Sources of organic aerosols in Europe: a modeling study using CAMx with modified volatility basis set scheme

Abstract

Source apportionment of organic aerosols (OAs) is of great importance to better understand the health impact and climate effects of particulate matter air pollution. Air quality models are used as potential tools to identify OA components and sources at high spatial and temporal resolution; however, they generally underestimate OA concentrations, and comparisons of their outputs with an extended set of measurements are still rare due to the lack of long-term experimental data. In this study, we addressed such challenges at the European level. Using the regional Comprehensive Air Quality Model with Extensions (CAMx) and a volatility basis set (VBS) scheme which was optimized based on recent chamber experiments with wood burning and diesel vehicle emissions, and which contains more source-specific sets compared to previous studies, we calculated the contribution of OA components and defined their sources over a whole-year period (2011). We modeled separately the primary and secondary OA contributions from old and new diesel and gasoline vehicles, biomass burning (mostly residential wood burning and agricultural waste burning excluding wildfires), other anthropogenic sources (mainly shipping, industry and energy production) and biogenic sources. An important feature of this study is that we evaluated the model results with measurements over a longer period than in previous studies, which strengthens our confidence in our modeled source apportionment results. Comparison against positive matrix factorization (PMF) analyses of aerosol mass spectrometric measurements at nine European sites suggested that the modified VBS scheme improved the model performance for total OA as well as the OA components, including hydrocarbon-like (HOA), biomass burning (BBOA) and oxygenated components (OOA). By using the modified VBS scheme, the mean bias of OOA was reduced from $-\mathrm{1.3}$ to $-\mathrm{0.4}$ $\mathrm{\mu }$g m${}^{-\mathrm{3}}$ corresponding to a reduction of mean fractional bias from $-\mathrm{45}$ % to$-\mathrm{20}$ %. The winter OOA simulation, which was largely underestimated in previous studies, was improved by 29 % to 42 % among the evaluated sites compared to the default parameterization. Wood burning was the dominant OA source in winter (61 %), while biogenic emissions contributed$\sim$ 55 % to OA during summer in Europe on average. In both seasons, other anthropogenic sources comprised the second largest component (9 % in winter and 19 % in summer as domain average), while the average contributions of diesel and gasoline vehicles were rather small ($\sim$ 5 %) except for the metropolitan areas where the highest contribution reached 31 %. The results indicate the need to improve the emission inventory to include currently missing and highly uncertain local emissions, as well as further improvement of VBS parameterization for winter biomass burning. Although this study focused on Europe, it can be applied in any other part of the globe. This study highlights the ability of long-term measurements and source apportionment modeling to validate and improve emission inventories, and identify sources not yet properly included in existing inventories.

1 Introduction

Pollution by atmospheric fine particulate matter (PM) exerts significant impacts on human health (Ciarelli et al., 2019; Cohen et al., 2017; Lelieveld et al., 2015; Tuet et al., 2017) and climate (Kanakidou et al., 2005), where organic aerosols (OAs) contribute 20 %–90 % (Jimenez et al., 2009; Kanakidou et al., 2005). Unlike the single-component pollutants such as ozone and sulfur dioxide, organic aerosols are composed of numerous compounds from different sources with distinct physical, chemical and toxicological properties (Hallquist et al., 2009; Shrivastava et al., 2017). The situation is even more complicated for secondary organic aerosol (SOA), which is generated from the oxidation of organic gases emitted from a wide range of biogenic and anthropogenic sources and accounts for a dominant fraction of OA (Hodzic et al., 2016; Srivastava et al., 2018). Understanding and identifying the OA sources is therefore important for understanding the implication of aerosols for health and climate and establishing effective mitigation policies.

Large efforts have been devoted to determine OA sources at different scales, mostly based on receptor modeling. Positive matrix factorization (PMF) analysis is often applied to aerosol mass spectrometer data to classify the measured organic mass spectra into different factors. Commonly retrieved factors include hydrocarbon-like OA (HOA) from traffic emissions, biomass burning (BBOA), cooking (COA) and oxygenated components (OOA) (Crippa et al., 2014). Recent studies with long-term offline aerosol mass spectrometer (AMS) measurements were able to further split the OOA component into biogenic SOA and anthropogenic SOA according to their seasonal variability (Daellenbach et al., 2016, 2017; Vlachou et al., 2018).

Air quality models (AQMs) provide another approach to quantify OA sources, with great advantages at high temporal and spatial resolution. Various tools have been developed and implemented in AQMs for regional-scale source apportionment of particulate matter, such as the Particulate Source Apportionment Technology (PSAT) for CAMx (Comprehensive Air quality Model with Extensions) (Koo et al., 2009), tagged species source apportionment (TSSA) (Wang et al., 2009) and the integrated source apportionment method (ISAM) (Kwok et al., 2013) for CMAQ (Community Multiscale Air Quality). However, the performance of these tools for OA source apportionment are always limited by substantial underestimation of SOA by the traditional AQMs (Hodzic et al., 2010; Tsigaridis et al., 2014). One of the most important reasons for the underestimation is the potentially high but unaccounted for contribution of non-traditional vapors. To take these vapors into consideration, the volatility basis set (VBS) scheme has been developed and implemented in several air quality models such as CAMx (Ciarelli et al., 2016, 2017a; Koo et al., 2014), CMAQ (Jathar et al., 2017; Koo et al., 2014; Woody et al., 2016), PMCAMx (Lane et al., 2008; Tsimpidi et al., 2010), CHIMERE (Cholakian et al., 2018; Zhang et al., 2013, 2015), EMEP (Bergström et al., 2012) and WRF-Chem (Ahmadov et al., 2012; Shrivastava et al., 2011, 2013, 2019). The VBS scheme classifies the first-generation oxidation products of vapors according to their volatility. The evolution of these products with aging through functionalization and fragmentation can also be presented by shifting the volatility of compounds (Donahue et al., 2006). It is widely reported that implementing VBS schemes improves the model performance for SOA (Ciarelli et al., 2016; Fountoukis et al., 2014; Koo et al., 2014; Robinson et al., 2007; Tsimpidi et al., 2010). It is however still a challenge to use the VBS scheme together with the source apportionment tools in AQMs. To our knowledge, the VBS scheme in CAMx is currently not enabled to be used with the source apportionment tool PSAT (Ramboll, 2018).

A number of studies have used air quality models with VBS to model OA components or sources on regional scales. Bergström et al. (2012) tested the EMEP model with different VBS setups to calculate the contributions of biogenic and anthropogenic SOA, residential wood combustion and wild fire emissions to OA during 2002–2007 over Europe, and evaluated the source apportionment results with observations at four sites with weekly or daily filter measurements of elemental and organic carbon (EC/OC) and one site with hourly AMS measurements. Yttri et al. (2019) used similar EMEP-VBS including reactions of semi-volatile organic compounds (SVOCs) and intermediate volatility organic compounds (IVOCs) to model OA sources at nine rural sites in Europe, and compared the model output with source apportionment of measured EC/OC by chemical and ${}^{\mathrm{14}}$C tracers. Both studies found that residential wood burning was largely underestimated, and required further improvement.

Based on recent chamber experimental studies on wood burning (Bruns et al., 2016), Ciarelli et al. (2017a, b) parameterized a hybrid volatility basis set which included a new set for the oxidation from SVOCs and implemented it in the air quality model CAMx to simulate winter OA sources in Europe. The new parameterization significantly improved the model performance for SOA (Ciarelli et al., 2017a). However, as the biomass burning and biogenic precursors are merged into the same set in the original VBS scheme of CAMx, the biogenic SOA is implicitly taken into account to react with OH in the gas phase, which could lead to overestimated SOA in summer when biogenic emissions are high (Ramboll Environ, 2016).

Skyllakou et al. (2017) extended the PSAT tool in the regional model PMCAMx with VBS to quantify the OA sources in Europe, with a major focus on the source–receptor relationship but less attention on the evaluation of OA predictions against observations. More modeling studies using AQMs with VBS to simulate OA sources are available at the city scale. Woody et al. (2016) used CMAQ-VBS to model the primary organic aerosol (POA) from meat cooking, gasoline and diesel vehicles, biomass burning and other sources in California in May–June in 2010, while the SOA was characterized by formation pathways (including first products of anthropogenic and biogenic VOCs, first products of IVOCs, aging reactions of secondary SVOCs, and anthropogenic and biogenic VOCs) instead of sources. Jathar et al. (2017) simulated the sources of POA and SOA in southern California with an updated CMAQ-VBS with special focus on gasoline and diesel vehicles, and predicted that gasoline vehicles contribute $\sim$ 35 % of the inland OA which is $\sim$ 13 times more than diesel sources. However, the contribution of gasoline and diesel vehicle emissions to the total OA in Europe, where the vehicle types have high spatial variations, remains unclear. Results from recent aging studies of diesel and gasoline exhaust (Gentner et al., 2016; Platt et al., 2017; Zhao et al., 2017) have not yet been implemented into models either. Despite all the progress, the parameterization of volatility basis sets in AQMs requires further improvement based on the advance of chamber experimental data, especially for wood burning which was highly underestimated; most modeling studies focus more on the SOA components differentiated by the formation pathways (i.e., from IVOCs, SVOCs), while the sources of SOA are equally important for emission reduction strategies; modeled source apportionment results need further evaluation using measurement data with higher time resolution and longer periods.

In this study, we (i) modified the regional air quality model CAMx with the VBS scheme to differentiate primary and secondary organic aerosols from various sources including gasoline vehicles, old and new diesel vehicles, biomass burning (excluding wild fires), other anthropogenic sources and biogenic sources, (ii) updated the parameterization of the VBS based on recent smog chamber experimental data for residential wood burning and diesel vehicles, with separate sets and parameterization for aging of secondary condensable gases from biomass burning and biogenic sources, (iii) conducted a whole-year simulation in 2011 to calculate the OA concentrations from different primary and secondary sources in Europe and (iv) evaluated the model performance on OA source apportionment by comparing the model results with the PMF analyses of hourly observational data covering nine Aerodyne aerosol chemical speciation monitor (ACSM)/aerosol mass spectrometer (AMS) stations in Europe with measuring periods ranging from 1 month to 1 year.

2 Method

2.1 Air quality model CAMx

The air quality model CAMx version 6.3 (Ramboll Environ, 2016) was used to simulate OA for the full year 2011. The model domain covers Europe (15${}^{\circ }$ W–35${}^{\circ }$ E, 35–70${}^{\circ }$ N), with a spatial resolution of $\mathrm{0.25}{}^{\circ }\times \mathrm{0.125}{}^{\circ }$ and 14 terrain-following vertical layers ranging from $\sim$ 20 m above ground level (first layer) going up to 460 hPa. The Carbon Bond 6 Revision 2 (CB6r2) gas-phase mechanism (Hildebrandt Ruiz and Yarwood, 2013) was selected in this study. The ISORROPIA thermodynamic model (Nenes et al., 1998) was used to simulate gas–aerosol partitioning of inorganic aerosols. For organic aerosols, a modified VBS module by the Paul Scherrer Institute (PSI-VBS) based on the 1.5-dimensional (1.5-D) VBS scheme (Koo et al., 2014) was used to model the formation and evolution of OA. The meteorological parameters were produced with the Weather Research and Forecasting model (WRF, version 3.7.1; Skamarock et al., 2008), based on the 6 h European Centre for Medium-range Weather Forecasts (ECMWF) reanalysis global data with a resolution of $\mathrm{0.72}{}^{\circ }\times \mathrm{0.72}{}^{\circ }$ (Dee et al., 2011). The initial and boundary conditions for the concentrations of chemical species were obtained from the global model MOZART-4/GEOS-5 (Horowitz et al., 2003). The Total Ozone Mapping Spectrometer (TOMS) data by the National 25 Aeronautics and Space Administration (NASA: https://acd-ext.gsfc.nasa.gov/anonftp/toms/omi/data/Level3e/ozone/, last access: 6 December 2019) was adopted for the input of ozone column densities, and the photolysis rates were calculated by the Tropospheric Ultraviolet and Visible (TUV) Radiation Model version 4.8 (NCAR, 2011). A more detailed description of the inputs of meteorology, photolysis and initial and boundary conditions is found in Jiang et al. (2019). The simulation period was between 1 January and 31 December 2011 with the first 2 weeks being used as spin-up. We performed the simulations using both the standard (BASE) and modified (NEW) parameterization in the VBS module.

2.2 VBS parameterization

2.2.1 Extended volatility basis sets

The 1.5-D VBS framework in CAMx is based the one-dimensional (1-D) VBS, in which the organic species are grouped only by their volatility (Donahue et al., 2006). The 1-D VBS was later extended to a second dimension (2-D) to include the oxidation state – specifically O : C ratio (Donahue et al., 2011). In order to reduce the high computational burden of the 2-D VBS when implemented in chemical transport models, the 1.5-D VBS was developed, which combines the 1-D VBS and the multiple reaction trajectories defined in the 2-D VBS space; it can therefore account for changes in both volatility and oxidation state (Koo et al., 2014). The default VBS scheme of CAMx version 6.3 includes five basis sets to describe the oxidation process of OA: three sets for freshly emitted OA from biomass burning (PFP, Particle Fire Primary), cooking (PCP, Particle Cooking Primary) and other anthropogenic (PAP, Particle Anthropogenic Primary), and two basis sets for chemically aged oxygenated OA from anthropogenic (PAS, Particle Anthropogenic Secondary) and biogenic (PBS, Particle Biogenic Secondary) emissions. For the PAS set, the OA generated from gasoline vehicles (GVs), diesel vehicles (DVs) and other anthropogenic activities (OP) are parameterized according to volatility distributions, yields and vaporization enthalpies reported in the literature, and then merged to PAS. A similar treatment is adopted for secondary OA from biomass burning, cooking and biogenic sources, which are merged together as PBS.

The VBS scheme has been modified in previous studies to improve the performance of air quality models. For example, Shrivastava et al. (2013, 2015) treated SOA as a non-absorbing semisolid with low “effective volatility” and added the fragmentation reactions. Using this method in the regional model CHIMERE, it was found that fragmentation could effectively reduce the SOA formation when further aging of biogenic SOA was allowed, leading to a better agreement with observations (Cholakian et al., 2018). Instead of a major modification of the chemical mechanism, our study aims at modifying the 1.5-D VBS framework of CAMx to enable source apportionment of OA in Europe. As the first step to separate the modeled OA components, the standard 5 basis sets were split into 11 basis sets including primary and secondary OA from five sources, i.e., new diesel vehicles (DNs, with and after Euro 4) equipped with diesel particle filter (DPF), old diesel vehicles (DOs, before Euro 4) without DPF, GV, biomass burning (BB) and other anthropogenic sources (OthA), as well as SOA from biogenic sources (BIO). The schematic diagram of the VBS with the modified basis sets (it will be referred to as PSI-VBS thereafter) is shown in Fig. 1. Due to lack of emission data, OA from cooking emissions was excluded in this study. Instead of merging OA from different sources as done in the default CAMx-VBS, we added the species in the 11 basis sets to the species list of model output to distinguish the OA sources. The new PSI-VBS scheme was then tested with the standard parameterization of CAMx version 6.3 (BASE) and a modified parameterization (NEW) based on recent experimental findings.

2.2.2 Standard parameterization: BASE

The BASE parameterization contains the default parameters of the CAMx (version 6.3) VBS module. The same parameterization (volatility distribution, yields, reaction rates of VOC/IVOC precursors and primary/secondary condensable gases) of the standard set PAS was used for the secondary OA from DO, DN, GV, OthA, as well as the PAP for the primary OA from DO, DN, GV, OthA. The primary and secondary BB and BIO followed the same parameterization as PFP and PBS, respectively. The aging of biogenic SOA was disabled in the standard parameterization as it led to significant overestimation of OA in rural areas in previous modeling studies (Ramboll Environ, 2016). As the biomass burning and biogenic SOA belong to the same PBS set and use the same parameterization in the default version, the aging of biomass burning SOA was also disabled in BASE.

2.2.3 Modified parameterization: NEW

In the NEW case, we used a modified parameterization based on smog chamber experimental studies. The major changes were made for diesel vehicles and biomass burning. Diesel vehicles constitute nearly half of the total passenger car registrations in Europe (ACEA, 2017). Diesel vehicle emissions were traditionally considered more efficient in generating SOA than gasoline exhaust (Gentner et al., 2012). However, the vehicles equipped with a diesel particle filter (DPF) were found to effectively reduce SOA production (Gentner et al., 2016; Gordon et al., 2013; Platt et al., 2017). In Europe, DPFs have been implemented in some diesel vehicles since Euro 4 in 2005 and have been required for all diesel vehicles since Euro 5 in 2009. Due to a large share of diesel vehicles equipped with DPF, we set the SOA yield for the basis set DN (diesel vehicle – new) to zero. Residential biomass burning is among the largest winter OA sources in Europe (Crippa et al., 2014; Lanz et al., 2010). Models generally underestimate OA from biomass burning (Ciarelli et al., 2017a; Hodzic et al., 2010; Jathar et al., 2014; Shrivastava et al., 2015). While the standard VBS of CAMx disables the aging of SOA for the basis set PBS (biomass burning and biogenic sources) to avoid overestimation of biogenic SOA, the separated sets for biomass burning (BB) and biogenic (BIO) allow us to implement individual parameterization schemes. Therefore, we kept the default parameterization (without aging of SOA) for BIO sources as a compromise for the lack of gas-phase fragmentation, and enabled the oxidation of secondary gases from biomass burning (see BB in Fig. 1) with a reaction rate of $\mathrm{4}\times {\mathrm{10}}^{-\mathrm{11}}$ cm${}^{\mathrm{3}}$ molec.${}^{-\mathrm{1}}$ s${}^{-\mathrm{1}}$ according to previous studies (Ciarelli et al., 2017a, b; Denier van der Gon et al., 2015; Fountoukis et al., 2014; Murphy and Pandis, 2009; Theodoritsi and Pandis, 2019). For other basis sets, the default parameters of CAMx v6.3 were used.

2.3 Emissions

The anthropogenic emissions were based on the high-resolution European emission inventory TNO-MACC (Monitoring Atmospheric Composition and Climate)-III, which is an extension of the TNO-MACC-II (Kuenen et al., 2014). The annual emissions of non-methane volatile organic compounds (NMVOCs), ${\mathrm{SO}}_{\mathrm{2}}$, ${\mathrm{NO}}_x$, CO, ${\mathrm{NH}}_{\mathrm{3}}$, PM${}_{\mathrm{10}}$ and PM${}_{\mathrm{2.5}}$ were hourly distributed using the TNO temporal variation profiles. The particulate matter emissions were split into POA, elemental carbon (EC), sodium (${\mathrm{Na}}^{+}$), particle sulfate (${\mathrm{PSO}}_{\mathrm{4}}^{\mathrm{2}-}$), and other primary particles in the fine (FPRM) and coarse (CPRM) size fractions according to the TNO PM-splitting profile. The NMVOC speciation was performed using the approach of Passant (2002) to generate emissions of 20 NMVOC species including toluene, xylene and benzene. The TNO-MACC-III emission inventory includes anthropogenic emissions from 10 SNAP (Selected Nomenclature for Air Pollution) source categories: energy industries (SNAP 1), residential and other non-industrial combustion (SNAP 2), industry combustion and processes (sum of SNAP 3 and SNAP 4), extraction and distribution of fossil fuels (SNAP 5), product use (SNAP 6), road transport (SNAP 7), non-road transport and other mobile sources (SNAP 8), waste treatment (SNAP 9) and agriculture (SNAP 10). Additional SNAP 7 emissions were provided by TNO, which include more detailed classification of gasoline and diesel vehicles (before Euro 4 and for Euro 4 and higher emission standards), liquefied petroleum gas vehicles, vehicles evaporation and brake wear. To match the requirement of the modified VBS module, we reclassified the diesel vehicles before and after (including) Euro 4 as OLD and NEW diesel vehicles (DO and DN), respectively. All the other emissions from road traffic such as vehicular evaporation and brake wear were attributed to the “other anthropogenic (OthA)” due to the lack of information to conduct a specific parameterization. The POA from residential combustion (SNAP 2) and agriculture (SNAP 10, mainly from on-field burning of stubble and straw) were summed up to represent POA from biomass burning as justified in Ciarelli et al. (2017a). The wildfires are not included in the biomass burning.

The intermediate-volatility and semi-volatile organic compounds (IVOCs and SVOCs) are considered as important precursors of SOA (Jathar et al., 2011, 2014), but they are generally absent in current emission inventories. Here we estimated the IVOC emissions from different sources based on literature (Table S1). The IVOCs from gasoline and diesel vehicles were calculated as 25 % and 20 % of NMVOC emissions from gasoline and diesel vehicles, respectively, according to the gas-phase carbon-balance analysis of Jathar et al. (2014). IVOC emissions from biomass burning were estimated as 4.5 times POA emissions based on Ciarelli et al. (2017a). For other anthropogenic sources, the IVOC emissions were calculated as 1.5 times POA as proposed by Robinson et al. (2007), which has been widely adopted by modeling studies (Jathar et al., 2017; Woody et al., 2016). The VBS scheme assumes that a certain fraction of POA can evaporate and be distributed in the semi-volatile range (saturation concentration between 0.1 and 1000 $\mathrm{\mu }$g m${}^{-\mathrm{3}}$), while most of the emission inventories only include POA in the particle phase. According to the partitioning theory (Donahue et al., 2006), the ratio between gas and particle phase in the semi-volatile range is roughly 3, and therefore many modeling studies increase the POA emissions by a factor of 3 to compensate for the missing SVOCs (Ciarelli et al., 2017a; Shrivastava et al., 2011; Tsimpidi et al., 2010). This approach agrees well with the emission study in Europe which shows that the revised residential wood combustion emissions accounting for the semi-volatile components are higher than those in the previous inventory by a factor of 2–3 on average (Denier van der Gon et al., 2015). However, Denier van der Gon et al. (2015) also pointed out that this factor presents substantial inter-country variability due to different combustion types, fuel parameters and operation conditions, indicating a potential over- or underestimation for a specific area by using the factor of 3 in the whole domain. To investigate the role of SVOC, we adopted the approach to increase POA emissions by a factor of 3 in NEW, while keeping the POA unchanged in BASE.

Biogenic emissions (isoprene, monoterpenes, sesquiterpenes, soil NO) were estimated by the PSI model developed at the Laboratory of Atmospheric Chemistry at the Paul Scherrer Institute (Andreani-Aksoyoglu and Keller, 1995) and further improved by Oderbolz et al. (2013) and Jiang et al. (2019). A comparison study with the widely used biogenic Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2.1 indicated that the PSI model produces higher monoterpene emissions in Europe than MEGAN, and leads to a better performance of CAMx for OA (Jiang et al., 2019).

2.4 Model evaluation

As the performance of CAMx strongly depends on the quality of the meteorological inputs, we first evaluated the meteorological parameters (surface temperature, wind direction, wind speed, precipitation) modeled by WRF as described in Jiang et al. (2019) using observations obtained from the UK Met Office Integrated Data Archive System (MIDAS) Land Surface Stations database (Meteorological Office, 2013). It covers $\sim$ 1000 stations in Europe and provides observations at 3 h time resolution. The model performance criteria used for meteorological parameters (Emery, 2001) are shown in Table S2. The general model performance for the main gas-phase species (i.e., ${\mathrm{O}}_{\mathrm{3}}$, ${\mathrm{SO}}_{\mathrm{2}}$, ${\mathrm{NO}}_x$, CO) and fine particulate matter PM${}_{\mathrm{2.5}}$ was also evaluated using the hourly measurements extracted from the European Environment Agency database, AirBase v7 (Mol and Leeuw, 2005). The statistical analysis was conducted by means of mean bias (MB), mean gross error (MGE), root-mean-square error (RMSE), mean fractional bias (MFB), mean fractional error (MFE), index of agreement (IOA) and correlation coefficient ($r$) between measured and modeled results. For ozone, only measurements at the background-rural stations were used in the model evaluation to reduce the possible uncertainties caused by the model resolution.

Table 1

Coordinates, observation periods and PMF related information of stations.

[Figure omitted. See PDF]

In order to further investigate the reasons for different model performance at different sites, the temporal variations of modeled OA for CAMx-BASE and CAMx-NEW were compared with the measurements (Figs. 2, S2). We selected four stations to represent different site types, i.e., Zurich as an urban (background) station, Paris SIRTA as a suburban station, SPC as a rural station and Montsec as a remote station. Among the four sites, measurements in Zurich cover almost the whole year of 2011, while measurements at other stations took place mostly in autumn and winter periods. The modeled OA concentrations in Zurich matched the measurements quite well except in winter, when the model underestimated OA with the BASE parameterization (Fig. 2a). In addition, there was a significant underestimation of temperature, which was expected to increase the modeled OA via enhanced condensation, indicating that the predicted OA concentrations would have been even lower if the modeled temperature had been higher. Including the reactions of SVOCs and further aging of secondary condensable gases for biomass burning sources in CAMx-NEW, the model performance in winter was effectively improved by increasing the OA concentrations by $\sim$ 100 % compared to CAMx-BASE during February and March. Nevertheless, CAMx-NEW led to an overestimation of OA in September and November, which influenced the overall performance of NEW. The two peaks in fall were largely due to the SOA from biogenic emissions, according to our previous work comparing the effects of two different biogenic emission models on OA (Jiang et al., 2019). Paris SIRTA is among the sites with best improvement of the predicted OA. In spite of the underestimation, the modeled OA reproduced the temporal evolution of the major peaks during the measurement period, and the OA by CAMx-NEW is about $\sim$ 1.4 times higher than CAMx-BASE (Fig. 2b) The highest peak during 19–24 November was related to an air mass from the south–southeast (SSE) with a recirculation over northern France, which is dominated by a wood burning source (Petit et al., 2015).

Figure 3

Seasonal variation of PMF analysis results and modeled OA components (with both BASE and NEW parameterizations). The average of modeled OA components was calculated based on the data during the same periods with measurements at each stations as shown in Table 1. ZRH: Zurich, SMEARII: SMEAR II Hyytiälä, MSA: Montsec, BLQ: Bologna, SPC: San Pietro Capofiume, SIRTA: Paris SIRTA, FKL: Finokalia, MRS: Marseille. Spring: March–April–May, summer: June–July–August, autumn: September–October–November, winter: December–January (2011)–February (2011). See Table S5 for seasonal-based statistical results.

[Figure omitted. See PDF]

At the rural site, SPC, a distinction can be seen before and after 23 November (Fig. 2c). Both CAMx-BASE and CAMx-NEW overestimated the OA before 23 November. Located in the Po Valley, OA at SPC is significantly influenced by fog scavenging processes during autumn when the high relative humidity and low temperature lead to frequent fog events (Gilardoni et al., 2014). Due to the coarse resolution and special geographical location of SPC, the meteorological model failed to reproduce the extremely high relative humidity during the fog events, leading to an overestimation of OA as particle activation into fog droplets, and eventually wet scavenging, are not adequately captured by the model. A better agreement with the observed OA was found for CAMx-NEW than CAMx-BASE after 23 November, when fog events were shorter.

For comparison with the observations at the high-altitude ($\sim$ 1570 m above sea level) station Montsec, we used the model results from the fifth vertical layer. Generally, air quality model results are poorly reproduced at high-altitude sites due to domain resolution. Although the NEW parameterization increased the OA by 35 %, the OA concentrations were still largely underestimated during July to September (Fig. 2d). The temporal variation of OA in Montsec is mainly influenced by the special meteorological conditions and the local planetary boundary layer (PBL) height. Especially in summer, the higher temperature and solar radiation enhances the growth of PBL and transport of OA to high elevation (Ripoll et al., 2015).

3.2 Comparison against PMF results

The modeled OA components were evaluated using AMS/ACSM measurements analyzed with PMF at different stations (Table 1). The statistical results are presented in Table 2. The CAMx-NEW led to a better agreement between the modeled and measured primary OA at most sites, except in Zurich (where HOA and BBOA tended to be overestimated by CAMx-NEW and underestimated by CAMx-BASE). In general, the NEW parameterization improved the underestimation of OOA by reducing the mean bias by 0.3 $\mathrm{\mu }$g m${}^{-\mathrm{3}}$ (SMEAR II) to 1.7 $\mathrm{\mu }$g m${}^{-\mathrm{3}}$ (Bologna), while increasing the mean bias at Zurich and SPC where OOA concentrations were overestimated by both CAMx-BASE and CAMx-NEW. To better understand the different model performance, the modeled and measured OA components were compared separately for the four different seasons (Fig. 3). The overestimation of OOA in autumn at SPC was largely due to the absence of fog droplet activation in the model as discussed in Sect. 3.1.2. In spite of the overestimation for the whole period, the winter OOA at Zurich was still underestimated by 73 % (BASE) and 64 % (NEW). The underestimation of winter OOA was also found at all the other studied sites (Marseille, Bologna and Paris SIRTA), while the summer OOA showed good agreement with the PMF results (Fig. 4a and d), indicating a possible underestimation of SOA from biomass burning (Gilardoni et al., 2014, 2016; Paglione et al., 2019; Qi et al., 2019). The NEW parameterization improved the OA modeling by increasing winter OOA by 29.3 % (SIRTA) to 41.7 % (Bologna). Compared to the modeled BBOA (Fig. 4b and e) and HOA (Fig. 4c and f) which were improved considerably by using the NEW parameterization, the SOA modeling still needs to be improved.

Figure 5

Comparison between modeled relative contributions of OA components and positive matrix factorization (PMF) analysis results at Zurich (urban site) and Montsec (remote site). GV: gasoline vehicles; DV: diesel vehicles; BB: biomass burning; OthA: other anthropogenic sources; BIO: biogenic sources.

[Figure omitted. See PDF]

Figure 6

Diurnal variations of modeled OA components by CAMx-NEW and PMF studies at urban sites in winter. The lines represent the average value of modeled and measured OA components during the same periods for each stations (as displayed in Fig. 2b), and the bars represent 25th and 75th percentiles of hourly data.

[Figure omitted. See PDF]

The model performance in reproducing the diurnal variations of the OA components varies with sites. Figure 6 shows the diurnal variations of HOA, BBOA and OOA by CAMx-NEW and PMF in winter. The simulation of primary organic aerosols (both HOA and BBOA) at site scale largely depends on the anthropogenic emission inputs. The modeled HOA shows dual peaks in the morning and evening rush hours for all sites as a consequence of the same diurnal variation factors in the emission model, while similar dual peaks only occur in Marseille and Bologna for the PMF HOA. The PMF HOA peaks are less pronounced in the urban background site Zurich and the suburban site SIRTA, where the magnitude of HOA is also lower compared to the other two urban stations. Similar to HOA, the modeled BBOA generally showed dual peaks in the morning and evening cooking and heating times. However, the diurnal pattern of PMF BBOA showed more variability over the sites. The largest difference between modeled and PMF BBOA occurred in Marseille, where the peaks of PMF BBOA were generally later than the modeled results and much higher at night. The high PMF BBOA during night mainly comes from the BBOA transported from the valleys near Marseille by the night land breeze (Bozzetti et al., 2017), whereas the local meteorology was poorly reproduced at Marseille, a common problem at coastal sites for meteorology modeling (Fig. S2b). Another important reason for the underestimated BBOA is that green waste combustion and agricultural fires comprise a large fraction of the BBOA in Marseille in February (Bozzetti et al., 2017); however, this part of BBOA was not modeled due to lack of emission data. In spite of a general underestimation for the modeled OOA in winter, both the CAMx-NEW and PMF OOA have rather flat diurnal patterns (Fig. 6). Better agreement between modeled and PMF results is found at rural/remote sites during spring to autumn (Fig. 7), especially at SMEAR II, Finokalia and Montsec, where OOA from biogenic sources contribute more than 90 % of the total OA according to the PMF results. The PMF OOA at San Pietro Capofiume shows a peak during the day when the liquid water content was low (Gilardoni et al., 2014), and the peak is higher during the less foggy period after 23 November (Fig. 7e) than during the highly foggy period (Fig. 7d). However, the modeled OOA is flat as the fog scavenging effect is poorly reproduced.

Figure 7

Diurnal variations of modeled OOA components by CAMx-NEW and PMF studies in rural or remote sites. The lines represent the average value of modeled and measured OOA. Bars represent 25th and 75th percentiles of hourly data.

[Figure omitted. See PDF]

Figure 8

Relative contributions of different sources (POA and SOA) to total OA in winter (a, b) and summer (c, d). The winter and summer results are the averages of December–January–February and June–July–August, respectively. OA from biogenic sources is all secondary so the POA panels for biogenic sources are empty.

[Figure omitted. See PDF]

Although the model performance was substantially improved by the NEW parameterization, some limitations still remain to be further improved. The winter OOA is still underestimated at most sites. It could come from the missing pathways of SOA formation such as the aqueous processing of water-soluble organics (Ervens et al., 2011), which was found to contribute up to $\sim$ 20 % of winter OA measured in Bologna (Gilardoni et al., 2016; Meroni et al., 2017). The uncertainties in the model parameters (reaction rate, yield, etc.) could also account for the underestimation. A recent study reported that the reaction rate of SVOC with hydroxyl radicals is highly uncertain without strong constraint of the whole set of parameters for various processes such as the vapor wall loss in chamber experiments (Bertrand et al., 2018). Another potential limitation is related to the uncertainties in SVOC/IVOC emissions. We adopted a factor of 3 for the SVOC and POA ratio for the whole domain; however, the substantial spatial and temporal variability of the factor could lead to over- or underestimation of SVOC emissions at site scale (Denier van der Gon et al., 2015) and therefore over- or underestimation of the SOA, as well as of POA. It could also partially explain the differences in model performance for the temporal variation of HOA and BBOA for each site. To further improve the model performance, it is necessary to continuously update the chemical mechanism in models by introducing missing processes and improving the parameterization based on advanced knowledge, as well as to improve the emissions by including more site-specific sources, IVOC and SVOC estimates and updated diurnal variation profiles.

3.3 Spatial and seasonal variation of OA components and sources

Due to the better performance of CAMx-NEW in predicting the OA components, the spatial and temporal variations of OA components and sources were investigated using the CAMx-NEW parameterization. The contributions of gasoline and diesel vehicles, biomass burning, other anthropogenic sources (including shipping, energy sector and industry) and biogenic sources in winter and summer are displayed in Fig. 8. Country-scale relative contributions of each source to the total OA can be found in Table S6. In winter, the largest contribution over the whole domain comes from biomass burning from residential and agriculture activities (Fig. 8a), and the biomass burning POA is the dominant component, with an average fraction of 43 % and the highest value reached in Slovenia (19.9 $\mathrm{\mu }$g m${}^{-\mathrm{3}}$, Fig. S4). The second largest OA component is the biogenic SOA (25 %), followed by the biomass burning SOA (19 %). A high contribution of biogenic SOA in winter is mainly found in Sweden, Ireland, and Spain (Table S6). Ireland and Spain have high emissions of the biogenic SOA precursor monoterpenes due to a high coverage of Norway spruce trees, the major emitters of monoterpenes, as well as comparatively high temperature. Our biogenic emission model assumes that grids with snow coverage of more than 50 % have zero monoterpene emissions in winter (Jiang et al., 2019). Therefore, there are still considerable monoterpene emissions in winter in areas with low snow coverage such as Ireland and Spain. The high biogenic SOA fraction in Sweden is mostly due to low SOA from anthropogenic activities.

Table 3

Modeled relative contribution of different sources to total OA in eight sub-regions in Europe. DJF: December–January–February; JJA: June–July–August.

[Figure omitted. See PDF]

3.4 Regional OA sources

In order to understand the regional variations of OA components and sources, we divided the model domain into eight sub-regions: the Iberian Peninsula (IP), the Mediterranean (MD), Po Valley (PV), eastern Europe (EE), central Europe (CE), Benelux (BX), Ireland and Great Britain (IG) and Scandinavia (SC). The regional division and average fractions of the POA and SOA components for each region are shown in Table 3 and Fig. 10. In winter, OA in most regions is dominated by biomass burning POA except for Ireland and Great Britain and the Iberian Peninsula, where the biogenic SOA contributes most with a fraction of 66 % and 45 %, respectively. Both of these regions have relatively high monoterpene emissions as explained in Sect. 3.3. Meanwhile, the high biogenic SOA fraction in Ireland and Great Britain could also come from underestimated contribution of biomass burning, in which the peat combustion covers a considerable portion (Lin et al., 2018) but is not considered in the emission inventory. The second largest POA source in all regions is other anthropogenic sources, with a similar contribution to total OA (4.9 %–9.6 %). The Po Valley and Benelux regions feature a comparatively higher contribution of POA from diesel vehicles (4.5 % for PV, 4.9 % for BX). For the winter SOA, the Mediterranean, Eastern Europe and Po Valley are dominated by biomass burning (20 %–29 %), while the other regions are dominated by biogenic SOA (30 % in CE – 66 % in BX). The summer OA is dominated by SOA for all the regions due to a considerable contribution of biogenic SOA. Scandinavia shows the highest fraction of biogenic SOA reaching up to 77 %, followed by Ireland and Great Britain (70 %) and the Iberian Peninsula (55 %). In summer, a significant increase is seen in the contribution of gasoline SOA compared to winter, especially in central and southern Europe. The Po Valley has the highest fraction of gasoline SOA (9.4 %), followed by the Mediterranean (7.3 %) and central Europe (4.0 %). The contributions of the POA components show more regional variation. In the coastal regions, e.g., the Mediterranean, Ireland and Great Britain, Benelux and Scandinavia, POA-OthA has the highest contribution, while in the other regions POA-BB contributes most (although with much lower values than in winter).

Among all the anthropogenic sources, OthA contributes from 9 % (EE) to 25 % (IG) in winter and from 33 % (EE) to 64 % (IG) to the anthropogenic OA (excluding the biogenic OA) in summer. This component mostly comes from the non-road transport, energy sector, industrial production and processes, solvent use, waste treatment, extraction of fossil fuels and brake wear. The CAMx-VBS is not able to further separate the OA sources from OthA. To give a general view about the contribution of specific sources in OthA, the contributions of PM${}_{\mathrm{2.5}}$ (as a proxy of POA) and NMVOC (as a proxy of gaseous OA precursor) emissions from specific sources are shown in Fig. S5. The highest ratio of PM${}_{\mathrm{2.5}}$ emissions from OthA reaches 82 % in the Iberian Peninsula due to high contribution of non-road transport (shipping), and the highest fraction of NMVOC emissions from OthA (major proxy of the gaseous precursors from OthA) reaches 86 % due to the solvent use. However, compared to the widely studied OA sources such as biomass burning and road traffic, knowledge about these sources is still quite limited.

4 Conclusions

This modeling study was conducted to identify the sources of organic aerosol (OA) components in Europe using the air quality model CAMx for the whole year of 2011. In order to improve the model performance for the organic aerosol which is generally underestimated by air quality models, we updated the VBS parameterization based on recent findings in chamber experiments with biomass burning and diesel vehicles (CAMx-NEW). A more source-specific PSI-VBS scheme compared with previous studies was used to calculate separately the OA contributions from old and new diesel and gasoline vehicles, biomass burning (residential wood burning and agricultural combustion), other anthropogenic sources (mainly shipping, industry and energy production) and biogenic sources in Europe. We modified the basis sets and emissions in order to be able to identify OA sources such as gasoline vehicles (GV), old diesel vehicles before Euro 4 emission standards (DO), new diesel vehicles with Euro 4 and higher (DN), biomass burning (BB), biogenic sources (BIO) and other anthropogenic sources (OthA). CAMx-NEW enhances the role of semi-volatile organic compounds (SVOC) and enables the further aging of secondary condensable gases from biomass burning, which significantly improved the model performance for the total OA.

Another important outcome of this study is the evaluation of the model results with measurements over a longer period than in other studies which strengthened our confidence on our modeled source apportionment. The model evaluation using measurements at nine ACSM/AMS stations in Europe showed that the CAMx-NEW reduced the mean fractional error (MFE) between the modeled and measured OA by 3 %–47 % compared to the standard VBS parameterization (CAMx-BASE). The model performance at reproducing the OA components (HOA, BBOA and OOA) was also improved by using CAMx-NEW. The MFE between modeled and PMF HOA was 24 %–72 % lower with CAMx-NEW than with CAMx-BASE, and the MFE of BBOA was reduced by 23 %–47 % with CAMx-NEW. The general underestimation of OOA by models was also improved with reduced mean bias of 0.3–1.7 $\mathrm{\mu }$g m${}^{-\mathrm{3}}$ by the modified VBS scheme.

The model results of CAMx-NEW suggested that biomass burning and biogenic sources are the major sources of OA in Europe in winter and summer, respectively. The highest contribution from biomass burning to OA was predicted in the Po Valley and Eastern European regions with $>$ 70 % in winter. The SOA from biogenic sources was calculated to cover more than 50 % of summer OA on average (highest with 77 % in Scandinavia). Over Ireland and Great Britain, the contribution of SOA from biogenic sources was more than 50 % even in winter, possibly due to underestimated biomass burning emissions in that area. The other anthropogenic component excluding biomass burning and road traffic (shipping, energy sector, industry, etc.) was identified as another group of important OA sources with an average contribution of 9 % and 19 % in winter and summer, respectively, and the highest contribution to total OA was predicted to be 29 % in the Mediterranean in summer. The contribution of road traffic (gasoline and diesel vehicles) was rather small on average ($\sim$ 5 %) but was higher in metropolitan areas. The highest contribution of gasoline and diesel vehicles to total OA was found in the Po Valley region with a value of 16 % in summer.

The results of this study provide information on the OA source apportionment with a regional perspective, which complements the current measurements and PMF analysis to understand regional differences of OA sources in Europe, and to identify needs for future studies. The modeling of OA source apportionment needs further evaluation based on measurements with wide spatial coverage, especially using advanced measuring techniques with improved capability of SOA identification. The considerable contribution of OA from shipping, industry and energy sectors highlights the importance of more experimental and model studies on such sources to provide explicit parameters (yields, volatility distributions, SVOC emissions, etc.) which are currently estimated by default parameters in most of the air quality models. The emission inventory remains to be improved for the highly uncertain cooking and wildfire emissions, as well as to include more site-specific emission sources, in order to further enhance the model performance at predicting the OA sources.

Data availability

Model data of this study are available online at 10.5281/zenodo.3540796 (Jiang and Aksoyoglu, 2019a). Source codes of the PSI-VBS module are available at 10.5281/zenodo.3540826 (Jiang and Aksoyoglu, 2019b).

The supplement related to this article is available online at: https://doi.org/10.5194/acp-19-15247-2019-supplement.

Author contributions

JJ and SA conceived the study. JJ carried out the model modification, simulation and data analysis. GC contributed to model setup and development of the source apportionment code. HACDG provided the anthropogenic emissions. IEH, FC, SG, MP, MCM, OF, YZ, NM, LH, AV, KF, COD and JO provided the measurement data and contributed to data interpretation. SA, ASHP and UB supervised the entire work development. The manuscript was prepared by JJ. All authors discussed and contributed to the final paper.

Competing interests

The authors declare that they have no conflict of interest.

Acknowledgements

We would like to thank the European Centre for Medium-range Weather Forecasts (ECMWF) for access to the meteorological data, the European Environmental Agency (EEA) for the air quality data, the National Aeronautics and Space Administration (NASA) and its data-contributing agencies (NCAR, UCAR) for the TOMS and MODIS data, the global air quality model data and the TUV model. We acknowledge the continuous support of CAMx by RAMBOLL. Simulation of WRF and CAMx models were performed at the Swiss National Supercomputing Centre (CSCS). Anna Alastuey and Andrés Ripoll from IDAEA-CSIC are acknowledged.

Financial support

This study was financially supported by the Swiss Federal Office for the Environment (FOEN). The measurements were supported by EPA-Ireland (2016-CCRP-MS-31), Emilia-Romagna region's Supersito project (DRG 428/10; DGR 1971/2013), Generalitat de Catalunya (AGAUR 2017 SGR41), the Spanish Ministry of Economy, Industry and Competitiveness (Ramón y Cajal fellowship awarded to María Cruz Minguillón), EU-FP7 ACTRIS project (grant agreement no. 262254), and the COST Action CA16109 Chemical On-Line cOmpoSition and Source Apportionment of fine aerosoL (COLOSSAL).

Review statement

This paper was edited by Andrea Pozzer and reviewed by Manish Shrivastava and two anonymous referees.