in this study.Eight water soluble ions (e.g., Na + , NH 4 + , K + , Mg 2+ , Ca 2+ , NO 3 -, SO 4 2-, and Cl -), nine target metal elements (e.g., Al, Fe, Ti, Mn, Cu, Zn, Sb, Pb, and Cr) as well as the PM 2.5 mass concentrations were used as the model input.In this study, we have tested five different sources in the PMF analysis.

Air Mass Backward Trajectories Cluster Analysis
The 48-h backward trajectories of air mass arriving at the sample site (39.97°N,116.37°E) were calculated using the HYSPLIT 4 model developed by the National Oceanic and Atmospheric Administration (NOAA).The 3-hourly archived meteorological data with a 1° × 1° latitude-longitude grid was used for backward trajectories calculation and were provided by the US National Centers for Environmental Prediction Global Data Assimilation System (GDAS).The arrival level was set at 500 m above the ground level (a.g.l.).The model was run one time at each hour during the sampling period.

Overview the Chemical Components in PM 2.5 and PM 10
The PM 2.5 mean mass concentration was 137.7 ± 124.8 µg m -3 during the whole observation period, accounting for 66% of PM 10 mass concentration.This level of PM 2.5 was nearly four times higher than the National Ambient Air Quality Standard (NAAQS-2012) in China for PM 2.5 (35 µg m -3 ).It was the frequently level observed in Beijing when haze often occurred.For example, PM 2.5 was approximately 126.8 µg m -3 on average during the period from 15 October to 2 November (before APEC Beijing) and 125.2 µg m -3 from November 13 to 30 (after the APEC) (Tang et al., 2015).The present level of PM 2.5 is comparable to the 2009-2010 level (135 µg m -3 ; Zhang et al., 2013).
The average concentrations of OC and EC were 29.1 ± 21.8 µg m -3 and 5.2 ± 4.6 µg m -3 , respectively, and accounted for 29.5 ± 13.7% and 4.7 ± 2% of PM 2.5 mass, respectively.OC and EC concentrations were both higher than values in 2008 (9.2 ± 3.3 µg m -3 for OC; 2.6 ± 1.3 µg m -3 for EC) on the campus of Peking University (Guo et al., 2013).This is due to the effective government mitigation measures enacted in 2008.Compared to south China, OC concentrations to the  (2001) north in Beijing, were higher than those in Chengdu between January and October 2011 (17 ± 8 µg m -3 ).This is usually attributed to the widespread heating and the lower rainfall in Beijing in winter compared with southern China.The 33 trace elements accounted for approximately 6.4% of the PM 2.5 mass.In general, 6 major elements accounted for 92% of the concentration sum for all elements and increased in abundance in the order of K > Fe > Ca > Na > Al > Mg.The six elements also accounted for 95% of the concentration sum for all elements in PM 2.5-10 , but increased in abundance in the order of Ca > Fe > Al > Mg > Na > K. Except for K, the elements of Ca, Fe, Al, Na and Mg were mainly present in the coarse fraction, and the ratios of their mass in coarse particles (PM 2.5-10 ) to that in the total mass (PM 10 ) were 86.3%, 70.9%, 77.8%, 50.6% and 77.2%, respectively.Due to the low EF value, Al and Mg were considered mainly from natural soil.Ca, Fe and Na in PM 2.5 were mixed source elements but those in PM 2.5-10 were mainly from natural crustal sources.K in PM 2.5 was mainly from biomass burning while K in PM 2.5-10 was substantially from natural soils (Table S1, Lin et al., 2016).Although the concentrations of other elements, such as Se, Sb, Cd, were low, their EF values were high, indicating that they primarily emitted by anthropogenic activities.

Chemical Properties of Aerosols at Different Pollution Levels
To gain a further insight into the chemical properties and the origins of PM 2.5 , the 33 samples were divided into four classes according to the pollution levels.Chemical components and source data were averaged within each pollution level.The four pollution levels were classified as follows: clean (with sampled PM 2.5 < 75 µg m -3 ), lightmedium pollution (with PM 2.5 from 75 to 150 µg m -3 ), heavy pollution (with PM 2.5 between 150 and 250 µg m -3 ) and severe pollution (with PM 2.5 > 250 µg m -3 ).These levels were defined by the "Technical Regulation on Ambient Air Quality Index (on trial)", which the Chinese Ministry of Environmental Protection promulgated in 2012.
Both PM 2.5 and PM 10 significantly increased with the elevated pollution levels (Fig. 2).At the same time, the percentage of PM 2.5 in PM 10 increased, from 52% to 74%.The ambient particles were dominated by fine particles, and decrease in atmospheric quality was mainly caused by the increase in fine particles.Along with an increase in PM 2.5 , SO 4 2-, NO 3 -and NH 4 + concentrations increased, and the secondary inorganic species in particles were produced much more and contributed a higher portion to PM 2.5 than in PM 2.5-10 .It was also found that the sulfur oxidation ratio (SOR) of SO 2 precursors to sulfate (SO 4 2- ) and the nitrate oxidation ratio (NOR) of NO 2 precursors to nitrate (NO 3 -) in PM 2.5 increased with an increase in haze pollution levels (Fig. S1).This was consistent with varying concentrations of SO 4 2-and NO 3 -.These results indicate the important role of secondary formation in the enhancement of haze pollution.Recent studies regarded high sulfate and nitrate concentrations as a result of enhanced heterogeneous formation in polluted environments with high relative humidity, because relatively high humidity can accelerate the conversion of SO 2 and NO x to salts (Sun et al., 2013b;Fan et al., 2015;Huang et al., 2016;Pan et al., 2016a;Fang et al., 2017;Ma et al., 2017).
In addition, NO 3 -concentrations were higher than SO 4 2concentrations with the increasing pollution level, because the nitrate oxidation ratios increased faster and were higher than sulfate oxidation ratios (Fig. S1).This observation may be ascribed to a higher NO x concentration in precursor gases than SO 2 (Fig. S1), likely suggesting the effects of SO 2 emission control over China in the recent years (Sun et al., 2015b).Previous studies also showed nitrates dominating the secondary inorganic species of aerosols (Sun et al., 2015b;Zheng et al., 2015;Wang et al., 2017).
With the increasing pollution level, Mg 2+ and Ca 2+ concentrations increased.However, their contribution to PM 2.5 decreased.Because Mg and Ca were largely originated from sources such as surface soils and construction dusts, their contributions decrease can be ascribed to stable weather conditions such as low wind speed and high relative humidity (Fig. S1), which made these elements difficult to emit into the air (Tian et al., 2014).In addition, the ratio of Ca 2+ concentration in PM 2.5 to PM 10 significantly decreased from 31% at the light-medium level to 28% at the heavy level and then to 10% at the severe level.Mg 2+ concentration ratio in PM 2.5 to PM 10 also showed a constitute decrease from 40% (light-medium level) to 32% (severe level).The results indicate that more Mg 2+ and Ca 2+ enriched in coarse particles with increasing pollution levels.Although Na + , K + and Cl -concentrations showed increasing trends by the pollution level, their relative contribution to PM were different.The relative contribution of K + and Cl -to PM and the relative contribution of Na + to PM 2.5-10 increased with increasing pollution levels.However, the relative contribution of Na + to PM 2.5 decreased.
The acidity of aerosol particles is a parameter that can describe not only the potential of hygroscopic growth, toxicity and heterogeneous reactions of the particles, but also the index identifying possible threats to both the public health and the ecological system (Sun et al., 2010;Gao et al., 2015;Tan et al., 2016).In general, the ratio of A (i.e., the anion micro-equivalent concentration) to C (i.e., the cation microequivalent concentration) is used to examine the ion balance of aerosol particles and, consequently, the acidity of aerosol particles (Gao et al., 2015;Tan et al., 2016).The anion and cation equivalents in this study were calculated as follow: Cation = NH 4 + /18 + Na + /23 + K + /39 + Ca 2+ /20 + Mg 2+ /12 Anion = NO 3 -/62 + SO 4 2-/48 + Cl -/35.5 (2) Particles were in a neutralization state when the A/C ratio was close to 1, and particles were in an acidic state when the A/C ratio was larger than 1.The higher the A/C ratio, the stronger the acidity of particles.
The average ratios of A/C at the clean, light-medium, heavy and severe pollution levels were 0.82, 1.01, 1.07 and 1.06, respectively, for PM 10 and 0.96, 1.08, 1.11, and 1.08, respectively, for PM 2.5 .The acidity of the aerosol particles increased with the pollution level.This was consistent with the results of a previous study in Beijing where PM 2.5 samplers has stronger acidity during haze days (A/C = 1.27) than that during clean days (A/C = 1) (Zhang et al., 2016).Moreover, the average A/C ratio for PM 2.5 at each pollution level was substantially higher than that in PM 10 , reflecting that the acidity of fine articles was stronger than that of coarse particles.The mean ratio of A/C for PM 2.5 at clean, light-medium pollution, heavy pollution and severe pollution levels were 14%, 7%, 4%, 2% higher, respectively, than those for PM 10 .The results indicate that the acidity of particles PM 10 was getting close to the acidity of PM 2.5 as the pollution increased.Particle acidity is significantly influenced by the secondary aerosols formation (Khlystov et al., 2005).The development of sulfate and nitrate formation from their gas precursors (e.g., SO 2 , NO x ) with the enhancement of pollution not only enhanced the pollution level, but also increased the acidity of the aerosol (Sun et al., 2006;Han et al., 2015;Ma et al., 2017).
In general, the concentrations of OC and EC increased with the increase of pollution, while their relative contribution to PM 2.5 decreased.The result was consistent with that in an recent study (Tao et al., 2017).In this study, the contributions of OC to PM 2.5 decreased from 37% (clean) to 19 % (severe level), and that of EC to PM 2.5 decreased from 6% to 4%.The concentration of trace elements and their EFs increased with the pollution level (Fig. 3).For example, the EF of Se, the largest one among the values of all measured elements in PM 2.5 , was 6422 at the clean level and 26892 at the severe level.In addition, EFs for most trace elements in Fig. 3. Average EF values of trace metals in PM 2.5 and PM 2.5-10 aerosols collected in Beijing at different pollution levels.PM 2.5 were higher and increased faster than in PM 2.5-10 (Table S1), indicating fine particles were more easily affected by anthropogenic emissions.

Source Apportionment
Five PM 2.5 sources were identified by the PMF analysis including (1) secondary inorganic aerosols, (2) industrial emissions, (3) soil dust, (4) traffic emissions and ( 5) coal combustion and biomass burning.Fig. 4 shows the resolved source profiles from the model.The contribution percentage from each source to the species is also presented (black dots in Fig. 4).The time series of modeled concentrations for identified main sources and chemical mass closure (CMC) of the PM 2.5 are shown in Fig. 5.As expected, the modeled concentration for identified main sources was roughly consistent with the CMC results (Y = 0.997X, R 2 = 0.988), suggesting that the PMF results are reasonable.
The source apportionment to the PM 2.5 concentration at different pollution levels and their relative contributions are shown in Fig. 6.The relative dominance of each identified main source differed at different pollution levels.
Secondary inorganic aerosols were typically characterized by the abundance of NH 4 + , SO 4 2-and NO 3 -.The production of these secondary inorganic aerosols contributed to 46% of PM 2.5 during this observation period.Five haze episodes with high relative humidity occurred during our sampling time (Fig. S4).Previous studies have found that aqueous processing closely associated with high humidity likely played an important role in the formation of secondary aerosols when the gas-phase photochemical production of secondary aerosols was less significant under low ozone conditions on haze days (Sun et al., 2013a).Nitrate and sulfate were formed from chemical conversions of the precursors NO x and SO 2 .These precursor gases could be from vehicle exhausts, coal burning and biomass combustion (Tao et al., 2014;Li et al., 2016a).It has been reported that 90% of NH 3 was from fossil fuel-based burning emissions during hazy days in urban Beijing (Pan et al., 2016b).
The modeled concentration due to the secondary formation of inorganic aerosols increased with the enhancement of pollution level (Fig. 6), which was consistent with the variations of NH 4 + , SO 4 2-and NO 3 -(Fig.2).The contribution of the secondary inorganic formation to PM 2.5 continuously increased from 2% (clean) to 29% (light-medium level), and subsequently increased to 40% (heavy level) and 78% (severe level).Thus, secondary inorganic aerosols production is the predominant process to PM 2.5 increases in the heavy and severe stages of haze pollution.This result is also  supported by previous studies, which show that the increase of PM 2.5 in later stages of haze occurrence are dominated by heterogeneous reactions on preexisting particles (Niu et al., 2012;Liu et al., 2016;Ma et al., 2017).This was the major formation pathway for secondary inorganic aerosols in polluted environment under humid conditions.
Industrial emissions were characterized by high Cr, OC and EC.This source has an average contribution of 20% to PM 2.5 during the whole observation period.Cr in ambient aerosols can be categorized as a steel-related metal (Tao et al., 2014;Chen et al., 2016), Waste incineration is also a contributor of Cr (Nriagu and Pacyna, 1988;Reimann and Caritat, 1998).Zhang et al. (2013) proposed that industrial pollution could be as a vital source of carbonaceous aerosols, which was widely ignored.Although modeled concentrations from industrial sources showed little fluctuation with the development of haze pollution (Fig. 6), the contribution of this source to PM 2.5 mass concentration decreased with an increase in pollution.In winter, industrial emissions were the largest contributor at the clean level, accounting for 60% of PM 2.5 .As the haze pollution developed, the contributions of industrial emissions to PM 2.5 decreased from 25% (lightmedium level) to 13% (heavy level) and 11% (severe level).Soil dust was typically characterized by crustal elements, such as Al, Fe and Ti.Particles from this source accounted for 10% of the PM 2.5 on average.This result is similar to previous studies in Beijing (15%) (e.g., Zhang et al., 2013).Al is a typical indicator of crustal dust.Sun et al. (2005) found that Al was correlated closely with construction dust and dust aerosols from arid and semi-arid areas to the north and northwest of Beijing.In addition, because the Cl - species were mainly from coal combustion and biomass burning.the particles in this category contained 21% of the total Cl -species, indicating that the particles likely included substantial resuspended dust (Zhang et al., 2005b).The modeled ratio of soil dust to PM 2.5 was higher at lightmedium and heavy pollution levels and lower at clean and severe levels, which was consistent with the variation of fine soil (FS) concentrations by pollution levels (Fig. S3).
Traffic emissions were characterized by Sb, Cu, Zn, Cr, Pb and Mn.This source explained 6% of PM 2.5 .It has been reported that the contribution of traffic emission to PM 2.5 was 4-7% (Zheng et al., 2005;Zhang et al., 2013).Cu and Zn could be from tire abrasion, brake linings, lubricants, and corrosion of vehicular parts and tailpipe emission (Li et al., 2016a;Tan et al., 2016).An earlier study found that Sb was also attributable to traffic emissions, particularly in urban areas (Sternbeck et al., 2002).Beijing, the political and cultural center of China, is one of the largest cities in the world with 5.7 million vehicles over an area of approximately 16410 square kilometers (Statistical Communique on National Economic and Social Development of Beijing, 2016).At the clean level, the contribution of traffic emissions to PM 2.5 mass concentration was biggest (9%).At the other three pollution levels, the contribution was stable (4%-5%).
Particles from coal combustion and biomass burning were identified by the high Cl -, K + and Na + content.It has been reported that Cl -is mainly from coal combustion and biomass burning, especially in the heating period of winter in Beijing, and K + was from biomass burning (Sun et al., 2006).This source category represented a mean contribution of 18% to PM 2.5 , which was lower than that found in earlier studies in Beijing (Zhang et al., 2013;Tian et al., 2016).The observation was made in late winter in Beijing when coal burning for heating decreased in comparison with that in the cold season.The contribution of this source factor to PM 2.5 was consistent with the variation of Cl -at different pollution levels (Fig. 2), suggesting the dominance of coal combustion in this category.Except for a minimal contribution (7%) at the severe pollution level, the contribution of this source factor was stable, accounting for 20%-25%.The low contribution at the severe level was likely due to the effective mitigation measures driven by the government.

Trajectory Clustering
Backward trajectory analysis was used to determine the effects of source regions on the chemical properties of PM 2.5 (Fig. 7).The trajectory clusters for each pollution levels, (a) clean, (b) light-medium pollution level), (c) heavy pollution level and (d) severe pollution level, were produced by the clustering algorithm.The proportion of the chemical components in PM 2.5 corresponding to each pollution level was also presented in Fig. 7. Significant changes in air masses and chemical components by source regions are observed for the different pollution levels.
With the increasing pollution level, the proportion of southern trajectories to each cluster increased, and the transport distance decreased.The slow movement of air Fig. 7. 48 h backward air trajectories at the sampling site (39.97°N,116.37°E) at a height of 500 m and chemical mass closures of PM 2.5 mass concentrations at different pollution levels.The backward trajectories were calculated using the NOAA HYSPLIT 4 (details in Section 2.4).parcels indicated that the meteorological conditions became stable.For the clean air periods, the 48 h backward trajectories came from the north and northwest areas far away from Beijing (such as Mongolia, Inner Mongolia and northwest Hebei Province).In those areas, there are few industrial activities and dust storms that frequently occur.These air masses moved very fast: the transported distance of the cluster was the longest among the clusters, and 62% of the air masses during the clean period moved more than 2000 km within 48 h.The concentration of PM 2.5 for this cluster was low, with an average of 30.7 µg m -3 , indicating that the air mass from those areas was clean and did not cause a significant increase of PM 2.5 mass in Beijing (Zhang et al., 1999(Zhang et al., , 2005a)).The largest contribution of FS to PM 2.5 occurred for this cluster (14%), and the secondary inorganic aerosols (SIA) contribution was the lowest (22%) in comparison with other clusters.These results correlated closely with the low RH because the air mass came from arid and semi-arid areas.PM 2.5 composition at the clean level was dominated by organic matter (47%), followed by SIA (sulfate 9%, nitrate 8%, and ammonium 6%) and FS.
A large portion of air masses (44%) circled over the northeast and moved southeast of Beijing before they arrived at the observation site, resulting in a sharply increase in PM 2.5 from the clean to light-medium level (Figs. 7 and S3).The PM 2.5 concentration at light-medium level was approximately 3 times higher than that at the clean level.Furthermore, the secondary inorganic species and water content both increased approximately 6 times from the clean to light-medium pollution level.The air mass from the south was humid and conducive for secondary formation.Although the organic matter (OM) and FS mass concentrations increased, their relative contribution to PM 2.5 decreased slightly by 17% and 2%, respectively.The sharply increase in PM 2.5 from clean to light-medium pollution could be attributable to the regional transport of polluted air mass from the south (Li et al., 2016 b;He et al., 2017;Ma et al., 2017).
In the subsequent pollution levels of heavy and severe, more than 50% of the air masses were from the south and their 48-h transport distance was smaller than 300 km (i.e., airmass movement was slower than 1.7 m s -1 ), indicating that the pollutants occurring on the local scale play an important role in the pollution elevation (Zhang et al., 2017).Elevated PM 2.5 concentrations with a high percentage (42% at the heavy level and 52% at the severe level) of secondary inorganic components and high RH (53% at the heavy level and 64% at the severe level) indicate the substantial contribution of heterogeneous reactions for the increase of these pollution levels (Ma et al., 2017).

CONCLUSIONS
In this study, multiple approaches, including a chemical mass balance, the ISORROPIA II thermodynamic equilibrium model, positive matrix factorization (PMF) and trajectory clustering were applied to characterize the chemical species and explore potential origins of PM 2.5 at different pollution levels in Beijing between February 24 and March 12, 2014, when severe air pollution episodes frequently occurred.
The average concentration of PM 2.5 was 137.7 ± 124.8 µg m -3 .Water-soluble ions were the major contributors of PM 2.5 .The contribution of SIA to the PM 2.5 continuously increased with the increase of both pollution level and relative humidity, indicating an important role of aqueousphase SIA production.SIA contributed a higher proportion in PM 2.5 than in PM 10 , indicating that secondary formation was more easily enriched in fine particles.The acidity of PM 2.5 was substantially higher than that of PM 10 .With the increase in pollution level, the acidity of PM 2.5 increased from 0.96 to 1.08.
Using the PMF model, five major source factors were identified to have contributed to the PM 2.5 : secondary inorganic aerosols, industrial emissions, soil dust, traffic emissions, coal combustion and biomass burning.On average they contributed 46%, 20%, 10%, 6% and 18% to the PM 2.5 , respectively.Industrial emissions were the predominant source of PM 2.5 during clean period, contributing to 60% of PM 2.5 .At the light-medium level, the contributions of four origins to PM 2.5 were approximately consistent, accounted for 17%-29% of PM 2.5 mass concentrations (traffic emission only contributed 4%).Of the five, secondary inorganic aerosols were a predominant source of PM 2.5 at the heavy and severe pollution levels, accounting for 40% and 78%, respectively.In addition, 48 h transported distances of air masses at a 500 m altitude was 2000 km at the clean level and 300 km at the severe level, indicating the near surface air became very stagnant.The proportion of air masses from south pollution areas to the total air masses at each pollution level increased from 0% to 97%.This result implies that the northerly transport of pollutants from the south at the regional and local scale played an important role in the pollution elevation.

Fig. 2 .
Fig. 2. Mass concentrations of water-soluble ions in PM 2.5 and PM 2.5-10 at different pollution levels.

Fig. 4 .
Fig. 4. Profiles of five sources profiles (bars) resolved from the PMF model (in units of µg µg -1 ) and contribution percentages (black dots) from each source factor.

Fig. 5 .
Fig. 5. Time series of half daily (D: daytime; N: nighttime) contributions from each identified source and CMC of PM 2.5 .

Fig. 6 .
Fig. 6.(a) Modeled concentration for each identified main source, (b) fraction of modeled concentration for each identified main source to PM 2.5 at different pollution levels.

Table 1 .
Formulas and references of major components in PM 2.5.