High contributions of fossil sources to more volatile organic aerosol

Sources of particulate organic carbon (OC) with different volatility have rarely been investigated, despite the significant importance for better understanding of the atmospheric processes of organic aerosols. In this study we develop a radiocarbon-based (¹⁴C) approach for source apportionment of more volatile OC (mvOC) and apply to ambient aerosol samples collected in winter in six Chinese megacities. mvOC is isolated by desorbing organic carbon from the filter samples in helium (He) at 200 ^∘C in a custom-made aerosol combustion system for ¹⁴C analysis. Evaluation of this new isolation method shows that the isolated mvOC amount agrees very well with the OC1 fraction (also desorbed at 200 ^∘C in He) measured by a thermal–optical analyzer using the EUSAAR_2 protocol. The mvOC, OC and elemental carbon (EC) of 13 combined PM_2.5 samples in six Chinese cities are analyzed for ¹⁴C to investigate their sources and formation mechanisms. The relative contribution of fossil sources to mvOC is 59±11 %, consistently larger than the contribution to OC (48±16 %) and smaller than that to EC (73±9 %), despite large differences in fossil contributions in different cities. The average difference in the fossil fractions between mvOC and OC is 13 % (range of 7 %–25 %), similar to that between mvOC and EC (13 %, with a range 4 %–25 %). Secondary OC (SOC) concentrations and sources are modeled based on the ¹⁴C-apportioned OC and EC and compared with concentrations and sources of mvOC. SOC concentrations (15.4±9.0 µg m⁻³) are consistently higher than those of mvOC (3.3±2.2 µg m⁻³), indicating that only a fraction of SOC is accounted for by the more volatile carbon fraction desorbed at 200 ^∘C. The fossil fraction in SOC is 43 % (10 %–70 %), lower than that in mvOC (59 %, with a range of 45 %–78 %). Correlation between mvOC and SOC from nonfossil sources (mvOC_nf vs. SOC_nf) and from fossil sources (mvOC_fossil vs. SOC_fossil) is examined to further explore sources and formation processes of mvOC and SOC.

1 Introduction

Aerosol particles are of importance for atmospheric chemistry and physics and exert crucial influence on the climate system, air quality and human health (Fuzzi et al., 2015). The carbonaceous fraction of aerosol comprises a large fraction of the fine aerosol mass (20 %–80 %) (Cao et al., 2007; Tao et al., 2017). Carbonaceous aerosols are often operationally subdivided into organic carbon (OC) and elemental carbon (EC). OC consists of a large variety of organic species that cover a great range of volatilities and are not or only weakly light absorbing, while EC is nonvolatile, resistant to chemical transformation and strongly light absorbing (Pöschl, 2005). It should be noted that since no clear distinction between OC and EC exists, OC and EC are operationally defined based on measurement techniques (Petzold et al., 2013). When a thermal–optical method is used to separate and determine OC and EC, EC is described as a thermally refractory carbon continuum and OC is weakly refractory (Pöschl, 2003, 2005; Petzold et al., 2013).

EC is exclusively emitted as primary aerosols from incomplete combustion of biomass (e.g., wood, crop residues and grass) and fossil fuels (e.g., coal, gasoline and diesel). OC can be of both primary and secondary origin. Primary OC (POC) is emitted directly from nonfossil (e.g., biomass burning, biogenic emissions and cooking) and fossil sources. Secondary OC (SOC) is formed in the atmosphere via atmospheric oxidation of volatile organic compounds (VOCs) (Jacobson et al., 2000; Hallquist et al., 2009; Bond et al., 2013). High concentrations of aerosol particles have been observed in many cities, especially in China, where carbonaceous aerosols dominate particle mass concentrations, with SOC responsible for a large fraction of OC (Guo et al., 2014; Huang et al., 2014; Fang et al., 2017). Thus, understanding the sources, formation and transformation of EC, OC and SOC is important for better understanding of air pollution.

For aerosol source apportionment, radiocarbon (¹⁴C) analysis of OC and EC has been used to quantify their fossil and nonfossil fractions, based on the fact that emissions from fossil sources are ¹⁴C-free due to long-time decay, whereas nonfossil emissions contain the contemporary ¹⁴C content (e.g., Szidat et al., 2004, 2006; Dusek et al., 2013, 2017). Earlier ¹⁴C measurements of OC and EC found that fossil sources contribute more to EC than OC (Heal, 2014 and references therein). However, direct ¹⁴C measurements of SOC are not possible yet, due to the technical and conceptual difficulties in isolating the SOC fraction from filter samples. In positive-matrix-factorization-based (PMF) source apportionment of an aerosol mass spectrometer (Jayne et al., 2000) dataset, oxidized organic aerosols (OOA, also referred to as secondary organic aerosol, SOA) can be separated into semi-volatile OOA (SV-OOA) and low-volatility OOA (LV-OOA) based on their volatility (e.g., Huffman et al., 2009a; Wang et al., 2017). Volatility is an important physical property as it determines the partitioning between gas and particulate phases of organic species (Donahue et al., 2006, 2009). Direct measurement of OA volatility is challenging because OA is a mixture of thousands of individual organics spanning a wide range of volatilities (Donahue et al., 2011). Different approaches have been used to estimate the OA volatility. For example, using a thermodenuder (TD) coupled with an aerosol mass spectrometer (Burtscher et al., 2001; Wehner et al., 2002), the volatility of different OA components, such as hydrocarbon-like OA (HOA), biomass burning OA (BBOA) and OOA, can be estimated for ambient aerosols. Source and ambient studies indicate that BBOA and HOA are generally more volatile than OOA. Meanwhile, the volatility of BBOA can be quite variable, depending on type of biomass and the combustion conditions, and either higher or lower than that of HOA (e.g., Grieshop et al., 2009c; Huffman et al. 2009a; Paciga et al., 2016; Cao et al., 2018). For example, Huffman et al. (2009b) measured the volatility of primary OA from biomass burning and found that the more volatile BBOA was generally dominated by smoldering combustion, while the less volatile BBOA was more influenced by flaming combustion. Paciga et al. (2016) found that the BBOA was more volatile than HOA in an urban background site in Paris, with 50 % semi-volatile organic components (SVOCs with effective saturation concentrations C^* of 1–100 µg m⁻³ range) and 30 % low-volatility organic compounds (LVOCs with C^* of 10⁻³–0.1 µg m⁻³). Gaston et al. (2016) found that levoglucosan (i.e., the biomass burning tracer) was semi-volatile and had the potential to form low-volatility products during the chemical aging of ambient biomass burning aerosol, applying high-resolution time-of-flight chemical ionization mass spectrometer (HRTOF-CIMS) coupled to a Filter Inlet for Gases and AEROsols (FIGAERO) instrument for aerosol in the northwestern US.

In some recent studies, a thermal desorption approach has been used for analysis of filter samples at lower and higher temperatures as an indicator of volatility (Holzinger et al., 2013; Lopez-Hilfiker et al., 2014, 2015; Timkovsky et al., 2015; Masalaite et al., 2017, 2018). Even though the desorption temperature is not a direct measure of the particle volatility, OC desorbed from the filter at lower temperature tends to be more volatile, whereas the less volatile (i.e., more refractory) OC tends to be desorbed at higher temperatures. Vodička et al. (2015) and Masalaite et al. (2017) found that OC in urban samples was desorbed at lower temperatures compared to coastal, forest and rural background samples in Eastern Europe. Both studies suggest that OC is more volatile in urban areas close to the emission sources. Keller and Burtscher (2017) found that after aging of biomass burning emissions, less OC desorbed at lower temperatures. The reduction in volatility after aging agrees with previous TD aerosol mass spectrometer studies (Grieshop et al., 2009a, b). Even though the thermal desorption and TD measurements are different methods, results of both methods are usually in qualitative agreement and show similar trends. Earlier studies (e.g., Grieshop et al., 2009c; Chen et al., 2007) show similar trends qualitatively. Ma et al. (2016) made first efforts to link the OC fractions desorbed at different temperatures from filter samples on a thermal-optical analyzer to the volatility basis set, showing that OC fractions desorbed at lower temperatures (e.g., 140 and 280 ^∘C) are semi-volatile.

Both POC and SOC can contribute to the OC fraction desorbed at lower temperatures. A number of previous studies have found that at least a part of POC from various sources is semi-volatile, including wood burning, gasoline and diesel vehicle exhausts and cooking (Lipsky and Robinson, 2006; Shrivastava et al., 2006; Robinson et al., 2007; Grieshop et al., 2009b, c; May et al., 2013a, b). Moreover, a significant part of freshly formed SOC is semi-volatile and will contribute to the more volatile OC fraction desorbed at lower temperatures (Holzinger et al., 2010; Salo et al., 2011; Meusinger et al., 2017; Gkatzelis et al., 2018). The relative contributions from fossil and nonfossil sources to OC can be dependent on the volatility. For example, recent studies during winter in Lithuania have shown that vehicular sources are associated with OC fractions desorbed at lower temperatures, while biomass burning (as a nonfossil source) contributes more strongly to less volatile OC fractions desorbed at higher temperatures (Masalaite et al., 2017, 2018).

Although an increasing number of studies has shown that OC from different emission sources and/or at different aging status may have different volatility, a quantitative study of the sources of OC with different volatility is scarce. In this study, wintertime fine particulate matter (PM_2.5, particles with aerodynamic diameter < 2.5 µm) samples from six Chinese megacities are studied. Sources of carbonaceous aerosols, including different carbon fractions such as OC, EC and SOC, are estimated by ¹⁴C source apportionment. In addition, we operationally define a more volatile OC fraction (mvOC) as the carbon fraction desorbed from the filter samples at 200 ^∘C in helium (He) and investigate sources of mvOC. The source contribution to mvOC is measured unambiguously by ¹⁴C.

2 Methods

2.1 Sampling

Simultaneous sampling was made during winter 2013/2014 in two northern (Beijing and Taiyuan) and three southern (Chongqing, Guangzhou and Shanghai) Chinese cities (Supplement Fig. S1). The samples from Xi'an in northern China were collected during winter 2015/2016. These sites were selected to represent urban-scale concentrations and located in the university or research center campus, > 100 m from local sources such as main roadways (Table S1 in the Supplement). The 24 h integrated PM_2.5 samples were collected from 10:00 to 10:00 the next day (local standard time, LST). Filter samplers were deployed on rooftops about 6 to 20 m above ground level. In Xi'an, Beijing, Guangzhou and Shanghai, PM_2.5 samples were collected on pre-baked (780 ^∘C for 3 h) quartz fiber filter (QM-A, Whatman Inc., Clifton, NJ, USA, 20.3 cm ×25.4 cm) using a high-volume sampler (TE-6070 MFC, Tisch Inc., Cleveland, OH, USA) at a flow rate of 1.0 m³ min⁻¹. In Taiyuan and Chongqing, samples were collected on pre-baked 47 mm Whatman quartz microfiber filters (QM-A) using mini-volume samplers at a flow rate of 5 L min⁻¹ (Airmetrics, OR, USA). After sampling, all filters were packed in a pre-baked aluminum foil (450 ^∘C for 3 h), sealed in polyethylene bags and stored in a freezer at −18 ^∘C until analysis.

2.2 Determination of carbon fractions by thermal-optical analysis

For the PM_2.5 samples collected in Xi'an, filter pieces of 1.5 cm² were taken for OC and EC analysis using a carbon analyzer (Model 5L, Sunset Laboratory, Inc., Portland, OR, USA) following the thermal-optical transmittance protocol EUSAAR_2 (European Supersites for Atmospheric Aerosol Research; Cavalli et al., 2010). The EUSAAR_2 protocol defines OC1 as the carbon fraction that desorbs in helium (He) at 200 ^∘C for 2 min. OC1 is compared with more volatile OC (mvOC) also extracted in He at 200 ^∘C but for 5 min using our aerosol combustion system (ACS) (Dusek et al., 2014). No charred OC is observed by the transmittance signal in the OC1 stage during the thermal–optical analysis. For PM_2.5 samples collected in Beijing, Taiyuan, Chongqing, Guangzhou and Shanghai, filter pieces of 0.5 cm² were used to measure OC and EC using a Desert Research Institute (DRI) Model 2001 Thermal/Optical Carbon Analyzer (Atmoslytic Inc., Calabasas, CA, USA), following the IMPROVE_A thermal–optical reflectance (TOR) protocol (Chow et al., 2007). OC fractions of EUSAAR_2 protocol are desorbed in He at 200 ^∘C (OC1), 300 ^∘C (OC2), 450 ^∘C (OC3) and 650 ^∘C (OC4). Different from EUSAAR_2 protocol, IMPROVE_A (Interagency Monitoring of Protected Visual Environments) protocol defines OC fractions at stepwise temperature of 140, 280, 480 and 580 ^∘C in He. Details of the carbon fraction measurements were described in our previous work (Cao et al., 2013; Zenker et al., 2017).

2.3 Radiocarbon (¹⁴C) measurements of mvOC, OC and EC

2.3.1 Sample selection for ¹⁴C analysis

For ¹⁴C analysis, we selected the samples carefully to cover periods of low and high total carbon (TC, the sum of OC and EC) and PM_2.5 concentrations to get samples representative of various pollution conditions that did occur in each city. Two or three composite samples from each city representing high (H), medium (M) and low (L) TC concentrations were selected for ¹⁴C analysis of EC, OC and mvOC (Fig. S2). Each composite sample consists of two to four 24 h filter pieces with similar TC loadings and air mass backward trajectories (Fig. S2, Table S1).

2.3.2 Extraction of mvOC, OC and EC

Three separate extractions were performed for mvOC, OC and EC on our aerosol combustion system (ACS) (Dusek et al., 2014). The ACS consists of a combustion tube and a purification line. Aerosol filter pieces placed on the filter boat are heated at different temperatures in pure He or O₂ and oxidized through the platinum catalyst in the combustion tube. The resulting CO₂ is isolated and separated from other gases (e.g., NO_x, water vapor) in the purification line. The purification line is equipped with an oven filled with copper grains and silver wire heated at 650 ^∘C to remove NO_x and liberated halogen, a U-type tube cooled with a dry ice–ethanol mixture for water removal, and a flask containing phosphorous pentoxide (P₂O₅) for removal of any trace water. The amount of purified CO₂ is manometrically quantified and subsequently stored in flame-sealed glass ampoules.

mvOC is desorbed by heating the filter pieces at 200 ^∘C in He for 5 min. After introducing the filter, the ovens are flushed with He for 10 min. Pilot tests show that the flushing time (10, 15 or 60 min) before heating the filter pieces does not affect the desorbed amount of mvOC (Table S2). OC is combusted by heating filter pieces at 375 ^∘C in pure O₂ for 10 min. EC is extracted from a separate filter piece after removing OC completely. First water-soluble OC is removed from the filter through water extraction (Dusek et al., 2014) to minimize the charring of organic materials (Yu et al., 2002; Zhang et al., 2012). Then, by heating the filter pieces at 375 ^∘C in O₂ for 10 min, most water-insoluble OC can be removed. Subsequently, the oven temperature is increased to 450 ^∘C for 3 min to remove most refractory OC that is left on the filter. However, during this step some less refractory EC might be lost. Finally, the remaining EC is combusted by heating at 650 ^∘C in O₂ for 5 min (Dusek et al., 2014, 2017; Zenker et al., 2017).

Contamination introduced by the isolation procedure is determined by following exactly the same isolating procedures with either an empty filter boat or with preheated filters (at 650 ^∘C in O₂ for 10 min). The contamination introduced by the combustion process yields 0.72±0.44 µgC EC, 0.85±0.49 µgC OC and 0.52±0.31 µgC mvOC per extraction. Compared with our sample size of 30–391 µgC EC, 30–445 µgC OC and 15–121 µgC mvOC, the blanks are ≤7 % of the mvOC sample amount and < 3 % of OC and EC sample amount and thus relatively small compared to our sample sizes. The mvOC concentrations reported in this study are corrected for contamination by subtraction. OC and EC concentrations are measured following thermal-optical protocols using carbon analyzers (Sect. 2.2), thus are not affected by the isolation procedure using the ACS. For ¹⁴C values (Eq. 1), the contamination is assessed but not used for further ¹⁴C data correction for mvOC, OC and EC because the corrections for the small blanks will not introduce large uncertainties to the data, as explained in Supplement Sect. S3.

Two standards with known ¹⁴C∕¹²C ratios are combusted using the ACS as quality control for the combustion process: an oxalic acid standard and anthracite. Small amounts of solid standard materials are directly put on the filter boat of the combustion tube and heated in O₂ at 650 ^∘C for 10 min. In the further ¹⁴C analysis, the CO₂ derived from combustion of the standards is treated exactly like the samples. Therefore, the contamination introduced by the combustion process can be estimated from the deviation of measured values from the nominal values of the standards (Table S3) (Dusek et al., 2014). The contamination inferred in this indirect way is below 1.5 µgC per extraction, which is slightly higher than the directly measured contamination of OC and EC separately but in the range of a TC contamination. It is also relatively small compared to the size of OC (30–445 µgC) and EC samples (30–391 µgC) in this study and thus can be neglected (Sect. S3). The F¹⁴C (Eq. 1) of the contamination can also be indirectly inferred and varied between 0.2 and 0.6 for standards measured on the ACS over the last 2 years. This is broadly within the range of our samples and implies that a correction for F¹⁴C will even be considerably smaller than 3 %.

2.3.3 ¹⁴C measurements by accelerator mass spectrometer (AMS)

AMS measurements were conducted at the Centre for Isotope Research (CIO) at the University of Groningen. The extracted CO₂ is released from the glass ampules and directed to a gas inlet system (Ruff et al., 2007), where the sample is diluted with He to 5 % CO₂ (Salazar et al., 2015). The CO₂∕He mixture is directly fed into the Cs sputter ion sources of the Mini Carbon Dating System (MICADAS) AMS at a constant rate (Synal et al., 2007).

The ¹⁴C∕¹²C ratio of an aerosol sample is usually reported relative to an oxalic acid standard (OXII) and expressed as fraction modern (F¹⁴C). The ¹⁴C∕¹²C ratio of the standard is related to the unperturbed atmosphere in the reference year of 1950 by multiplying it by a factor of 0.7459 (Mook and Van Der Plicht, 1999; Reimer et al., 2004):

$$\begin{array}{}\text{(1)}& {\mathrm{F}}^{\mathrm{14}}\mathrm{C}={\displaystyle \frac{{(}^{\mathrm{14}}\mathrm{C}{/}^{\mathrm{12}}\mathrm{C}{)}_{\mathrm{sample},[-\mathrm{25}]}}{\mathrm{0.7459}\times {(}^{\mathrm{14}}\mathrm{C}{/}^{\mathrm{12}}\mathrm{C}{)}_{\mathrm{OXII},[-\mathrm{25}]}}},\end{array}$$

where the ¹⁴C∕¹²C ratio of the sample and standard are both corrected for machine background and normalized to δ¹³C $=-\mathrm{25}$ ‰. Aerosol carbon from fossil sources has F¹⁴C = 0 due to the extinction of ¹⁴C after long-time decay. Aerosol carbon from contemporary (or nonfossil) sources should have F¹⁴C ∼1 in an atmosphere undisturbed by human influences. However, the F¹⁴C values of contemporary carbon are higher than 1 due to the nuclear bomb tests that nearly doubled the ¹⁴CO₂ in the atmosphere in the 1960s. The ¹⁴CO₂ produced by the nuclear bomb tests has been taken up by oceans and the biosphere and diluted by the ¹⁴C-free CO₂ emitted by the fossil fuel burning. Currently, F¹⁴C of the atmosphere CO₂ is ∼1.04 (Levin et al., 2010).

The F¹⁴C values are corrected for cross contamination (also known as memory effect) (Wacker et al., 2010) using alternate measurements of OXII and ¹⁴C-free material as gaseous standards. Correction for instrument background (Salazar et al., 2015) is done by subtracting the memory-corrected F¹⁴C values of the ¹⁴C-free standard. Finally, the values are normalized to the average value of the (memory- and background-corrected) OXII standards. All standards used for the corrections are measured on the same day as the samples.

2.4 Estimation of source contributions to different carbon fractions

F¹⁴C of EC, OC and mvOC (i.e., F¹⁴C_(EC), F¹⁴C_(OC) and F¹⁴C_(mvOC), respectively) were directly measured. We define “more refractory organic carbon” (mrOC) as the difference between OC and mvOC. F¹⁴C of mrOC (F¹⁴C_(mrOC)) was calculated by isotope mass balance:

$$\begin{array}{}\text{(2)}& {\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{mrOC}\right)}={\displaystyle \frac{{\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{OC}\right)}\times \mathrm{OC}-{\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{mvOC}\right)}\times \mathrm{mvOC}}{\mathrm{OC}-\mathrm{mvOC}}}.\end{array}$$

F¹⁴C_(EC) was converted to the fraction of biomass burning (f_bb(EC)) by dividing with an F¹⁴C value representative of typical biomass burning emissions (F¹⁴C_bb). Analogously, the fraction of nonfossil OC, mvOC and mrOC (i.e., f_nf(OC), f_nf(mvOC) andf_nf(mrOC), respectively) were estimated from their corresponding F¹⁴C values and F¹⁴C of nonfossil sources (F¹⁴C_nf). F¹⁴C_bb and F¹⁴C_nf are estimated as 1.10±0.05 and 1.09±0.05 (Lewis et al., 2004; Mohn et al., 2008; Palstra and Meijer, 2014), respectively, based on tree-growth models and the assumption that wood burning dominates biomass burning. F¹⁴C_bb for EC is slightly bigger than F¹⁴C_nf for OC because, besides biomass burning, biogenic emissions also contribute to OC but have a smaller F¹⁴C than that of biomass burning. The estimation of F¹⁴C_bb and F¹⁴C_nf has been reported in our previous study (Ni et al., 2018).

Then, f_bb(EC) was used to determine EC mass concentrations from nonfossil biomass (EC_bb) and fossil fuel combustion (EC_fossil):

$$\begin{array}{}\text{(3)}& {\displaystyle }{\mathrm{EC}}_{\mathrm{bb}}=\mathrm{EC}\times f_{\mathrm{bb}}\left(\mathrm{EC}\right),\text{(4)}& {\displaystyle }{\mathrm{EC}}_{\mathrm{fossil}}=\mathrm{EC}-{\mathrm{EC}}_{\mathrm{bb}}.\end{array}$$

Analogously, mass concentrations of OC, mvOC and mrOC from nonfossil sources (OC_nf, mvOC_nf and mrOC_nf, respectively) and fossil sources (OC_fossil, mvOC_fossil and mrOC_fossil, respectively) were determined.

Secondary OC (SOC) includes SOC from fossil (SOC_fossil) and nonfossil sources (SOC_nf):

$$\begin{array}{}\text{(5)}& \mathrm{SOC}={\mathrm{SOC}}_{\mathrm{fossil}}+{\mathrm{SOC}}_{\mathrm{nf}}.\end{array}$$

Fraction fossil in total SOC (f_fossil(SOC)) was formulated as follows:

$$\begin{array}{}\text{(6)}& f_{\mathrm{fossil}}\left(\mathrm{SOC}\right)={\displaystyle \frac{{\mathrm{SOC}}_{\mathrm{fossil}}}{{\mathrm{SOC}}_{\mathrm{fossil}}+S{\mathrm{OC}}_{\mathrm{nf}}}}.\end{array}$$

For the following calculations we assume that SOC_nf can be approximated by OC from nonfossil sources excluding primary biomass burning (OC_o,nf; OC other nonfossil). In principle, OC_o,nf includes SOC_nf and nonfossil primary OC from vegetative detritus, bioaerosols, resuspended soil organic matter, or cooking. But the contributions from plant detritus, bioaerosols and spores to PM_2.5 are likely small (Song et al., 2006; Hu et al., 2010; Guo et al., 2012). If cooking sources are significant then this assumption results in an upper limit of SOC_nf.

$$\begin{array}{}\text{(7)}& {\mathrm{SOC}}_{\mathrm{nf}}\cong {\mathrm{OC}}_{\mathrm{o},\mathrm{nf}}={\mathrm{OC}}_{\mathrm{nf}}-{\mathrm{POC}}_{\mathrm{bb}},\end{array}$$

where OC_o,nf was calculated by the difference between OC_nf and primary OC from biomass burning (POC_bb). POC_bb was calculated by multiplying EC_bb with the primary OC∕EC ratio of biomass burning (r_bb):

$$\begin{array}{}\text{(8)}& {\mathrm{POC}}_{\mathrm{bb}}={\mathrm{EC}}_{\mathrm{bb}}\times r_{\mathrm{bb}}.\end{array}$$

SOC_fossil was calculated by subtracting primary fossil OC (POC_fossil) from OC_fossil. POC_fossil was estimated by multiplying EC_fossil with primary OC∕EC ratio of fossil fuel combustion (r_fossil):

$$\begin{array}{}\text{(9)}& {\displaystyle }{\mathrm{SOC}}_{\mathrm{fossil}}={\mathrm{OC}}_{\mathrm{fossil}}-{\mathrm{POC}}_{\mathrm{fossil}},\text{(10)}& {\displaystyle }{\mathrm{POC}}_{\mathrm{fossil}}={\mathrm{EC}}_{\mathrm{fossil}}\times r_{\mathrm{fossil}}.\end{array}$$

Fossil sources in China are almost exclusively from coal combustion and vehicle emissions, thus r_fossil was estimated as follows:

$$\begin{array}{}\text{(11)}& r_{\mathrm{fossil}}=r_{\mathrm{coal}}\times p+r_{\mathrm{vehicle}}\times (\mathrm{1}-p),\end{array}$$

where r_coal and r_vehicle are the primary OC∕EC ratio of coal combustion and vehicle emissions, respectively. The r_bb, r_coal and r_vehicle varies with the fuel types and properties, combustion conditions, sampling and analysis methods, etc. Best estimate of r_bb(4±1; average ± standard deviation), r_coal(2.38±0.44) and r_vehicle(0.85±0.16) is done through a literature search and described in our earlier study (Ni et al., 2018). p is the fraction of EC from coal combustion (EC_coal) in EC_fossil.

We used two different methods to estimate p. (1) Since both coal combustion and vehicle emissions do not contain ¹⁴C, they can not be distinguished by ¹⁴C measurements alone. Therefore, p is randomly chosen from 0 to 1, that is no constraint on p values. (2) EC from coal combustion is on average more enriched in the stable carbon isotope ¹³C compared to vehicle emissions. Therefore, complementing ¹⁴C results of EC with measurements of the ¹³C∕¹²C ratios of EC (expressed as δ¹³C_EC in Eq. S1; Sect. S1) allows separation of EC_fossil into EC_coal and EC from vehicle emissions (EC_vehicle). Samples taken from Beijing, Shanghai and Guangzhou using high-volume samplers had enough material for analysis of both F¹⁴C_EC and δ¹³C_EC, thus EC_coal and EC_vehicle are separated as described in detail in Sect. S2. In brief, the fraction of coal combustion and vehicle emissions in EC was calculated from measured F¹⁴C_EC and δ¹³C_EC for ambient EC combined with the source signatures. Bayesian Markov chain Monte Carlo (MCMC) calculations were used to account for the uncertainties in the source signatures and the measurement uncertainties (Andersson, 2011; Andersson et al., 2015). The results of the Bayesian calculations are the posterior probability density functions (PDFs) for the relative contributions of each source to EC (f_coal and f_vehicle; Fig. S3). The p was formulated as follows:

$$\begin{array}{}\text{(12)}& p={\displaystyle \frac{f_{\mathrm{coal}}}{f_{\mathrm{fossil}}}}={\displaystyle \frac{f_{\mathrm{coal}}}{f_{\mathrm{coal}}+f_{\mathrm{vehicle}}}}.\end{array}$$

The PDF of p was derived from the PDF of f_coal and f_vehicle and shown in Fig. S4.

To propagate uncertainties, a Monte Carlo simulation with 10 000 individual calculations was conducted. For each individual calculation, F¹⁴C_(EC), F¹⁴C_(OC), F¹⁴C_(mvOC), and concentrations of EC, OC and mvOC were randomly chosen from a normal distribution that is symmetric around the measured values with the experimental uncertainties as standard deviation (SD; Table S4). For F¹⁴C_bb, F¹⁴C_nf, r_bb, r_coal and r_vehicle, random values were chosen from a triangular frequency distribution, which has its maximum at the central value and at 0 at the lower limit and upper limit. For p ranging from 0 to 1 (no ¹³C constraints), p was randomly chosen from a uniform distribution. For p constrained by F¹⁴C and δ¹³C using MCMC (hereafter, ¹³C-constrained p), random values from the respective PDF of p were used (Fig. S4). In this way 10 000 different estimation of f_bb(EC),f_nf(OC), f_nf(mvOC), f_nf(mrOC), EC_bb, EC_fossil, OC_nf, OC_fossil, mvOC_nf, mvOC_fossil, mrOC_nf, mrOC_fossil, SOC_nf, SOC_fossil, SOC and f_fossil(SOC) were calculated (Tables S5, S6, S7, S8). The derived average represents the best estimate and the SD represents the combined uncertainties.

3 Results

3.1 Method evaluation and quality control for mvOC extraction

The separation of OC and EC for ¹⁴C analysis using our aerosol combustion system (ACS) were thoroughly evaluated by Dusek et al. (2014, 2017). It is thus necessary to validate the new extraction method for isolating mvOC. The reproducibility of the extracted mvOC amount was tested for two independent test filters with mvOC loadings of 6 and 18 µg cm⁻², respectively (Fig. S5). The coefficient of variation was determined as a measure of reproducibility. The reproducibility was found to be ∼5 % (n=9).

Figure 1 (a) OC1 loading (µg cm⁻²) measured with a thermal–optical method (EUSAAR_2 protocol using a Sunset analyzer) and mvOC loading measured using our custom-made aerosol combustion system. The outliers are shown in red circle and square with their sample name on the top x axis. (b) F¹⁴C_(mvOC) of winter-M and winter-H with high and low recoveries with respect to OC1. The red numbers in brackets in panels (a) and (b) give the respective outliers' mvOC recoveries to OC1.

Since carbon fractions (e.g., OC1) at different desorption temperatures have mostly been measured using the EUSAAR_2 protocol in many previous studies (e.g., Vodička et al., 2015; Keller and Burtscher, 2017), our goal is to define the mvOC fraction as representative of OC1. Therefore, the mvOC is desorbed at 200 ^∘C, the same temperature as used for OC1 in the EUSAAR_2 protocol. However, the extracted amounts of the ACS system might differ due to different heating rates and length of the temperature step. The winter samples from Xi'an as well as the two test filters described above are used to compare mvOC concentrations from the ACS system to OC1. For most samples, excellent agreement was found between mvOC and OC1 (Fig. 1a) and most data points fall close to the 1:1 line. However, there are three data points deviating largely from the 1:1 line (red circle and square in Fig. 1). The two red squares represent the mvOC extraction for sample winter-H and winter-M using larger filter pieces (i.e., more mvOC in µgC per extraction). With larger filter pieces, the mvOC (µgC cm⁻² filter area) extracted by the ACS system is significantly lower than that measured by the Sunset analyzer. The recoveries of these two outliers are 0.59 for sample winter-H and 0.74 for winter-M, calculated by dividing mvOC mass by OC1 mass. For winter-H, the low recovery is also repeatable with the same filter area (Fig. S5). The low recoveries for large filter pieces may result from the lower temperature (< 200 ^∘C) towards the ends of the filter boat. At the relatively low temperature of 200 ^∘C, only a 3.5 cm length at the center of a 12 cm combustion tube is maintained at 200 ^∘C and outside of that 3.5 cm the temperature is lower than 200 ^∘C, e.g., ∼170 ^∘C at the end of the combustion tube as measured by a thermocouple. When filter pieces are large and placed outside the centered 3.5 cm, the desorption temperature for part of the filter pieces will lower than 200 ^∘C, leading to lower desorbed mvOC amount. Another possibility is saturations of the catalyst (platinum, Pt) in the combustion line of ACS system. Pt works as catalyst by collecting oxygen atoms on the surface as has been demonstrated by direct observation of an ultra-thin oxygen layer on the Pt surface by a microscope (Spronsen et al., 2017). This is used to oxidize CO and hydrocarbons to CO₂ in a reducing atmosphere. Thus, for large sample amounts it is possible that the oxygen on the catalyst could not be sufficient to oxidize all desorbed CO and hydrocarbons to CO₂. However, we observed recoveries near 100 % for mvOC amounts up to 120 µg, which was higher than the total amount of desorbed mvOC for the winter-H sample. Therefore, limited catalyst capacity is not the likely explanation for the low recoveries. For subsequent experiments we consequently placed the filter pieces carefully in the 3.5 cm, 200 ^∘C section of the combustion tube and avoided stacking multiple filter pieces to ensure a desorption temperature of 200 ^∘C and sufficient He supply.

To examine the effect of the low recoveries of mvOC on the F¹⁴C_(mvOC), we compare the F¹⁴C_(mvOC) of winter-M and winter-H samples for both high and low mvOC recoveries. F¹⁴C_(mvOC) for low recoveries is roughly 0.05 (absolute value, Fig. 1b) higher than for high recoveries, which is non-negligible compared to the measurement uncertainty. In addition, to validate the measured F¹⁴C_(mvOC) for the combined sample winter-M, we also extracted mvOC separately from those three filters that were combined for the composite winter-M sample. Figure 2 shows the F¹⁴C_(mvOC) of combined winter-M and those of the three individual filter samples. From the F¹⁴C_(mvOC) and mvOC mass of the individual filters, we can estimate the expected F¹⁴C_(mvOC) of the combined winter-M sample using the isotope mass balance equation. The F¹⁴C_(mvOC) of winter-M calculated from individual filter pieces is 0.524, which is quite similar to the measured F¹⁴C_(mvOC) of 0.529±0.007 for the combined winter-M with recovery close to 1.

Figure 2 F¹⁴C_(mvOC) (shown in circle) and mvOC loading (vertical bar) of the three daily filters (“filter1”, “filter2” and “filter3” shown in dark green) that comprise winter-M and the composite winter-M (shown in blue in the gray area) with mvOC recovery of ∼1. The dashed dark green line indicates the expected mvOC loading and the horizontal solid green line indicates the F¹⁴C_(mvOC) for the combined winter-M (weighted mean ± standard deviation; 0.524±0.028). The latter is calculated from the mvOC loading (µg cm⁻²) and F¹⁴C_(mvOC) of the three daily filters using the isotope mass balance equation. F¹⁴C_(mvOC) uncertainties are indicated but are too small to be visible.

We conclude that F¹⁴C_(mvOC) of samples with mvOC recoveries of ∼1 are reliable and used in the following discussion. The mvOC recovery of sample winter-L (red circle in Fig. 1) is also low (0.51), and we could not repeat it due to the limited filter material. For sample winter-L, we take the OC1 concentrations as mvOC and the measured F¹⁴C_(mvOC) values but assign a bigger absolute uncertainty of 0.05, due to its low mvOC recoveries. This is based on the difference in F¹⁴C_(mvOC) for winter-H and winter-M with low and high mvOC recoveries, which is roughly 0.05 (Fig. 1b).

3.2 mvOC, OC and EC concentrations

Figure 3 shows the concentrations of mvOC, OC and EC and the mvOC contributions to OC (%) for the selected samples in the six studied Chinese megacities. mvOC and OC concentrations averaged 3.3±2.2 µg m⁻³ (range of 0.7–7.4 µg m⁻³) and 30.0±13.8 µg m⁻³ (8.8–50.4 µg m⁻³), respectively. EC concentrations ranged from 2.5 to 14.8 µg m⁻³, with an average of 6.9±3.6 µg m⁻³. High TC concentrations were found in Taiyuan (60 µg m⁻³ for sample Taiyuan-H), Chongqing (59 µg m⁻³ for Chongqing-H), Beijing (57 µg m⁻³ for Beijing-H) and Xi'an (57 µg m⁻³ for winter-H) in descending order. Of these cities, Chongqing is located in southern China, where there is no official heating season using coal in winter. This study nevertheless indicates severe pollution of carbonaceous aerosols in Chongqing. TC concentrations in the other southern Chinese cities (Shanghai and Guangzhou) were much lower than those in Chongqing (Fig. 3).

Figure 3 Concentrations of mvOC, mrOC and EC (µg m⁻³) and contribution of mvOC to total OC (%) in three cities in northern China (Xi'an, Beijing and Taiyuan) and three in southern China (Shanghai, Chongqing, Guangzhou). mrOC concentrations are calculated by the difference between OC and mvOC. For details, see Table S4.

The fraction of mvOC in total OC (mvOC∕OC in Fig. 3) gives an indication of OC volatility. The mvOC contributed on average 10.5±3.3 % to OC, ranging from 3 % to 15 %. The mvOC∕OC varies between samples within the same city and between cities, indicating complicated sources and atmospheric processing of OC. The variations might also be partially attributed to the different protocols used for OC quantification. OC in Xi'an is measured with the EUSAAR_2 protocol (up to 650 ^∘C for desorbing OC), whereas the IMPROVE_A protocol (up to 580 ^∘C for OC) was used for the other five cities. However, Han et al. (2016) found that the absolute OC concentrations determined by EUSAAR_2 do not differ much from those determined by IMPROVE_A (22.6±12.0 µg m⁻³ vs. 19.7±10.7 µg m⁻³) for 1-year PM_2.5 samples in Xi'an during 2012/2013. Because of the small differences of OC between the two protocols, we think the comparison of OC concentrations and mvOC∕OC amongst the six cities is justified. The rest of OC (∼90 %) was contributed to the mrOC. mrOC concentrations averaged 26.8±12.0 µg m⁻³, ranging from 7.9 µg m⁻³ for sample Guangzhou-L to 43.1 µg m⁻³ for Beijing-H.

Direct comparison of our results with previous works is somewhat difficult because different thermal–optical protocols were used. The averaged mvOC concentration (3.3±2.2 µg m⁻³) in winter for the six studied sites in China is higher than winter concentrations at an urban background site (1.6±1.7 µg m⁻³), a rural background site (0.7±0.6) in Prague (Vodička et al., 2015), where 200 ^∘C in He was also applied to desorb this OC fraction using EUSAAR_2 protocol. The mvOC∕OC ratio in our study is smaller than that of ambient samples from various other locations and much smaller than that of fresh source samples. For example, the urban background site in Prague had a mvOC∕OC ratio of 28 % and the rural background site of 17 % (Vodička et al., 2015). The contribution of OC1 to total OC was as high as ∼60 % for primary biomass burning measured by EUSAAR_2 thermal–optical protocol (Keller and Burtscher, 2017). For vehicle emissions, the OC fractions desorbed at 140 and 280 ^∘C are the major OC fractions measured by the IMPROVE_A protocol, contributing ∼30 % and ∼20 % to total OC in tunnels in Taiwan, respectively (Zhu et al., 2014, 2010). Using the same protocol for OC analysis, Tian et al. (2017) found that OC fraction desorbed at 140 and 280 ^∘C contributed ∼13 % and ∼20 %, respectively, to primary OC from residential coal combustion in China. The mvOC desorbed at 200 ^∘C falls between 140 and 280 ^∘C, thus the mvOC fraction in total OC should be higher than the faction of OC desorbed at 140 ^∘C in total OC and lower than the fraction of OC desorbed up to 280 ^∘C in total OC.

OC from different emission sources has different volatility and the atmospheric processing can also alter its volatility. Keller and Burtscher (2017) found that aging reduces the volatility of OC from biomass burning; i.e., the contribution of OC1 fraction to total OC decreases from 60 % to 25 % after aging. The photochemical processing of OA can lead to accumulation of carbon in the more refractory organic fraction and also larger organic compounds (Masalaite et al., 2017). The mvOC fraction in OC of our ambient samples is much smaller than that of the primary sources, suggesting that atmospheric aging of OC plays an important role on modifying the volatility of OC.

3.3 Nonfossil and fossil fractions of different carbon fractions

Figure 4a shows the fraction of nonfossil carbon in mvOC, OC and EC (f_nf(mvOC), f_nf(OC) and f_bb(EC), respectively). There are no considerable changes in f_nf(mvOC), f_nf(OC) and f_bb(EC) between polluted days (“H” samples) and clean days (“L” samples) within each site (Fig. 4a), despite the very different concentrations of carbonaceous aerosols (Fig. 3). However, f_nf(mvOC), f_nf(OC) and f_bb(EC) varied significantly among different sites: the lowest values are always found in Taiyuan and the highest are found in Chongqing. This implies that different pollution patterns exist in individual Chinese cities. The smallest f_nf(mvOC), f_nf(OC) and f_bb(EC) being in Taiyuan suggests that fossil sources are the main contributor to mvOC, OC and EC, whereas the largest values being in Chongqing shows that the nonfossil contributions to mvOC, OC and EC are evidently higher in Chongqing than in other sites.

Figure 4 (a) Fraction of nonfossil (f_nf) in mvOC, OC and EC. f_nf(mrOC) is calculated by the differences between OC and mvOC using isotope mass balance equation. (b) mvOC∕OC ratios for nonfossil, fossil and total OC.

Ranging from 22 % to 55 %, f_nf(mvOC) (41±11 %) is consistently smaller than f_nf(OC) (52±16 %, with a range of 29 %–77 %) and larger than f_bb(EC) (27±9 %, with a range of 10 %–42 %), despite their variations among the cities. The absolute difference in the nonfossil fractions between mvOC and OC is 13 % (7 %–25 %), similar to that between mvOC and EC (13 %, with a range of 4 %–25 %). Consistently smaller f_nf(mvOC) than f_nf(OC) suggests that mvOC is more fossil (less nonfossil) than the total OC. Liu et al. (2017) also found that F¹⁴C of OC desorbed at lower temperature (up to 200 ^∘C) in He was 0.389, smaller than F¹⁴C of total OC desorbed in the He phase of EUSAAR_2 protocol (F¹⁴C =0.495, up to 650 ^∘C) for a single test sample collected in winter in Xinjiang, China. Agrios et al. (2016) had the qualitatively similar finding that F¹⁴C of OC extracted in O₂ at 200 ^∘C was smaller than F¹⁴C of OC extracted in O₂ at higher temperatures for samples collected from an urban and rural site in Switzerland and from the Los Angeles Basin, USA. We can unambiguously conclude from ¹⁴C measurements of mvOC and OC that mvOC is more fossil than OC in six Chinese cites. The fraction of nonfossil carbon in mrOC (f_nf(mrOC)) is calculated by the differences between OC and mvOC using the isotope mass balance. Since mvOC is only a small fraction of OC, f_nf(mrOC) is very similar to f_nf(OC).

Primary OC from vehicular emissions is generally more volatile (i.e., less refractory) than OC from biomass burning, which is in line with the ambient results presented. For example, Grieshop et al. (2009c) constrained the volatility distribution of primary OA from a diesel engine and wood burning using measurements of TD coupled to aerosol mass spectrometer and found that OA from wood burning was less volatile than from diesel exhaust. Chen et al. (2007) measured OC∕EC from fresh biomass burning emissions and found that high-temperature OC fractions (desorbed from 450 to 550 ^∘C) are the major fraction of OC, in contrast to gasoline and diesel exhausts, where OC fractions desorbed at temperature lower than 250 ^∘C are more abundant (Watson et al., 1994; Chow et al., 2004). Zhu et al. (2010, 2014) also found that OC desorbed at 140 ^∘C is the major fraction of OC in fresh vehicle emissions in tunnel experiments. A more recent study of Masalaite et al. (2017) found that OC desorbed at 200 ^∘C in He has higher contributions from vehicular emissions than OC desorbed at higher temperatures (250–350 ^∘C) at three sites (urban, coastal and forest) in Lithuania, based on δ¹³C of OC desorbed at a different temperature step. However, the conclusions of this study remained qualitative because δ¹³C of organic carbon can be changed significantly by isotopic fractionation during atmospheric processing, whereas the ¹⁴C source apportionment used in our study is independent of atmospheric processing.

By comparing (mvOC∕OC)_fossil (i.e., the fraction of mvOC_fossil in OC_fossil) and (mvOC∕OC)_nf, we can get a measure of the volatility of OC from fossil and nonfossil sources, respectively:

$$\begin{array}{}\text{(13)}& \begin{array}{rl}{\displaystyle \frac{(\mathrm{mvOC}/\mathrm{OC}{)}_{\mathrm{fossil}}}{(\mathrm{mvOC}/\mathrm{OC}{)}_{\mathrm{nf}}}}& ={\displaystyle \frac{{\mathrm{mvOC}}_{\mathrm{fossil}}/{\mathrm{mvOC}}_{\mathrm{nf}}}{{\mathrm{OC}}_{\mathrm{fossil}}/{\mathrm{OC}}_{\mathrm{nf}}}}\\ & ={\displaystyle \frac{\frac{\mathrm{1}}{f_{\mathrm{nf}}\left(\mathrm{mvOC}\right)}-\mathrm{1}}{\frac{\mathrm{1}}{f_{\mathrm{nf}}\left(\mathrm{OC}\right)}-\mathrm{1}}}.\end{array}\end{array}$$

Asf_nf(mvOC) is smaller than f_nf(OC), this is equivalent to (mvOC∕OC)_fossil >(mvOC∕OC)_nf (Eq. 13). That is, OC from fossil sources contains a larger, more volatile fraction than OC from nonfossil sources. This is true for all studied cities, despite the variation in f_nf(mvOC) and f_nf(OC) in different cities. (mvOC∕OC)_fossil(14±6.6 %, with a range of 3.7 %–28 %) is consistently higher and more variable than (mvOC∕OC)_nf (7.5±2.9 %, with a range of 2.6 %–11.6 %; Fig. 4b). (mvOC∕OC)_fossil and (mvOC∕OC)_nf in general tracks the variation in mvOC∕OC, except that (mvOC∕OC)_fossil for Chongqing (28 % for Chongqing-H and 21 % for Chongqing-L) is much higher than the other sites (averaged 12 %). When comparing the mvOC∕OC within each city, it is found that (mvOC∕OC)_fossil changes more strongly between H and L samples than (mvOC∕OC)_nf, indicated by the bigger absolute differences of (mvOC∕OC)_fossil between H and L samples than that of (mvOC∕OC)_nf. Correlations between f_nf(mvOC) and f_nf(OC) and between f_nf(mvOC) and f_bb(EC) are examined. f_nf(mvOC) correlates more closely with f_nf(OC) (r²=0.89) than with f_bb(EC) (r²=0.62) (Fig. 5).

Figure 5 Correlation between f_nf(mvOC), f_nf(OC), f_nf(mvOC) and f_bb(EC) for Xi'an, Beijing, Taiyuan, Shanghai, Chongqing and Guangzhou during winter.

It is likely that some of mvOC are of secondary origin, since SOC is formed in the atmosphere and subsequently condenses onto the aerosol particles (Hallquist et al., 2009; Jimenez et al., 2009). It has been shown that SOC formed in chambers initially desorbed at relatively low temperatures (e.g., 200 ^∘C) (Holzinger et al., 2010; Salo et al., 2011; Meusinger et al., 2017; Gkatzelis et al., 2018). However, there is also evidence that the formation of highly oxidized low-volatility compounds can occur in SOA formation (Ehn et al., 2014). We thus compared concentrations and sources of SOC to those of mvOC. SOC concentrations and f_fossil(SOC) are estimated based on the ¹⁴C-apportioned OC and EC in combination with p values of 0–1 (Sect. 2.4). Where δ¹³C measurements were available for high-volume filter samples (Beijing, Shanghai and Guangzhou), we compare the SOC concentrations and f_fossil(SOC) derived from p values of 0–1 with p values constrained by δ¹³C. In these three cities, SOC concentrations and f_fossil(SOC) without constraints on p value (0–1) is very similar to that with constraints on p value using δ¹³C, only with relatively larger uncertainties (Fig. 6). This indicates that the unconstrained p values do not lead to significant bias in SOC concentrations and f_fossil(SOC) and the main gain in using δ¹³C is currently a decrease in uncertainty.

Figure 6 (a) Observed mvOC concentrations (shown as a green bar) and estimated SOC concentrations by ¹⁴C-apportioned OC and EC in combination with p values without (p=0–1) and with constraints by δ¹³C (¹³C-constrained p), shown as solid orange bars and patterned orange bars, respectively. Error bars indicate the uncertainties of mvOC and SOC concentrations calculated by Monte Carlo error propagation. (b) For Chongqing, lower p values of 0–0.5 are used to estimate SOC concentrations and are compared with p values of 0–1. (c) Fraction of fossil (f_fossil) in mvOC (blue square) and SOC. f_fossil of SOC is estimated using p values without constraints (p=0–1) are shown as red circles and with constraints by δ¹³C (¹³C-constrained p) as purple triangles, respectively. For Chongqing, lower p values of 0–0.5 are used (purple circles). Interquartile ranges (25th–75th percentile) of the median f_fossil(SOC) are shown as vertical bars, uncertainties of f_fossil(mvOC) are indicated but are too small to be visible.

When choosing p randomly from 0 to 1, the calculated median f_fossil(SOC) is smaller than 0 for samples in Chongqing, reflecting negative SOC_fossil values. Since SOC_fossil is calculated by subtracting POC_fossil from OC_fossil, this indicates POC_fossil is overestimated in Chongqing. Because the primary OC∕EC ratio is higher for coal (r_coal) than vehicle (r_vehicle) emissions, an overestimate of POC_fossil is the direct result of overestimated p values, i.e. a too high fraction of coal burning in EC_fossil (Eq. 11). Chongqing is in southern China, where there is no official heating using coal in winter and smaller contribution of coal combustion is expected than in northern China. Further, earlier studies found that vehicle emissions contributed more to EC than coal combustion in Chongqing (Chen et al., 2017). We thus tried to lower p values to 0–0.5. Lower p values lead to higher f_fossil(SOC) (Fig. 6c) and, to a lesser extent, higher SOC concentrations (Fig. 6b). For Chongqing, we took p from 0 to 0.5 as a conservative estimate. For other cities except Chongqing, we kept p values of 0–1 due to the very similar f_fossil(SOC) and SOC concentrations derived from p values with and without constraints by δ¹³C, especially for the two southern cities in China, e.g., Shanghai and Guangzhou.

f_fossil(mvOC) ranged from 45 % to 78 %, with an average of 59 %, whereas f_fossil(SOC) ranged from 10 % to 70 % and averaged 43 %. f_fossil(mvOC) is always larger than f_fossil(SOC) (Fig. 6c), showing that mvOC is more fossil than SOC. However, both SOC and mvOC have a larger contribution of fossil sources than the total OC. The smallest differences between f_fossil(SOC) and f_fossil(mvOC) are found in Beijing for both samples Beijing-H and Beijing-L, with an absolute difference of 3 % and 2 %, respectively, within the uncertainty of the measurements. Much higher differences are found in other cities, with the absolute differences ranging from 13 % to 35 %, well outside the measurement uncertainties.

Median values for SOC, SOC_fossil and SOC_nf are very close to their mean values (Fig. S6, Tables S7, S8). To compare with mvOC, here we use mean (±SD) as well for SOC, SOC_fossil and SOC_nf. The lowest SOC concentrations are found to be 2.95 µg m⁻³ for Guangzhou-L and the highest are found to be 35.9 µg m⁻³ for Beijing-H. SOC concentrations (15.4±9.0 µg m⁻³) are consistently higher than those of mvOC (3.3±2.2 µg m⁻³, Fig. 6a), showing that the contribution of SOC to total OC is considerably larger than that of mvOC. The major fraction of SOC has desorption temperatures above 200 ^∘C and falls into our operationally defined mrOC. SOC_fossil ranged from 0.90 to 25.3 µg m⁻³, with an average of 6.8±6.4 µg m⁻³, which is consistently higher than mvOC_fossil ranging from 0.5 to 5.4 µg m⁻³ with an average of 2.1±1.4 µg m⁻³ (Table S6). mvOC_nf is smaller than SOC_nf as well (Fig. S7), averaged mvOC_nf concentrations (1.5±1.1 µg m⁻³) are only one-fifth of that SOC_nf(8.5±5 µg m⁻³).

Figure 7 Correlation between (a) OC_nf and mvOC_nf (blue cross), (b) POC_bb and mvOC_nf (orange diamond), SOC_nf and mvOC_nf (green circle), (c) POC_fossil and mvOC_fossil(diamond in dark orange), and SOC_fossil and mvOC_fossil(circle in dark green) for Xi'an, Beijing, Taiyuan, Shanghai, Chongqing and Guangzhou during winter. In panel (c), the data points marked in red are from sample Beijing-H with the considerably higher SOC_fossil ∕ POC_fossil ratio (3.1) than the other samples (an average of 0.8) and are not included in the regression.

We examined the correlation between mvOC and SOC from nonfossil sources (mvOC_nf vs. SOC_nf) and from fossil sources (mvOC_fossil vs. SOC_fossil). Comparing absolute concentrations between different cities and pollution levels is expected to result in reasonably positive correlations for all species because it can be expected that all carbon fractions are higher in polluted cities than in clean cities. However, it is still interesting to compare the correlation coefficients relative to one another. mvOC_nf includes primary and secondary mvOC from nonfossil sources. Good correlation was found between mvOC_nf and SOC_nf(r²=0.87, bootstrapped 95 % confidence interval of [0.21, 0.94]; Fig. 7), which is not significantly different from the correlation between mvOC_nf and POC_bb (r²=0.71, bootstrapped 95 % confidence interval of [0.60, 0.99]). SOC_nf is estimated by subtracting POC_bb from OC_nf, where POC_bb is estimated by multiplying EC_bb with the OC∕EC ratio of fresh biomass burning plumes (r_bb) (Eq. 7). Due to the large uncertainty in r_bb, separation between POC_bb and SOC_nf is rather uncertain and the r² for a correlation of mvOC_nf and SOC_nf is therefore affected by r_bb. For example, with lower r_bb of 3 (mean), the correlation between SOC_nf and mvOC_nf is stronger (r²=0.92) than for our best estimate of r_bb(4) (r²=0.87). For r_bb of 5, it is slightly weaker (r²=0.76, Fig. S8). OC_nf, the sum of POC_bb and SOC_nf, is well constrained by F¹⁴C_OC and not affected by r_bb. mvOC_nf correlates strongly with OC_nf (r²=0.91), suggesting strong impacts on the variability of mvOC_nf from nonfossil emissions, including the secondary formation from nonfossil sources and primary biomass burning emissions.

SOC_fossil is produced by the oxidation of VOCs from fossil origins, including coal combustion and vehicle emissions, whereas mvOC_fossil can have both primary and secondary sources. No correlation is found between mvOC_fossil and SOC_fossil (r²=0.01, Fig. 7c), implying that in cities or pollution conditions characterized by high concentrations of fossil SOC, mvOC_fossil is not concurrently elevated. However, as demonstrated by the good correlation between POC_fossil and mvOC_fossil (r²=0.69; Fig. 7c) we conclude that primary fossil sources may play an important role on the variation in mvOC_fossil. Larger f_fossil(mvOC) than f_fossil(SOC) also indicates that fossil sources are associated more with mvOC than SOC. In Fig. 7c there are two outlier data points from sample Beijing-H with strongly elevated SOC_fossil and lower POC_fossil (shown in red circle and diamond, respectively). The highest SOC_fossil ∕ POC_fossil ratio (3.1) is found in Beijing-H, suggesting the largest fossil SOC formation during polluted days in Beijing compared to clean days in Beijing (Beijing-L with SOC_fossil ∕ POC_fossil ratio of 1.4) and samples in other cities (average SOC_fossil ∕ POC_fossil ratio of 0.7). This is also found in an early study that during winter 2013, SOC_fossil ∕ POC_fossil ratios in Beijing (with an average of 4.2) were much higher than those in Xi'an, Shanghai and Guangzhou (average SOC_fossil ∕ POC_fossil ratio of 1.3) (Zhang et al., 2015). This pollution event of Beijing-H is thus not included in the regression.

4 Discussion

In this study, samples are collected on single bare quartz (BQ) filters. mvOC may be influenced by adsorption of organic vapors to the quartz fibers (positive sampling artifact). In other words, the measured mvOC in the earlier discussion includes particulate mvOC and gaseous carbon adsorbed to the quartz filters that can also be desorbed from the filters up to 200 ^∘C. This artifact can be estimated by using a backup quartz filter behind the BQ filter (quartz behind quartz, QBQ). The QBQ filter is assumed to be exposed to the same organic vapor concentrations as the front BQ filter and the carbon measured on the QBQ filter is mainly due to adsorbed organic vapor and can serve as an estimate of the positive artifact (Chow et al., 2010; Zhu et al., 2012).

Figure 8 (a) The predicted particulate fraction of OC fractions (i.e., OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$, OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$, OC${}_{\mathrm{480}{}^{\circ }\mathrm{C}}$, OC${}_{\mathrm{580}{}^{\circ }\mathrm{C}}$, where the subscript refers to the desorption temperature of the IMPROVE_A protocol) as function of OC concentrations (Sect. S4). Hollow shapes are used to indicate samples collected by high-volume samplers (HiVol) at a flow rate of 1.0 m³ min⁻¹ and filled shapes by mini-volume samplers (MiniVol) at a flow rate of 5 L min⁻¹. Particulate fraction of mvOC (X_{p_mvOC}, blue cross) is estimated by interpolating the particulate fraction of OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$ and OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$. The blue solid line is the fitted line for the X_{p_mvOC}. (b) The X_{p_mvOC} for each sample corresponding to the fitted line in (a) for each sample; (c, d) fraction of nonfossil in mvOC collected on the single bare quartz (BQ) filter (f_nf(mvOC), blue line and solid circle). The blue area indicates the possible range of the nonfossil fraction of particulate carbon after correction for the positive artifact (f_nf(p_mvOC)), using an upper limit estimate for the positive artifact. The upper and lower limits of the range correspond to the extreme assumption that the F¹⁴C of the adsorbed organic vapor is 50 % lower and 50 % higher than measured F¹⁴C_(mvOC), respectively. f_nf(p_mvOC) is compared with f_nf(OC) (solid green triangle) in (c) and f_nf(SOC) (red circle) in (d). Interquartile range (25th–75th percentile) of the median f_fossil (SOC) is shown as vertical bars.

Without the QBQ backup quartz filters in this study, we estimate the particulate fraction of mvOC (X_{p_mvOC}) on the BQ filters using the OC fractions measured by the thermal–optical carbon analyzer in combination with an empirical volatility basis set (VBS) model, following the method of Ma et al. (2016) (Sect. S4). The VBS bins the organic aerosol (OA) compounds according to the effective saturation concentrations (C^*). For example, if C^* is equal to the concentrations of OA (C_OA), then 50 % of OA mass will be in the particle phase (Donahue et al., 2006). Ma et al. (2016) provide empirical C^* values for each desorption temperature corresponding to the measured particulate fraction on the filter, determined by the comparison of BQ and QBQ filters. They show that these empirical C^* values (10^1.6 for OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$, 10^1.1 for OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$, 10^0.6 for OC${}_{\mathrm{480}{}^{\circ }\mathrm{C}}$, 10^0.1 for OC${}_{\mathrm{580}{}^{\circ }\mathrm{C}}$, where the subscript refers to the desorption temperature of the IMPROVE_A protocol) are useful to predict particulate fractions for various OC loadings and sampling environments. Assuming OA∕OC mass ratio is 1.8 for urban environment (Xing et al., 2013), the particulate fraction of the collected OC for each of the four IMPROVE_A temperature steps i(X_pi) can be estimated as a function of OC concentration, as shown in Fig. 8a. The particulate fraction of OC increases considerably from OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$ to OC${}_{\mathrm{580}{}^{\circ }\mathrm{C}}$ and with increasing OC concentration. The desorption temperature for mvOC is 200 ^∘C, falling within the 140 ^∘C for OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$ and 280 ^∘C for OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$. We thus estimated the X_{p_mvOC} will fall within the X_p1 for OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$ and X_p2 for OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$. Particulate fraction of mvOC is estimated by interpolating the particulate fraction of OC${}_{\mathrm{140}{}^{\circ }\mathrm{C}}$ and OC${}_{\mathrm{280}{}^{\circ }\mathrm{C}}$ (Fig. 8a). X_{p_mvOC} ranged from 0.54 for sample winter-L with the lowest OC concentration of 15 µg m⁻³ to 0.79 for sample Beijing-H with the highest OC concentration of 50 µg m⁻³ (Fig. 8b). However, for our study we consider this as a lower estimate of particulate fraction of mvOC (i.e., higher limits of the positive artifact).

Our sampling duration (24 h) is much longer than that of Ma et al. (2016), which is 45 min or 1 h. With longer sampling duration, the loading on the quartz filter of both particulate OC and adsorbed organic vapors increase initially; however, the latter will reach a saturation point and level out. Thus, the percent bias by the adsorbed organics is smaller for longer sampling duration (Turpin et al., 2000). For example, Turpin et al. (1994) found that the loading of adsorbed organics reached around 1.7 µg cm⁻² and remained constant after 200 min of sampling time, but the aerosol loading increased linearly with the sampling time.

The empirically determined C^* is affected by the sampling conditions, e.g., temperature. $C_i^{*}$ is defined at 300 K, higher than the average temperature in winter in China. The lower temperature drives the semi-volatile OC into the particle phase and leads to higher particulate fraction of each OC fraction (Donahue et al., 2006).

The adsorption of organic vapors is also influenced by the face velocity (the volumetric flow rate divided by the exposed filter area). As the face velocity decreases, the adsorption of organic gaseous species increases (McDow and Huntzicker, 1990; Turpin et al., 2000; Viana et al., 2006). In this study, most of samples were collected using high-volume samplers (HiVol) with a flow rate of 1 m³ min⁻¹, including samples collected in Xi'an, Beijing, Guangzhou and Shanghai. Only samples in Taiyuan and Chongqing were collected using mini-volume samplers (MiniVol) at a flow rate of 5 L min⁻¹. The face velocities for the HiVol and MiniVol sampler were 42 and 7 cm s⁻¹, respectively. Turpin et al. (1994) found that the apparent OC concentrations on the front quartz filter were 22 % greater with a face velocity of 20 cm s⁻¹ than those with a face velocity of 40 cm s⁻¹. A more recent study of Kim et al. (2016) found that the positive artifacts for OC using sampler with the face velocity of 42 cm s⁻¹ is 7.2 %, which is smaller than 12.2 % using sampler of a face velocity of 24 cm s⁻¹ in spring and fall of 2014 in Korea. We do not take this into consideration and apply the same $C_i^{*}$ for HiVol and MiniVol samples, due to the lake of $C_i^{*}$ for different face velocities.

Due to the lack of information on how sampling duration, temperature and face velocity affect the C^* for each carbon fraction, we treat the particulate fraction in Fig. 8b as a rough estimate. But with long sampling duration, low temperature and high face velocity for most of samples as well as high OC concentrations in winter in this study, the particulate fraction of mvOC on the collected filters should be considerably higher than the estimate derived from VBS. For example, Cheng and He (2015) found that the positive artifacts in OC were lowest (∼10 %) in winter, much lower than in summer (∼23 %) in Beijing, China. With the same face velocity of ∼10 cm s⁻¹, the smaller positive artifacts in winter than that in summer can be attributed to higher OC concentrations (35 µg m⁻³ in winter vs. 12 µg m⁻³ in summer) and lower temperatures in winter. We consider the positive sampling artifact in this study is small and the collected mvOC on the bare quartz filter is dominated by the particulate mvOC.

Even though the positive artifact introduced by the adsorbed organic vapors is likely small, it is worth investigating its potential effect on the ¹⁴C results. The measured F¹⁴C_(mvOC) on the BQ filter can be explained by the F¹⁴C of particulate mvOC (F¹⁴C_(p_mvOC)) and adsorbed organic vapors (F¹⁴C_(g_mvOC)). The F¹⁴C_(p_mvOC) can be calculated as follows:

$$\begin{array}{}\text{(14)}& \begin{array}{rl}& {\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{p}\mathrm{\_}\mathrm{mvOC}\right)}\\ & ={\displaystyle \frac{{\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{mvOC}\right)}-(\mathrm{1}-X_{\mathrm{p}\mathrm{\_}\mathrm{mvOC}})\times {\mathrm{F}}^{\mathrm{14}}{\mathrm{C}}_{\left(\mathrm{g}{\mathrm{\_}}_{\mathrm{mvOC}}\right)}}{X_{\mathrm{p}\mathrm{\_}\mathrm{mvOC}}}},\end{array}\end{array}$$

where X_{p_mvOC} is the estimated particulate fraction of mvOC, as shown in Fig. 8b, which is a much lower limit, as discussed earlier. (1–X_{p_mvOC}) is thus an estimate of positive artifact. How close the measured F¹⁴C_(mvOC) is to the F¹⁴C_{(p_mvOC)} depends on the F¹⁴C_{(g_mvOC)}.

Therefore, we conducted a sensitivity study based on the following assumption: the positive artifact is dominated by semi-volatile gases that actively partition between gas and particle phase. As the base case, we assume that F¹⁴C_{(g_mvOC)}= F¹⁴C_(mvOC) and consider extreme cases of F¹⁴C${}_{\left(\mathrm{g}\mathrm{\_}\mathrm{mvOC}\right)}=\mathrm{1.5}\times$ F¹⁴C_(mvOC) and F¹⁴C${}_{\left(\mathrm{g}\mathrm{\_}\mathrm{mvOC}\right)}=\mathrm{0.5}\times$ F¹⁴C_(mvOC), i.e. a variation of ±50 %. This allows F¹⁴C values for the adsorbed gas phase to be as low as for EC and higher than for OC in most cases. When the lower limit of F¹⁴C_{(g_mvOC)} is used (50 % lower than F¹⁴C_(mvOC)), the f_nf(p_mvOC), calculated by dividing F¹⁴C_{(p_mvOC)} with F¹⁴C_nf, is 0.04 to 0.19 (absolute value) higher than f_nf(mvOC) (Fig. 8) but still lower than f_nf(OC) and f_nf(SOC) for most of samples, at least for highly polluted (H) samples with a higher particulate fraction of mvOC. Consequently, our conclusion that mvOC is more fossil than OC and SOC is still valid for most of samples, especially since this sensitivity study assumes (i) an upper limit of the positive artifact as well as (ii) a very large maximum difference between the F¹⁴C of the adsorbed gas phase and the semi-volatile particle phase. The upper limit of f_nf(p_mvOC) corresponds to a semi-volatile gas phase with F¹⁴C values in the range of EC, which is probably not realistic.

5 Conclusions

Radiocarbon measurements are conducted for EC and OC of PM_2.5 samples collected in six Chinese megacities during winter. In addition, the sources of mvOC are measured unambiguously using ¹⁴C. mvOC is isolated by desorbing organics in He at 200 ^∘C using our custom-made aerosol combustion system with a reproducibility of ∼5 %. This new isolation method is evaluated by comparing with the OC1 fraction of the EUSAAR_2 protocol using a thermal–optical analyzer. The mvOC amount agrees very well with the OC1, despite the different heating rate and length of the temperature step, showing the representativeness of the isolated mvOC. The measured F¹⁴C_(mvOC) for a combined sample (0.529±0.007), which consists of three individual daily filters, is quite similar to the weighted average of the individual filters. This further validates our isolation method for mvOC that the F¹⁴C_(mvOC) is representative of the combined samples. With long sampling duration, low temperature and high face velocities as well as high OC concentrations in winter in this study, the positive artifacts to mvOC should be relatively small and play a limited role in the F¹⁴C_(mvOC).

mvOC∕OC ranges from 3 % to 15 %, with an average of 10.5±3.3 %. The ambient samples in this study have a smaller mvOC∕OC than ambient samples from other places and have a much smaller mvOC fraction than fresh source samples, likely due to the effects from atmospheric aging processes. mvOC∕OC from fossil sources averages 14±6.6 %, consistently higher and more variable than that from nonfossil sources (7.5±2.9 %). That is, OC from fossil sources contains a larger more volatile fraction than that from nonfossil sources.

The nonfossil contribution to mvOC, OC and EC (f_nf(mvOC), f_nf(OC), f_bb(EC)) varies significantly among samples in different sites: the lowest values are always found in Taiyuan and the highest values are found in Chongqing. This implies that different pollution patterns exist in individual Chinese cities. Despite the variations in f_nf(mvOC), f_nf(OC) and f_bb(EC) in different cities, mvOC is more fossil than OC and less fossil than EC in all locations. The fossil contribution to mvOC, OC and EC are, on average, 59±11 %, 48±16 % and 73±9 %, respectively. Higher fossil contribution to mvOC than to OC is consistent with conclusions from δ¹³C of OC desorbed at different temperatures for ambient samples in other locations. f_nf(mvOC) correlates better with f_nf(OC) (r²=0.89) than f_bb(EC) (r²=0.62).

SOC concentrations and sources are modeled based on the ¹⁴C-apportioned OC and EC. SOC concentrations (15.4±9.0 µg m⁻³) are consistently higher than those of mvOC (3.3±2.2 µg m⁻³), showing that the contribution of SOC to total OC is considerably larger than that of mvOC. A major fraction of SOC has desorption temperatures above 200 ^∘C and falls into our operationally defined mrOC. mvOC is more fossil than SOC. The relative contribution of fossil sources to mvOC is 59 % on average (range of 45 %–78 %), which is higher than that to SOC (43 %, with a range of 10 %–70 %).

Good correlation is found between mvOC_nf and OC_nf(r²= 0.91). This indicates strong impacts on the variability of mvOC_nf from nonfossil emissions, including the secondary formation of OC from nonfossil sources and primary biomass burning. The variation in mvOC_fossil is mainly explained by a good correlation with POC_fossil (r²=0.69) and no correlation is found between mvOC_fossil and SOC_fossil (r²=0.01). This indicates that primary mvOC from fossil sources may play an important role in the variation in mvOC_fossil.