We used the UCD/CIT regional chemical transport model (Kleeman and Cass, 2001) to simulate the emissions, transport, chemistry, and deposition of air pollutants over the state of California at a grid resolution of 24 km and over southern California (see Fig. S1 in the Supplement) using a nested 8 km grid from 20 July to 2 August 2005. The results and analysis were focused on model predictions over southern California because the region, with approximately 15 million people, is home to one of the most polluted cities in the United States (Los Angeles; ALA, 2017). The time period for simulation was primarily chosen because the model has been previously evaluated for this time period (Jathar et al., 2016) and applied to examine important sources and formation pathways of OA (Cappa et al., 2016; Jathar et al., 2015, 2016, 2017b). The recent literature describes the latest version of the UCD/CIT model but we provide a very brief description of the models and inputs used in this work. Anthropogenic emissions for California were developed using the California Regional PM10/PM2.5 Air Quality Study (CRPAQS) inventory of 2000 but scaled to match conditions in 2005. Wildfire emissions were based on the model FINN (Fire Inventory from National Center for Atmospheric Research) (Wiedinmyer et al., 2011) although they were not found to significantly contribute to OA during the simulated time period (Docherty et al., 2011). Biogenic emissions were based on the model MEGAN (Model of Emissions of Gases and Aerosols from Nature) (Guenther et al., 2006). The Weather Research and Forecasting (WRF) v3.4 model (https://www.mmm.ucar.edu/weather-research-and-forecasting-model, last access: 25 March 2019) was used to produce hourly meteorological fields. National Centers for Environmental Prediction's (NCEP) NAM (North American Mesoscale) analysis data were used to set the initial and boundary conditions for WRF. The gas- and particle-phase initial and hourly varying boundary conditions were based on the results from the global model MOZART-4/NCEP (Emmons et al., 2010). The gas-phase chemistry was modeled using SAPRC-11 (Carter, 2010).

The Base simulation – representing our most comprehensive simulation – included the updates described in Sect. 2.2.2: a source-resolved semi-volatile and reactive treatment of POA, source-resolved SOA formation from SVOCs and IVOCs, and correction of the subsequent SOA formation for vapor wall losses in chambers. The Base simulation included sub-simulations at two resolutions (24 and 8 km) with two NOx parameterizations (low and high NOx).

Additional simulations were designed and performed with two objectives in mind: (i) to examine the influence of each update included in this work and (ii) to test the sensitivity in model predictions to uncertainties inherent in the updates and other model inputs. A set of four simulations was performed to systematically study the influence of model updates. These included the following simulations where only one update was made over the previous configuration: (1) Traditional – non-volatile POA, no IVOCs, SOA from VOCs, and no correction for chamber vapor wall losses; (2) SVOC – semi-volatile POA, no IVOCs, SOA from SVOCs and VOCs, and no correction for chamber vapor wall losses; (3) IVOC – semi-volatile POA; IVOCs; SOA from SVOCs, IVOCs, and VOCs; and no correction for chamber vapor wall losses; and (4) Base – semi-volatile POA; IVOCs; SOA from SVOCs, IVOCs, and VOCs; and correction for chamber vapor wall losses. Successive differences in model predictions between the Traditional, SVOC, IVOC, and Base simulations were used to systematically examine the influence of the semi-volatile and reactive POA, IVOCs, and chamber vapor wall losses respectively.

A set of six simulations were performed to study uncertainties in model inputs. The SVOCmax (5) simulation assumed that POA from all sources (all POA except marine POA) was semi-volatile and modeled using the volatility distribution for diesel exhaust POA. Diesel POA was chosen since it was the most volatile of the volatility distributions used in this work. This simulation bounded the maximum loss in POA mass to evaporation. The IVOCmax (6) simulation assumed that all sources (combustion and non-combustion except biogenic sources) emitted IVOCs, which were estimated using an IVOC:NMOG ratio of 0.2 and allowed to form SOA equivalent to a C15 alkane. This simulation provided an upper-bound estimate to the contribution of IVOCs to ambient SOA although the IVOC emissions and their potential to form SOA could be even higher than that assumed here. The No Aging (7) simulation assumed no multigenerational aging or, in other words, the emitted precursor was allowed to react with OH and form four functionalized products with no further oxidation. This simulation investigated the influence of multigenerational aging on ambient SOA. The VOCspec (8) simulation updated the VOC speciation for on- and off-road gasoline and diesel vehicles based on a comprehensive set of measurements performed on an in-use fleet (May et al., 2013a, b). This simulation examined the influence of updated emissions profiles on the non-IVOC contribution to SOA. The Aromatic (9) simulation assumed that the oxidation of SVOCs and IVOCs to form SOA was modeled using toluene. There were two reasons for choosing toluene. First, both mono- and polycyclic aromatic compounds are found in gasoline and diesel fuel (Gentner et al., 2012) and in tailpipe emissions from mobile sources (Zhao et al., 2015, 2016), and oxygenated aromatic compounds such as phenols, guaiacols, and syringols are found in biomass burning emissions (Schauer et al., 2001; Stockwell et al., 2015). Second, aromatic compounds, similar to alkanes, have been studied in detail for their potential to form SOA and are recognized to form more SOA than linear alkanes for the same carbon number. This simulation provided an upper-bound estimate for SOA formation from the oxidation of SVOCs and IVOCs. Finally, the SVOCcooking (10) simulation used a hydrocarbon/linear alkane distribution based on the measured volatility distribution of Louvaris et al. (2017) to represent POA from food cooking sources. This simulation examined the effect of a more realistic volatility distribution for food cooking POA on mass concentrations of POA and SOA from SVOCs.

The UCD/CIT model was run on the High Performance Computing Cluster run by Engineering Network Services at Colorado State University. Although the number of cores varied based on availability, on average each simulation used 96 cores and required 5 days to execute 19 simulated days. Since each set included four sub-simulations, each simulation required ∼5 days and all simulations in this work required ∼180 days of computational time.

Model predictions were evaluated against gas-phase measurements of SOA precursors and particle-phase measurements of OA mass concentrations and composition. Here, we briefly describe the primary measurement data and any post-processing of the data we performed prior to undertaking the model evaluation.

Gas-phase measurements of SOA precursors were from two different sources. The first source was routine daily-averaged measurements of single-ring aromatics made by the South Coast Air Quality Management District (SCAQMD, 2017) in southern California at three different sites: north Los Angeles, Riverside, and Long Beach. While measurement data were available at three other sites, data were not available for 2005, our modeled year, and hence not included. These gas-chromatography-based measurements were available every 12th day and included the following aromatic species: benzene, toluene, o/m/p-xylene, ethyl-benzene, and styrene. Since there was little overlap between the modeled episode (14-day period over July–August) and available aromatic data, the measurement data were averaged over a 3-month period in the summer (15 May to 15 September) and then compared to the episode-averaged model predictions. The second source was gas-chromatography mass-spectrometry measurements of single-ring aromatics (Borbon et al., 2013) and IVOCs (Zhao et al., 2014) made at the Pasadena ground site in the months of May and June of 2010 as part of the CalNex campaign. The single-ring aromatics were measured every hour and included the following species: benzene, toluene, o/m/p-xylene, ethyl-benzene, and styrene. The IVOCs were measured every 3 h and included most of the reduced and oxidized organic species with a carbon number larger than 12. Since these measurements were from a different time period, we compared campaign-averaged measurements against episode-averaged model predictions.

Particle-phase measurements were from two different sources as well. The first source was routine daily-integrated measurements of organic carbon (OC) in southern California from four sites in the Chemical Speciation Network (CSN; central Los Angeles, Riverside, Simi Valley, and Escondido) and six sites in the Interagency Monitoring of Protected Visual Environments (IMPROVE) network (San Rafael, Riverside-Rubidoux, San Gorgonio Wilderness, Joshua Tree NP, Agua Tibia, and San Gabriel). The CSN is a network of ∼50 urban measurement sites across the United States where pollutant concentrations are typically higher, more variable, and representative of local sources, and measurements are made once every 3 days. IMPROVE is a network of ∼200 rural/remote continental sites typically located in national parks across the United States where pollutant concentrations are lower, less variable, and representative of regional influences, and measurements are made once every 3 days. Over the 14-day episode modeled in this work, three measurements from the CSN and five measurements from the IMPROVE network were available for comparison. We used an organic-aerosol-to-organic-carbon ratio (OA:OC) of 1.6 to calculate OA at the CSN sites (Docherty et al., 2011, measured an OA:OC ratio of 1.77 during the SOAR-1 campaign, after correction with the updated calibration of Canagaratna et al., 2015) and a ratio of 2.1 to calculate OA at the IMPROVE sites (Turpin and Lim, 2001). The CSN data are artifact corrected but we subtracted 0.5 µg m-3 from the calculated OA mass concentrations to blank correct the data (Subramanian et al., 2004). The IMPROVE data are both blank and artifact corrected. We note that a negative evaporation artifact has been reported at IMPROVE sites in the southeast US (Kim et al., 2015), but it is not known whether such an artifact may be present in this region and no correction has been made. The second source was particle measurements made at the ground site in Riverside as part of the SOAR-1 campaign during the summer of 2005 (Docherty et al., 2008, 2011). These measurements included hourly-averaged mass concentrations and elemental ratios of H:C and O:C for OA, as well as estimates of the POA–SOA split based on results from a positive matrix factorization analysis.

Gas- and particle-phase emissions of organic compounds in the 8 km southern California domain, averaged over the 14-day episode, are shown in Fig. 1. The 8 km domain, shown in Fig. S1 in the Supplement, includes the entire Los Angeles metropolitan statistical area, parts of the Pacific Ocean, and forested areas surrounding the urban area. The emissions are color-coded by source type and include all species that contribute to direct emissions and atmospheric formation of OA. These do not include emissions of marine POA since those were calculated in line in the UCD/CIT model. Since the POA repartitioned between the gas and particle phases after emission, POA was split into POA and SVOC that represented the particle and gas portions of POA partitioned at an urban OA mass concentration of 9 µg m-3. We chose 9 µg m-3 to partition POA because the campaign-averaged OA mass concentration at Riverside during SOAR-1 was 9 µg m-3. If one discounts the POA emissions in the “other” category (which is mostly made of road, agricultural, and construction dust), the repartitioning results in about 60 % of the POA emitted evaporating as SVOC vapors; these vapors oxidize in the atmosphere to form SOA. As noted earlier, a relatively more volatile treatment compared to that described in the recent literature suggests that we may have overestimated the POA evaporation from food cooking sources. Mobile sources accounted for 20 % of the POA and 35 % of the SVOC vapors and competed with food cooking as an important source of primary emissions and one which accounted for 15 % of the POA and 44 % of the SVOC vapors. IVOC, long-alkane, and aromatic emissions were roughly on the same order of magnitude but taken together were approximately an order of magnitude larger than the POA emissions. This suggests that even at low SOA mass yields (say < 10 %), the OA formed from the oxidation of these precursors could quickly exceed direct emissions of POA.

Emissions of total IVOCs were slightly lower than those for long alkanes (by ∼30 %) and aromatics (by ∼40 %) but a factor of 2 higher than the sum of POA and SVOCs. Previously, IVOC emissions have been estimated by scaling POA emissions by a factor of 1.5 to 3 derived from gas–particle partitioning calculations (Dzepina et al., 2009; Shrivastava et al., 2008) and from atmospheric measurements (Ma et al., 2017). While our estimates for IVOC emissions are within the previously used range, our estimates were informed by a broader suite of source measurements, which will help reduce the uncertainty in IVOC emissions and related SOA formation in atmospheric models. IVOC emissions from mobile sources were similar to aromatic emissions but twice the long-alkane emissions from the same source. We note that in this work we only considered IVOC emissions from combustion sources, but recent work suggests that volatile chemical products present in sources such as pesticides, coatings, cleaning agents, and personal care products may be a large source of IVOCs in urban environments (McDonald et al., 2018).

Mobile sources – dominated by gasoline use – accounted for a much larger fraction of the anthropogenic SOA precursors (85 % of IVOCs, 27 % of long alkanes, and 55 % of aromatics) in this study. Hence, mobile source regulation on precursor emissions from gasoline vehicles (e.g., limits on emissions of unburned hydrocarbons) has and could have a much larger influence on controlling ambient OA than regulating direct emissions of POA, although this ultimately depends on the extent of conversion of these species to SOA. Finally, biogenic precursor emissions of isoprene, monoterpenes, and sesquiterpenes were about a factor of 3 higher than the combined emissions of SVOCs, IVOCs, long alkanes, and aromatics and will continue to be an important source of SOA in southern California. However, their impact on urban OA/SOA will be smaller since these emissions are primarily limited to regions outside the urban areas.

In Fig. 2 we plot predictions of the 14-day-averaged mass concentrations for OA, POA, SOA, and contributions from three lumped SOA precursors (long alkanes and aromatics, SVOC and IVOCs, and biogenic VOCs) from the Base case simulation. We used the terminology developed by Murphy et al. (2014) to describe the SOA from the different sources. To reiterate, the Base case simulation included a semi-volatile treatment of POA; SOA formation from oxidation of SVOCs, IVOCs, and VOCs; multigenerational aging; and SOA parameterizations that accounted for the influence of chamber vapor wall losses. The mass concentrations in Fig. 2 account for SOA formation under varying NOx levels as per Eq. (2) (logarithmic dependence on the VOC:NOx ratio). We chose Eq. (2) because it produced the highest SOA mass concentrations and presented an upper bound on SOA formation.

The highest OA mass concentrations were found in three general regions: the densely populated Los Angeles–Orange–Riverside County region, likely attributed to heavy transportation emissions; along the coast as a result of sea spray emissions; and in biogenic-VOC-dominated areas. In central Los Angeles (grid cell containing the CSN site), OA accounted for 38 % of the modeled non-refractory PM2.5 mass concentration with 20, 25, and 18 % contributions from sulfate, nitrate, and ammonium aerosol. A sensitivity simulation that turned emissions of marine POA off suggested that the marine POA mass concentrations in central Los Angeles were ∼0.9 µg m-3, which were considerably higher than the coastal measurements made during CalNex in 2010 (Hayes et al., 2013). Measured mass concentrations of POA over the open ocean west of California were ∼0.2 µg m-3 during CalNex in 2010, and it was expected that these mass concentrations would be substantially lower by the time they were transported to central Los Angeles (Hayes et al., 2013). Sea spray emissions in the UCD/CIT model are based on the parameterization of Gong et al. (2003) and may need to be revisited in the future.

The broader spatial trends of OA, POA, and SOA were in line with results from earlier chemical transport model studies that have treated POA as semi-volatile and modeled SOA formation from SVOCs and IVOCs (Ahmadov et al., 2012; Jathar et al., 2017a; Koo et al., 2014; Robinson et al., 2007; Tsimpidi et al., 2010). POA mass concentrations were highest in upwind (e.g., 3.4 µg m-3 in central Los Angeles) and lower in downwind (e.g., 2.7 µg m-3 in Riverside) locations as the POA emissions that were transported away from the source region evaporated with dilution. SOA mass concentrations, in contrast to POA, had a more regional presence with lesser differences between the upwind and downwind regions (e.g., 2.4 µg m-3 in Riverside vs. 2.2 µg m-3 in central Los Angeles) or in regions with high emissions of biogenic VOCs (e.g., 2.5 µg m-3 inside the Los Padres National Forest). To assess the relative contribution of POA and SOA to total OA, we plot the POA:SOA ratio in Fig. S2, which suggests a POA:SOA ratio of ∼1.6 in near-source regions and lower elsewhere, e.g., ∼0.4, 0.8, and 1.2 in representative marine, biogenic-VOC-dominated, and urban downwind regions. These POA:SOA splits qualitatively aligned with the hydrocarbon-like and oxygenated organic aerosol (HOA and OOA) splits estimated in aerosol mass spectrometer datasets in urban locations worldwide (Jimenez et al., 2009; Zhang et al., 2007). However, we predict POA:SOA ∼1 for Riverside during SOAR-1, compared to a measured ratio of ∼0.25 (Docherty et al., 2008), which indicates that SOA may still be underestimated in the model. A comparison of the OA composition predictions with the aerosol mass spectrometer measurements is described in Sect. 4.

Panels (d) through (f) show contributions of three distinct SOA precursor classes to total SOA. Alkane and aromatic VOCs – included as SOA precursors in most atmospheric models – appeared to contribute a maximum of 1.2 µg m-3 of what we refer to as aV-SOA downwind of the source region. The majority of this aV-SOA (75 %) originated from aromatic precursors, implying that alkane VOCs are unlikely to contribute much to the anthropogenic SOA or total OA burden in urban areas, consistent with our earlier work (Cappa et al., 2016; Jathar et al., 2016). We note that emissions inventories typically only include alkane species with carbon numbers less than 12 (Pye and Pouliot, 2012), and longer alkanes with carbon numbers larger than 12 are included as part of the POA, SVOC, and IVOC emissions. Together aS-SOA and aI-SOA mass concentrations exhibited a similar spatial pattern over the domain but were substantially lower than the aV-SOA mass concentrations – reaching a maximum of only 0.5 µg m-3. The lower aS-SOA and aI-SOA mass concentrations were somewhat contrary to earlier work that has argued that SVOCs and IVOCs are an equal or dominant precursor of anthropogenic SOA when compared to aV-SOA, especially in urban areas (Jathar et al., 2014, 2017a; Woody et al., 2016). The reason for these lower concentrations can be partially attributed to the precursor-dependent influence of accounting for vapor wall losses in chamber experiments (probed in greater detail in Sect. 3.4). Biogenic SOA or bV-SOA mass concentrations exceeded 3.2 µg m-3 in regions with high biogenic emissions but were slightly less than 1 µg m-3 in urban regions where the POA mass concentrations were the highest. Previous work has suggested that the bV-SOA in urban regions is formed outside but later transported to the urban region (Hayes et al., 2015; Heo et al., 2015). Overall, the averaged results over the urban areas appeared to be split evenly between POA, anthropogenic SOA (aV-SOA + aS-POA + aI-SOA), and biogenic SOA (bV-SOA).

We examined the absolute OA mass concentrations and precursor contributions to SOA in central Los Angeles across four different simulations to better understand the effect of successive updates: semi-volatile and reactive POA, IVOCs, and accounting for vapor wall losses. We chose central Los Angeles (grid cell containing the CSN site) as our study area as it is representative of an urban location with a large population density and suffers from some of the poorest air quality in the United States (ALA, 2017); results from the sensitivity simulations in Sect. 3.5 are also discussed at this specific site. Results at other urban locations (e.g., Riverside, Simi Valley) had similar SOA precursor fractional contributions although the absolute concentrations did vary a little (see Fig. S3). In Fig. 3, we plot the 14-day-averaged, precursor-resolved OA mass concentrations and precursor contributions to SOA in Los Angeles from two pairs of four different simulations. The two pairs represent model predictions based on the low- and high-NOx parameterizations.

Semi-volatile and reactive POA. Differences in the Traditional and SVOC simulations were used to highlight the influence of including a semi-volatile and reactive treatment of POA. The semi-volatile POA treatment resulted in evaporation of the primary POA emissions from combustion sources (on- and off-road gasoline and diesel, woodsmoke, biomass burning, and food cooking) and reduced POA mass concentrations by 35 % in central Los Angeles. A ratio of the POA mass concentrations from the SVOC simulation to those from the Traditional simulation suggested that the POA mass was reduced by approximately 30 % to 50 % in the urban environment around the central Los Angeles site (Fig. S4). Overall, the POA reductions appeared to be smaller than those implied by the volatility distributions of May and coworkers (May et al., 2013a, b, c) and those simulated in other atmospheric models (Robinson et al., 2007). For gasoline, diesel, and biomass burning, May and coworkers (May et al., 2013a, b, c) proposed a 45 % to 80 % reduction in POA mass concentrations at ambient OA mass concentrations between 1 and 10 µg m-3. This difference was mainly because we only modeled certain combustion-related POA to be semi-volatile (i.e., gasoline, diesel, biomass burning, and food cooking sources) while earlier modeling work has considered POA from all sources to be semi-volatile (e.g., marine, dust). The use of a less volatile and more realistic food cooking POA than that used in this work (informed by the works of Woody et al., 2016, and Louvaris et al., 2017) would tend to further increase the discrepancy between our work and the findings of May and coworkers. Hu et al. (2014) found that the combustion sources considered to be semi-volatile in this work accounted for about half of PM2.5 mass concentrations in Los Angeles. The POA mass reductions shown here are conservative and might have been larger if there was evidence that sources other than those considered here (e.g., marine, dust) produced POA that was semi-volatile too, although this scenario seems unlikely.

Allowing the POA vapors or SVOCs to react resulted in only a small fraction of their oxidation products condensing back as aS-SOA. For example, of the 1.75 µg m-3 of POA lost at the central Los Angeles site, only 0.082 µg m-3 for the low-NOx simulations and 0.068 µg m-3 for the high-NOx simulations was regained as aS-SOA from oxidation reactions. This implied a very low chemical conversion efficiency (∼4 %) for the POA-to-SVOC-to-aS-SOA pump within the urban area (Miracolo et al., 2010). The SVOCs, at an ambient concentration of 9 µg m-3, from gasoline exhaust, diesel exhaust, and biomass burning emissions had an average carbon number between 18 and 20. Calculations with a box model version of the SOM suggested that the SOA mass yields for C18 and C20 alkanes were between 33 % and 86 % where the range includes yields for low-NOx and high-NOx parameterizations. One possible explanation for the difference between the chemical conversion efficiency in the 3D model and box model yields was that only a small fraction of the SVOCs had the opportunity to react with OH and form SOA before they were transported out of the urban area. If we assume that most of the sS-SOA in the grid cell that contains the Los Angeles site was from the oxidation of SVOCs released in that grid cell and from grid cells that are up to two grid cells away, our results do not appear unrealistic. For example, for an SOA precursor with an OH reaction rate constant of 2.4×10-11 cm-3 molecules-1 s-1 (average value from a C18 and C20 linear alkane) and an SOA mass yield of 60 % (average from the SOA mass yield range described earlier for a C18 and C20 linear alkane), the chemical conversion efficiency would be 3.5 %–15 % with a daily-averaged OH concentration of 1.5×106 molecules cm-3 and a reaction time of 0.5–2.3 h. A reaction time of 0.5 to 2.3 h corresponds to a transport of 4 km (half a grid cell) and 20 km (2.5 grid cells) at an average wind speed of 2.4 m s-1 (Weather Spark).

The low- and high-NOx parameterizations had little effect on the aS-SOA mass concentrations presumably because the n-dodecane-based parameterization used for semi-volatile POA exhibited marginal differences in SOA production under low- and high-NOx environments (Loza et al., 2014). Finally, SOA parameterizations based on including the vapor wall loss effect only marginally increased the aS-SOA mass concentrations, especially when viewed in light of the SOA increases from other precursors. We examine the precursor-resolved vapor wall loss effect in more detail in Sect. 3.4. For the Base simulations, the aS-SOA mass concentrations were a factor of 10 and 2 lower than the aV-SOA mass concentrations for the low- and high-NOx parameterizations respectively.

IVOC. Differences in the SVOC and IVOC simulations were used to determine the influence of including SOA formation from IVOCs. For both the low- and high-NOx simulations, IVOCs contributed marginally to the aI-SOA mass concentrations in Los Angeles (∼0.045 µg m-3) and elsewhere too (see Figs. S3 and S4). The aI-SOA mass concentrations were about half of the aS-SOA mass concentrations for both the low- and high-NOx simulations. When compared to the aV-SOA mass concentrations, the aI-SOA mass concentrations were slightly lower for the high-NOx simulations (∼35 %) but about a factor of 3.3 lower for the low-NOx simulations. The inclusion of vapor wall losses seemed to make aI-SOA as or more important than aS-SOA but still less important than aV-SOA; the aI-SOA mass concentrations were a factor of 3.2 and 2.9 lower than the aV-SOA mass concentrations for the Base simulations for the low- and high-NOx simulations respectively. Our simulations imply that IVOCs might be as influential as SVOCs as a bulk class of SOA precursors, but they were still less important than the traditional SOA precursors (that included long alkanes and aromatics) in contributing to ambient SOA levels. In this work, the IVOC contribution to SOA was smaller compared to that from traditional SOA precursors mostly because IVOC emissions were only about a third of the traditional SOA precursors (see Sect. 3.1 for details on emissions). So although IVOCs have higher SOA yields than most of the traditional SOA precursors, the significantly lower IVOC emissions more than offset the increased SOA formation from higher yields. While there are exceptions (e.g., Tsimpidi et al., 2010; Jathar et al., 2017a), our results did not align with previous box (e.g., Dzepina et al., 2009; Hayes et al., 2015; Ma et al., 2017) and 3-D (e.g., Bergström et al., 2012; Zhang et al., 2013) modeling literature that has found IVOCs to be similar to or more important than traditional SOA precursors in contributing to ambient SOA levels. Below we discuss three main reasons for this inconsistency.

First, some previous estimates of IVOC emissions are likely to be less representative of the in-use gasoline- and diesel-powered sources and unconstrained for biomass burning sources. IVOC emissions in most atmospheric models have previously been determined by scaling emissions of POA or by calculating partitioning with the measured POA, with scaling factors typically on the order of 1.5 (e.g., Shrivastava et al., 2008) but as large as 3 (e.g., Dzepina et al., 2009). These factors have been calculated from emissions data from two medium-duty gasoline vehicles built more than two decades ago and a POA volatility distribution from a small off-road diesel engine (Robinson et al., 2007). Additionally, since POA is semi-volatile the POA mass in the particle phase will change with OA loading, which can complicate the use of a scaling based on POA (but this is addressed by the partitioning method used in some studies). Zhao et al. (2015) provided some evidence for this where they found that the POA-based scaling did not work that well for modern diesel vehicles and instead recommended the use of an NMOG-based scaling. We note that Ma et al. (2017) used the IVOC estimates of Zhao et al. (2015) and still found IVOCs to be comparable to VOCs in terms of SOA production in the Los Angeles area. Second, the SOA formation from IVOCs in most models to date has not been experimentally constrained. Most schemes to model SOA formation from IVOCs have relied on an ad hoc aging scheme where IVOCs and their oxidation products react with the OH radical to form lower-volatility products with ultimate SOA yields of 100 % (Robinson et al., 2007). These schemes do not account for fragmentation reactions and have not been comprehensively validated against experimental data. Jathar et al. (2016) showed that such schemes may significantly overestimate the net aerosol production from SOA precursors. Finally, most models use SOA parameters that do not account for the effect of vapor wall losses in chamber experiments. This effect and its particular influence on the IVOC contribution to SOA is discussed in Sect. 3.4. In this work, we (i) rely on a comprehensive set of IVOC emissions estimates made from measurements performed on more representative sources, (ii) model fragmentation reactions during IVOC oxidation, (iii) to some degree constrain SOA formation from IVOCs with chamber experiments, (iv) to some degree account for the influence of vapor wall losses in chamber experiments, and (v) include all of the previously mentioned updates in a chemical transport model. Hence, we argue that our findings on the IVOC contribution to SOA might be more robust than those modeled in earlier studies.

Traditional VOCs. For the Base simulations in Los Angeles, aromatics accounted for 33 % of the total SOA in Los Angeles and were the most important anthropogenic precursor of SOA. Alkane contributions to SOA were less than 10 % for both the low- and high-NOx simulations. Biogenic VOCs accounted for 46 % and 55 % of the total SOA for the low- and high-NOx simulations respectively and were clearly the most important precursor of SOA at the central Los Angeles site. After accounting for the influence of NOx based on Eq. (2), the isoprene, monoterpene, and sesquiterpene contributions to bV-SOA were 23 %, 68 %, and 9 % respectively, suggesting a strong monoterpene contribution to SOA in southern California. As biogenic VOCs react very quickly with OH and O3 (chemical lifetimes of a few hours), most of the biogenic SOA at this site was likely formed outside the urban airshed and transported to this location, as suggested by Kleeman et al. (2007), Hayes et al. (2015), and Heo et al. (2015).

SOA parameterizations that accounted for the influence of vapor wall losses in chambers seemed to have had a large effect on the absolute mass concentrations of SOA. This can be seen by comparing model results between the IVOC and Base simulations in Fig. 3. The SOA mass concentrations were enhanced by a factor of 10.1 and 2.6 for the low- and high-NOx simulations respectively and consistent with previous 3D simulations (Cappa et al., 2016). However, they were slightly higher than the range of enhancements reported by Zhang et al. (2014) and estimated by Krechmer et al. (2016) based on analyses of chamber data. The SOA enhancements resulted in an OA enhancement of 1.66 and 1.14 in the low- and high-NOx simulations, which were lower than the SOA enhancements since SOA only accounted for a fraction of the OA mass. Differences in enhancements in the low- and high-NOx simulations suggest that the vapor wall loss effect was modified by the NOx level where the enhancement may be lower in urban source regions with higher NOx but higher in rural/remote continental regions with lower NOx. Since urban SOA mass concentrations are usually higher than those in rural/remote continental regions, an implication of this NOx-modified enhancement is that accounting for vapor wall loss artifacts will tend to reduce gradients in SOA mass concentrations between urban and rural/remote continental regions and make SOA more of a regional pollutant similar to ozone (O3).

Different precursors contributed in varying degrees to the SOA enhancement. The precursor-resolved enhancements are visualized in Fig. 4 where we plot the ratio of the 14-day-averaged model predictions of the SOA mass concentrations from the Base simulation to those from the IVOC simulation for each grid cell in the southern California domain (dots) and overlay box–whisker plots based on those data. For all precursors the enhancements were higher for the low-NOx simulations compared to the high-NOx simulations. SVOCs showed the smallest enhancement at both the low and high NOx levels (median of 1.6 and 1.2) and hence their fractional contribution to total SOA was reduced in the Base simulation when compared to the IVOC simulation. Alkanes showed the largest enhancement in the low-NOx simulations (median of 94) and the second largest enhancement in the high-NOx simulations (median of 4.5). Despite the large enhancements, alkanes still contributed marginally to total SOA in the Base simulations because the baseline contribution of alkanes to SOA was small in the IVOC simulations (< 3 %). IVOCs exhibited a larger enhancement (median of 17 and 2.9) compared to SVOCs and a smaller enhancement compared to alkanes in both simulations, despite using the same surrogate (i.e., n-dodecane) to model SOA formation. The reason for varying enhancements in SVOC, IVOCs, and alkanes, despite using the same surrogate (i.e., n-dodecane), was that the vapor-wall-loss-related enhancement was inversely related to the carbon number where larger carbon number precursors (e.g., SVOC that had an average carbon number of 18 to 20) showed smaller enhancements and smaller carbon number precursors (e.g., alkanes that included species between carbon numbers of 6 to 12) showed larger enhancements. The simplest explanation for this inverse relationship is that larger precursors and their oxidation products, relatively speaking, have shorter chemical lifetimes and undergo fewer chemical reactions before condensing, which make them less susceptible to being lost to the walls (see Fig. S5 where we plot the vapor-wall-loss-related enhancement in SOA yields as a function of the carbon number at an OA mass concentration of 9 µg m-3). Of the two other important precursors, aromatics displayed the largest enhancement in the high-NOx simulations (median of 6.6) and were tied with IVOCs for the second largest enhancement in the low-NOx simulations (median of 16) while biogenic VOCs showed the lowest enhancement after SVOC in both the low-NOx and high-NOx simulations. Accounting for vapor wall loss artifacts is expected to result in an increase in the aromatic contribution to SOA when compared against biogenic VOCs. Vapor wall loss rates in Teflon chambers might be much higher (∼ factor of 5) than those used in this work to develop the SOM parameterizations (Huang et al., 2018; Krechmer et al., 2016; Sunol et al., 2018), the use of which will tend to increase SOA mass concentrations even further. This new understanding will need to be considered in the future.

Results from the sensitivity simulations that examined uncertainties in select model inputs are shown in Fig. 5 where we plot the 14-day-averaged model predictions from these simulations at the central Los Angeles site. We also plot model predictions from the Base simulations as all the sensitivity simulations have been performed using the Base simulation as the reference (see Table 3 for details about the simulations). Model predictions from the low- and high-NOx simulations are shown separately. The No Aging simulations decreased the SOA mass concentrations by almost an order of magnitude, demonstrating the importance of modeling multigenerational aging in the SOM. The inclusion of oligomerization reactions that may enhance the partitioning of semi-volatile species may alter this finding. The No Aging simulations produced a very different precursor contribution to total SOA compared to the Base simulations and the changes in the precursor contribution were also different between the low- and high-NOx simulations. For instance, the aV-SOA contributions to total SOA increased from 39 % to 41 % for the low-NOx simulations but decreased from 26 % to less than 5 % in the high-NOx simulations. This implied that the treatment of multigenerational aging in the SOM did not proportionately enhance the SOA mass concentrations from the different precursors but rather produced varying levels of enhancement for the different precursors that were further modified by the NOx levels. This finding is of note because chemical transport models that have employed schemes such as the volatility basis set (VBS) have typically assumed that multigenerational aging has an approximately similar effect on SOA mass concentrations from different precursors, regardless of the NOx levels, and one which does not significantly change the precursor contribution to SOA (Robinson et al., 2007). With the VBS, one may observe some differences with multigenerational aging from the use of different starting VBS distributions for SOA from different precursors.

The SVOCmax simulations that assumed all POA (except marine POA) to be semi-volatile saw POA mass concentrations decrease by 36 % compared to the Base simulations and by 56 % compared to the Traditional simulations (not shown here but inferred from results in Fig. 3). The increase in SVOCs from the additional evaporation of POA mass resulted in about a 3-fold increase in the aS-SOA mass concentrations and a proportionate increase in the SVOC contribution to total SOA. Similar to the findings discussed in Sect. 3.3, only a fraction of the evaporated POA mass lost was regained as aS-SOA mass concentrations. For instance, when compared to the Traditional simulations, of the 2.9/3.3 µg m-3 of POA mass lost 0.32/0.22 µg m-3 was regained as aS-SOA reflecting a chemical conversion efficiency of 11 %/7 % for the low-/high-NOx simulations. These simulations predicted the maximum decrease in POA mass concentrations from treating all POA as semi-volatile and reactive, but the results still found POA to be 40 % and 69 % of the total OA in the low- and high-NOx simulations respectively. Direct emissions of POA were still a sizeable fraction of the ambient OA and PM burden using the current state-of-the-science treatment.

Estimating IVOCs to be 20 % of the NMOG emissions for all combustion sources and modeling the SOA formation from IVOCs using a C15 linear alkane – as modeled in the IVOCmax simulations – resulted in an approximately 4-fold increase in the aI-SOA mass concentrations over the Base simulations. The increases were partly attributed to additional IVOC emissions from sources other than mobile and biomass burning (factor of 2.8 compared to IVOC emissions from the Base simulations) and partly to using a larger alkane (C15 linear alkane) with a higher SOA mass yield to model SOA formation from IVOCs emitted by gasoline sources. Simulating SOA formation from IVOCs using an aromatic surrogate in the S-IVOCaromatic simulations had the same effect as the IVOCmax simulations and increased aI-SOA mass concentrations by a factor of 2.6 and 6.3 for the low- and high-NOx simulations respectively. The aI-SOA mass concentrations were higher because aromatics for the same carbon number have a higher SOA mass yield than alkanes. The IVOCmax and S-IVOCaromatic simulations potentially present an upper-bound contribution of IVOCs to SOA formation, and in both these simulations they were ∼30 % of the total SOA and a factor of ∼1.5–2 larger than the aromatic VOC contribution. While the IVOCmax and S-IVOCaromatic simulations dramatically increased the aI-SOA mass concentrations, these simulations only modestly increased the total OA mass concentrations over the low- and high-NOx simulations (average increase of 10 %). Over the urban area, the OA mass concentrations in the IVOCmax and S-IVOCaromatic simulations were on average 10 %–12 % higher compared to the Base simulations (see Fig. S6). Updating the emissions profiles based on the work of May et al. (2014) had a negligible effect on the SOA mass concentrations and its precursor contribution, implying that the emissions profiles from more than a decade and a half ago may be sufficient to model the modern mobile source fleet. Finally, a lower-volatility (i.e., more realistic) POA in the SVOCcooking simulations, informed by the measurements of Louvaris et al. (2017), resulted in a 20 % increase in POA mass concentrations when compared to both the low- and high-NOx Base simulations. POA mass concentrations in these low- and high-NOx simulations accounted for approximately 55 % and 85 % of the OA respectively. The SOA mass concentrations between the SVOCcooking and Base simulations remained the same.

The SOM currently does not model the continuous evolution of SOA under varying NOx concentrations. One of the challenges in modeling the NOx influence on SOA formation has been in quantifying the branching of the VOC oxidation under low- and high-NOx conditions. Most commonly used schemes in atmospheric models use the NO:HO2 ratio to determine the initial branching of the precursor to form SOA via the low- or the high-NOx pathway. However, this scheme depends on an accurate prediction of NO and HO2. To assess, at least qualitatively, the ability of the model to capture NO and HO2 concentrations, we compare 14-day-averaged diurnal profiles from this work to those measured in Pasadena in 2010 during the CalNex campaign in Fig. S7. We found that the model predictions were within a factor of 2 for NO concentrations but were about a factor of 10 lower than the measured HO2* concentrations. We should note that the HO2* measurements included HO2 and a fraction of RO2 radicals, where RO2 radicals contributed to ∼30 % of the HO2* measurements (Griffith et al., 2016). The inclusion of RO2 should not change the findings reported here. If the results from our modeling are representative of results from other atmospheric models that use SAPRC or other gas-phase chemical mechanisms, underestimating the HO2 concentrations may lead NO:HO2-ratio-based schemes to overestimate the SOA formed via the high-NOx pathway. Given this limitation and the fact that the SOM does not model the continuous evolution of SOA under varying NOx concentrations, we attempted to model the NOx-dependent SOA formation using VOC:NOx ratios and NOx concentrations.

Four different methods – described in Eqs. (1) through (4) – were used to adjust the SOA mass concentrations from each individual precursor to account for the influence of NOx. To remind the reader, Eqs. (1) and (2) assume a linear and logarithmic dependence respectively between the SOA mass concentration and the VOC:NOx ratio. Equations (3) and (4) assume a linear and logarithmic dependence respectively between the SOA mass concentration and the NOx concentration. The adjusted SOA mass concentrations, referred to as SOAeff, were summed to calculate the total SOA mass concentrations. Equation (2) produced the highest SOA mass concentrations while Eq. (3) produced the lowest SOA mass concentrations amongst the four equations. Scatter plots comparing the SOA mass concentrations calculated using Eq. (2) to those calculated using other equations, in Fig. S8, show that the SOA mass concentrations based on Eq. (2) were, on average, a factor of 1.27, 3.19, and 1.92 higher than those with Eqs. (1), (3), and (4) respectively. This meant that a calculation based on the VOC:NOx ratio produced a stronger response of NOx on SOA mass concentrations than the NOx concentrations themselves. In the subsequent sections, where we evaluate the model predictions (Sect. 4) and predicted future changes in the OA burden (Sect. 5), we used the SOAeff calculations based on Eq. (2) since they represented an upper-bound estimate of the NOx effect on SOA mass concentrations. The validity of Eq. (2) needs to be examined in future work.

In Fig. 6, we plot the ratio of the total SOAeff mass concentrations based on Eq. (2) to the total SOA mass concentrations from the (a) high-NOx and (b) low-NOx Base simulations. The SOAeff mass concentrations were higher than the SOA mass concentrations predicted using the high-NOx parameterizations, with an average factor of 2 increase in urban areas and a maximum factor of 4 increase in non-urban areas. This was because the model-predicted VOC:NOx ratios in the urban areas were higher than the VOC:NOx ratios produced in the high-NOx chamber experiments, and based on Eq. (2) the SOA mass concentrations were adjusted upwards to include the SOA predicted using the low-NOx parameterizations. The adjustments increased the SOA mass concentrations because the SOA mass concentrations from each precursor were universally higher with the use of the low-NOx parameterizations compared to the high-NOx parameterizations. The SOAeff mass concentrations were 30 %–40 % lower than the SOA mass concentrations predicted using the low-NOx parameterizations in urban areas, suggesting that the SOAeff mass concentrations were approximately midway between the SOA predictions using the high- and low-NOx parameterizations. In contrast, the SOAeff mass concentrations were only marginally lower (10 %–20 %) in the non-urban areas, implying that the VOC:NOx ratios in these regions were very similar to the VOC:NOx ratios produced in the low-NOx chamber experiments. In summary, a modest fraction of the SOA mass may be formed through the low-NOx pathway in high-NOx urban areas, which may result in substantial increases in the predicted SOA mass concentration when compared against predictions purely based on the use of high-NOx parameterizations. This low-NOx SOA will continue to increase in the future as NOx concentrations are reduced in urban areas through controls on mobile sources. In contrast, only a small fraction of the SOA mass may be formed through the high-NOx pathway in low-NOx non-urban areas, and the use of a low-NOx parameterization in these regions will only marginally bias model predictions of SOA mass concentrations.

In Fig. 7a, we compare 14-day-averaged model predictions of aromatic concentrations for our 2005 episode against measured temporal trends in summer-averaged single-ring aromatic concentrations at three different sites in southern California (Los Angeles-North Main Street, Riverside-Rubidoux, and Long Beach) (SCAQMD, 2017); model predictions of aromatic concentrations are a sum of the benzene, ARO1, and ARO2 concentrations. On the same figure, we also plot model predictions of aromatic concentrations at Pasadena for our 2005 episode and measured single-ring aromatic concentrations made at the Pasadena ground site in 2010 as part of the CalNex campaign (Zhao et al., 2014). The summertime single-ring aromatic concentrations in southern California have decreased by a factor of 2 to 3 between 2000 and 2011 presumably from regulations that have targeted emissions from mobile sources. These reductions agreed well with reported temporal trends in carbon monoxide, nitrogen oxides, and non-methane organic compounds for Los Angeles over the same time period (Warneke et al., 2012; McDonald et al., 2013). Aromatic measurements at Pasadena in 2010 compared well with the 2010 measurements made ∼12 km southwest of Pasadena at the Los Angeles-North Main Street location, suggesting that the summer/campaign-averaged aromatic concentrations were spatially homogeneous over urban Los Angeles, and findings from the model–measurement comparison at a particular site could be generalized for the larger modeled domain. The model–measurement comparison for aromatics in 2005 was mixed. Concentrations were overpredicted by a factor of ∼1.5 at the Los Angeles-North Main Street and Long Beach sites but agreed well with measurements at Riverside-Rubidoux. The predictions might have been overestimated because we were using an older emissions inventory developed for the year 2000 but adapted for use for the year 2005 based on activity data (Hu et al., 2015). Another possibility for the overprediction was that the lumped model species ARO1 and ARO2 in SAPRC-11 also included emissions from oxygenated aromatic (e.g., phenols) and aromatic-like compounds (e.g., furans) while the measurements were limited to a handful of single-ring reduced aromatic compounds. Despite differences in the absolute concentrations, the model seemed to capture the measured spatial differences between the three sites; i.e. Los Angeles-North Main Street > Riverside-Rubidoux > Long Beach.

In Fig. 7b, model predictions of aromatics and IVOCs in Pasadena in 2005 are compared against measurements made at the Pasadena ground site in 2010. The model predictions in Pasadena were calculated by averaging predictions from the grid cell that contained the Pasadena ground site and the grid cell immediately to the south. This was done because the ground site location was very close to the cell boundary to the south and the grid cell containing the Pasadena ground site included mountains to the north of Pasadena that tended to dilute the concentrations in that grid cell. The measurements in Fig. 7b included primary IVOCs but did not include the oxygenated IVOCs measured by Zhao et al. (2014) since the primary IVOCs, according to the authors, relate most closely to IVOC emissions from mobile sources. The IVOCs included in this work were mostly (> 95 %) from mobile sources (see Fig. 1) and hence the comparison with primary IVOCs was appropriate. The model-predicted aromatic concentrations at Pasadena in 2005 were twice the measured aromatic concentrations at Pasadena in 2010. This 2005 (modeled) to 2010 (measured) ratio was slightly higher but still consistent with the measured 2005-to-2010 ratio in aromatic concentrations at the Los Angeles-North Main Street site (1.67). That the 2005 (modeled) to 2010 (measured) ratio for IVOCs in Pasadena was ∼1.0 is some evidence that the model predictions of IVOCs might be underpredicted in 2005, assuming that the ambient IVOC-to-aromatic ratio did not change between 2005 and 2010. The IVOCmax sensitivity simulation (the only sensitivity simulation that modeled an increase in IVOC emissions) predicted a 2005 (modeled) to 2010 (measured) ratio of 3.15 for IVOCs in Pasadena, which was closer to the measured aromatic concentrations ratios between 2005 and 2010 at the Los Angeles-North Main Street site. This provides additional evidence for higher IVOC emissions to be included in the model. While this model–measurement comparison validates the aromatic SOA precursors and to some extent the mobile source IVOC SOA precursors, our model does not account for the oxygenated IVOCs that Zhao et al. (2014) measured and we recommend that future work investigate the sources, composition, and the SOA potential for these IVOCs.

Scatter plots comparing model predictions of OA from the Base simulations to (a) CSN and (b) IMPROVE measurements in southern California are shown in Fig. 8a and b. Predictions from the low- and high-NOx simulations are presented in grey while predictions accounting for the influence of NOx are shown in color. The colors denote different sites and the site locations are shown in Fig. 8c. The model–measurement performance is also captured using statistical metrics of fractional bias, fractional error, and the coefficient of determination in Table 4. At all CSN sites, model predictions of OA that included SOA mass concentrations adjusted for the influence of NOx were in between those predicted between the low- and high-NOx simulations. As explained earlier, this was because the VOC:NOx ratios at all these sites (see Fig. S9a) were always higher than those in the high-NOx chamber experiments (see Table 2), and hence the SOA mass concentrations calculated using Eq. (2) were always higher than those predicted in the high-NOx simulations. At all the CSN sites, correcting for NOx improved model performance compared to the high-NOx experiments but was still inferior compared to the predictions from the low-NOx simulations (see Table 4). The mean predicted OA mass concentration across all the CSN sites was about 30 % lower than the measurements (5.96 vs. 8.86 µg m-3). Model predictions of OA were very similar to those predicted in the low-NOx simulations at the IMPROVE sites where the VOC:NOx ratios were higher (e.g., San Rafael, green square). But, similar to the finding at the CSN sites, model predictions of OA were in between the predictions between the low- and high-NOx simulations at the IMPROVE sites where the VOC:NOx ratios were lower as a result of their proximity to urban areas (e.g., Agua Tibia, blue square; and Riverside, brown square). Accounting for NOx seemed to improve the model performance at the IMPROVE sites when compared to predictions from the high-NOx simulations and they were slightly inferior to those from the low-NOx simulations (see Table 4). Of the 27 IMPROVE measurements available for comparison, 22 or ∼80 % of the model predictions corrected for NOx were within a factor of 2 of measurements with little bias (fractional bias = -16.63 %). The model skill, captured by the R2 values, for all model simulations at both the CSN and IMPROVE sites was quite poor but still slightly better than that found in earlier work for the southern California region with the CMAQ model (Baker et al., 2015). However, the model skill was much worse than that reported in earlier work with CMAQ (e.g., Murphy et al., 2017) and WRF-Chem (e.g., Ahmadov et al., 2012) over regions other than southern California, suggesting that there might be missing emissions sources and/or chemical pathways or meteorological considerations that contribute to the poor model skill in southern California.

Given the differences in the model–measurement comparison between the CSN (or urban) and IMPROVE (rural/remote continental) sites, the underprediction at the CSN sites might be indicative of a missing urban source or pathway of OA formation. Recently, McDonald et al. (2018) found that volatile chemical products such as pesticides, coatings, cleaning agents, and personal care products may contribute substantially to IVOC emissions and account for more than half of the anthropogenic SOA formation in southern California. Our underprediction at urban sites might be evidence of missing SOA from volatile-chemical-product-related IVOC emissions. However, it is also possible that the urban versus rural/remote continental difference is an artifact of how the SOM models the oxidation chemistry and/or accounts for the influence of vapor wall losses. Within the CSN and IMPROVE sites, we did not find the model–measurement comparison to vary systematically by location. The model–measurement comparison over all of California using the 24 km simulations produced a similar result (Fig. S10).

Model predictions of the OA:ΔCO diurnal profile and daytime OA versus CO (between 10:00 and 20:00 local time) are compared against measurements made at the Riverside site during the SOAR-1 campaign in Fig. 9a and b; SOA mass concentrations have been adjusted for the influence of NOx using Eq. (2). The ΔCO for the measurements was calculated by assuming a background concentration of 105 ppbv (Hayes et al., 2013) while the ΔCO for the model predictions was calculated by using the model-predicted background concentration of CO over the ocean to the west of Los Angeles. This model–measurement comparison was not completely coincident in time since the model results were between 20 July and 2 August while the SOAR-1 campaign spanned from 15 July to 15 August. The measurements did not point to any substantial differences in results between the coincident and non-coincident time, and hence we did not anticipate any issues in our comparisons here. The model predictions were able to capture the general trends in the measured diurnal profile in Fig. 9a with low ratios during the night, high ratios attributed to photochemistry in the mid-afternoon, and a peak between 13:00 and 14:00 (local time). However, the modeled OA:ΔCO ratios at all times in the diurnal profile in Fig. 9a and the slope of the OA:CO ratios in Fig. 9b were approximately a factor of 2 to 3 lower than the measured ratios, indicating a significant underprediction of urban SOA, which was consistent with the much higher POA / SOA ratios predicted by the model compared to the observations, as discussed above. This underprediction cannot be blamed on the model grid resolution since a ratio with CO should to first order account for the influence of dilution in the grid cell. Cappa et al. (2016) showed much better model performance than this work when they assumed a non-volatile POA and SOA formed under low-NOx conditions. In this work, despite forming additional SOA from SVOCs and IVOCs, the evaporation of the POA mass and an SOA estimate adjusted for NOx meant that the model performance was worse in comparison to Cappa et al. (2016). The sensitivity simulations of IVOCmax and S-IVOCaromatic produced slightly higher OA mass concentrations (∼10 %–15 %) compared to the Base simulations but not dramatically different to influence the comparison in Fig. 9a and b. As mentioned earlier, SOA formation from IVOC emissions from volatile chemical products, or other future improvements in the SOM, has the potential to reduce the model underprediction at Riverside during the SOAR-1 campaign.

The 14-day-averaged results predicted POA and SOA mass concentrations of 3.4 and 2.2 µg m-3 and an approximate 60:40 POA–SOA split at Riverside. Docherty et al. (2011) estimated average POA and SOA mass concentrations of 1.9 and 7.0 µg m-3 and a POA–SOA split of 20:80 at Riverside during the SOAR-1 campaign. On an absolute basis, model predictions of POA mass concentrations were overpredicted by ∼80 %. A sensitivity simulation that turned sea spray emissions off suggested that the 14-day-averaged marine POA mass concentrations at Riverside were ∼0.8 µg m-3, which are very likely to be overestimated (Hayes et al., 2013). If the emissions of marine POA were updated to align better with the observations and in the limiting case where the marine POA mass concentrations at Riverside were negligible, model-predicted POA mass concentrations at Riverside (3.4-0.8=2.6 µg m-3) would compare well with the measured values (1.9 µg m-3). As the POA mass concentrations in the SVOCcooking simulations increased and the SOA mass concentrations remained the same compared to the Base simulations, a low-volatility and more realistic treatment of the POA from food cooking sources increased the discrepancy in the modeled and measured POA : SOA ratio at Riverside. It is also possible that the model might be overpredicting POA because we only considered POA from certain sources (gasoline and diesel use, woodsmoke, and food cooking) to be semi-volatile.

Figure 1 shows that more than half of the partitioned POA (that excludes marine POA) in southern California belonged to other sources (e.g., road and construction dust) and this POA was treated as non-volatile in the Base simulations. Model predictions from the SVOCmax simulations that treated all POA except marine POA as semi-volatile predicted a 14-day-averaged POA mass concentration of 2.1 µg m-3, which was much closer to the measured value of 1.9 µg m-3. This suggests that all POA, regardless of source, might be semi-volatile and could be modeled so in atmospheric models. While these results are in better agreement with measurements, PM2.5 from road and construction dust sources is not created in a high-temperature process followed by rapid cooling and so it is unknown whether the POA portion in it would evaporate with atmospheric dilution. We also compared the hydrocarbon-like OA estimate from the measurements, which was more representative of POA from mobile sources, against model predictions of POA from mobile sources. We did not model POA from mobile sources separately, but if we assumed that mobile sources only accounted for about a quarter of the partitioned POA mass in southern California (based on Fig. 1), our estimated Base model predictions of POA mass concentrations from mobile sources of 0.85 µg m-3 (=3.4×0.25) would compare reasonably with the measured HOA mass concentrations of 1.20 µg m-3.

On an absolute basis, SOA mass concentrations were underpredicted by a factor of 3 compared to measurements. Based on the discussion in the previous paragraph, if we added the non-mobile source POA to SOA, the net SOA mass concentration (3.4×0.75+2.2=4.75 µg m-3) was still 33 % lower than the measured value. The SOA mass concentrations in the IVOCmax simulations – sensitivity simulations that modeled a fixed IVOC:NMOG ratio of 20 % for all sources except biogenic sources, assumed IVOCs formed SOA similar to a C15 linear alkane, and which produced the maximum SOA mass concentrations amongst all the simulations – were 33 % higher than those in the Base simulation but still ∼60 % lower than the measured SOA mass concentration of 7 µg m-3. A combination of the two, i.e., adding the non-mobile source POA to the SOA formation in the IVOCmax simulations, resulted in a net SOA mass concentration that was only 22 % lower than the measured SOA value. Since the IVOCmax simulations produced ambient IVOC concentrations that were more in line with the measurement trends (see Sect. 4.1), it is likely that the IVOCmax simulations were better in predicting IVOC concentrations and their contribution to SOA. However, there are no bottom-up (i.e., source) or top-down (i.e., atmospheric) data to directly constrain the emissions of and SOA formation from IVOCs in the IVOCmax simulations, and hence this finding provides motivation for more detailed studies of IVOCs in the future.

The SOM tracks the carbon and oxygen numbers for the OA species, and hence we were able to compare model predictions of the diurnal profiles for the OA H:C and O:C ratios to measurements made at the Riverside site during the SOAR-1 campaign. The comparisons are shown in Fig. 9c and d. For the Base simulations (shown as orange box plots), model predictions of H:C were significantly overpredicted and those for O:C were significantly underpredicted although the predictions did capture dips in the H:C and the peaks in the O:C ratios in the mid-afternoon, coincident with peak photochemical activity. The model predictions did not capture the slight increase in H:C and the decrease in O:C in the early morning attributed to emissions from rush-hour traffic. The high H:C and low O:C predictions were a result of OA being dominated by POA (∼60 %), which in this work was modeled as a hydrocarbon distribution that had an H:C slightly larger than 2.0 and an O:C of 0. Docherty et al. (2011) found that POA had a campaign-averaged H:C of 1.92 and an O:C of 0.078. If the POA O:C ratios were fixed to the values estimated by Docherty et al. (2011), model predictions (shown as blue box plots) improved – as shown in Fig. 9c and d – but still over- and underpredicted the H:C and O:C, respectively; since SOM only tracks carbon and oxygen numbers for an organic species and determines the hydrogen number based on the remaining valence, specifying the O:C dictates the H:C. To assess the ability of the model to predict the elemental composition of SOA, we plot the diurnal profile of H:C and O:C of the SOA in Fig. 9c and d. Model predictions of SOA H:C and O:C (shown as green box plots) compared well with the measured range of values but did not reproduce the diurnal changes. Docherty et al. (2011) argued that the H:C and O:C of OA at Riverside was mostly controlled by the SOA composition, which did not change dramatically during the day, and was modified by POA at certain times when POA emissions dominated over SOA production (e.g., nights, rush-hour traffic). This suggests that if absolute predictions of the SOA mass concentrations and the POA–SOA splits were improved, our model would be able to predict both the magnitude and diurnal changes in OA H:C and O:C ratios. We found that the SOA H:C and O:C ratio predictions did not vary significantly and produced similarly flat diurnal profiles across a subset of sensitivity simulations performed (Fig. S11), suggesting that the modeled elemental composition of SOA was not very sensitive to the distribution of precursor contributions to SOA.