Jathar et al. (2017a) performed photooxidation experiments using an OFR to measure SOA production from the exhaust of a 4.5 L John Deere diesel engine. The stock engine met Tier 3 emissions standards for off-road diesel engines. The OFR used therein was described in detail by Friedman et al. (2016) and the experimental setup and OA measurements from these experiments were described in detail by Jathar et al. (2017a). We briefly summarize the experimental setup, measurements, and findings from Jathar et al. (2017a). The engine was run at two different loads (idle and 50 % load) with two different fuels (diesel and biodiesel) and with and without an aftertreatment system. The aftertreatment system included a diesel oxidation catalyst (DOC) to oxidize CO and total hydrocarbons (THC) and a diesel particle filter (DPF) to trap fine particles. Diesel exhaust was diluted by a factor of 45–110 before entering the OFR. The intensity of the mercury lamps (at wavelengths of 185 and 254 nm) inside the OFR was varied to produce different hydroxyl radical (OH) concentrations and simulate different photochemical exposures. The OFR had a residence time of 100 s. A suite of instrumentation was used to measure gas-phase (CO2, CO, THC, NOx, O2, oxygenated organic compounds) and particle-phase (aerosol size and composition) concentrations. A total of 13 experiments (see Table 1 for more details) were performed at varying engine loads and with varying fuels and aftertreatment configurations. The OH exposure was varied between 0 and a maximum of 9.2 × 107 molecules-h cm-3 (equivalent to 2 days of photochemical aging at an OH concentration of 1.5 × 106 molecules cm-3). On average, each experiment included measurements at six to seven different photochemical exposures. The mass concentrations and elemental composition of the POA (measured when OFR lights were off) and SOA (at varying OH exposures) were measured by a high-resolution aerosol mass spectrometer (HR-AMS). In addition to the measurements reported by Jathar et al. (2017a), the gas-phase concentrations of oxygenated organic compounds were measured by an acetate reagent ion-based chemical ionization mass spectrometer (CIMS) (Link et al., 2016). At all engine configurations, SOA production exceeded the POA emissions after the equivalent of a few hours of atmospheric photochemical aging. SOA production was particularly strong at idle (or less fuel-efficient) engine loads and/or when exhaust temperatures were low and proper functioning of the aftertreatment devices was limited. Further, POA emissions and SOA production were nearly identical between diesel and biodiesel fuels. A synopsis of experiments performed and the THC, which includes all SOA precursors CO, NOx, POA, SOA, O : C, OH, and size distribution, is presented in Table 1.

Although the diesel exhaust was diluted with clean air to produce atmospherically relevant concentrations of POA, the initial THC, CO, and NOx concentrations in the OFR were still quite high. Peng and Jimenez (2017), using a detailed gas-phase model, argued that the high external OH reactivity from high THC, CO, and NOx concentrations might lead to non-OH chemistry in the OFR and NO could quickly be consumed in the OFR leading to low NO conditions for SOA formation. Peng and Jimenez (2017) quantified the potential influence of NO on the oxidation chemistry by calculating the ratio of the reactive flux of the peroxy radicals with NO to the reactive flux of the peroxy radicals with HO2 (rRO2+NO/rRO2+HO2). A ratio greater than 1 was considered as “high NO” while a ratio less than 1 was considered “low NO”. For the relative humidity, photon flux, initial NO, and external OH reactivity values in Jathar et al. (2017a), the model of Peng and Jimenez (2017) predicted that the OFR mostly ran in a high NO mode at all photochemical exposures when the engine was run at load conditions or with an aftertreatment device in place. However, the model predicted that the OFR mostly ran in a low NO mode especially at the high photochemical exposures when the engine was run at idle conditions and without an aftertreatment device (i.e., idle–diesel–none and idle–biodiesel–none). The rRO2+NO/rRO2+HO2 ratio and low versus high NO mode for each photon flux and experiment combination is listed in Table S1 in the Supplement. Based on these results, we accordingly used the low and high NOx parameterizations to perform the model simulations.

The model of Peng and Jimenez (2017) also indicated that most of the experiments performed by Jathar et al. (2017a) were run under, in what Peng and Jimenez (2017) refer to as, “risky” or “bad” conditions. These conditions refer to situations in the OFR where the initial NO concentrations and external OH reactivity are high enough to suppress OH exposure and lead to non-tropospheric photolysis at 185 and 254 nm, which could compete with OH exposure to determine the fate of the SOA precursors and its oxidation products. Such conditions could be avoided by ensuring low initial NO concentrations and external OH reactivity that for combustion emissions would require substantial dilution with clean air before they are oxidized in the OFR. Future studies on combustion sources should be cognizant of this fact to avoid artifacts linked to non-tropospheric photolysis of organic compounds in OFRs.

In this work, we used two different OA models to predict the mass concentrations and chemical composition of SOA and compare predictions against the SOA measurements from Jathar et al. (2017a) and Friedman et al. (2017). In this section, we briefly describe the two model frameworks, namely the VBS and the SOM, used to simulate the coupled chemistry, thermodynamic properties, and kinetic gas–particle partitioning of OA. Neither model accounted for photolysis of organic compounds in the gas phase at 185 or 254 nm, which may need to be considered in the future when modeling the OFR chemistry from combustion emissions. The VBS model was chosen as it is widely used in contemporary air quality models; the SOM was chosen to examine the influence of improved representation of OA processes (e.g., fragmentation reactions) on model predictions.

Palm et al. (2016), Ahlberg et al. (2017), and Jathar et al. (2017a) have argued that the short residence times and small condensation sinks in the OFR may not permit all low-volatility products formed from precursor oxidation to condense onto preexisting aerosol. Hence, unlike earlier work that has assumed equilibrium partitioning to model SOA in OFRs (Tkacik et al., 2014; Chen et al., 2013), we modeled the kinetic gas–particle partitioning of OA using Eq. (7) (Zhang et al., 2014): dCipdt=2πDiDpNpFFSCig-CipCi*COA, where Cip is the particle-phase mass concentration for the ith organic species (µg m-3), Di is the gas-phase diffusion coefficient of the ith organic species (m2 s-1), Dp is the number mean particle diameter (m), Np is the total particle number concentration (m-3), FFS is Fuchs–Sutugin correction for non-continuum mass transfer, Cig is the gas-phase mass concentration of the ith organic species (µg m-3), Ci* is the effective saturation concentration of the ith organic species, and COA is the total OA mass concentration (µg m-3). The ith organic species refers to the organic compounds tracked in the VBS bins and the SOM grids. The gas-phase diffusion coefficient was calculated for each organic species as follows: Di=DCO2MWCO2MWi, where DCO2 is the gas-phase diffusion coefficient of CO2 (1.38 × 10-5 m2 s-1), MWCO2 (g mole-1) is the molecular weight of CO2, and MWi (g mole-1) is the molecular weight of the ith organic species. In the VBS model where we do not track the molecular composition of the SOA species, we assumed all condensing species to have a molecular weight of 200 g mole-1. This formulation to calculate the gas-phase diffusion coefficient underpredicted the measured gas-phase diffusion coefficients compiled by Tang et al. (2015) by ∼ 20 %. However, doubling the gas-phase diffusion coefficient calculated in Eq. (8) resulted in very small change (< 1 %) in the OA mass predictions for a representative experiment. Hence, we decided to use the formulation in Eq. (8) for the rest of this work. The Fuchs–Sutugin correction was calculated as follows: FFS=0.75α(1+Kn)Kn2+Kn+0.283⋅Kn⋅α+0.75α, Kn=2λiDp,λi=3DiCj,Ci=8NAkTπMWi, where Kn is the Knudsen number, α is the mass accommodation coefficient, λi is the mean free path of the ith organic species in air (m), Ci is the root mean square speed of the gas (m s-1), NA is Avogadro's number (molecules mole-1), k is the Boltzmann constant (m2 kg s-2 K-1), and T is the temperature (K).

The VBS and SOM models were run separately for each photochemical exposure simulated for each experiment listed in Table 1. In the VBS simulations, POA was tracked in one basis set while products from each SOA precursor were tracked in separate basis sets, allowing us to distinguish between POA and SOA. In the SOM simulations, all precursor molecules with the same surrogate (e.g., all n-alkanes) were tracked in the same SOM grid. Model simulations were performed in phases to answer specific questions and inform model inputs for later simulations:

To provide a general overview of the model predictions and model–measurement comparison, and to orient the reader to the results thereafter, we performed simulations with the VBS and SOM models using the base set of inputs for one of the idle–diesel–none experiments. Our base case included Profile 3161 for VOC emissions, 30 % IVOC mass fraction, kinetic gas–particle partitioning with a mass accommodation coefficient of 0.1, and monodisperse aerosol inputs based on measured data at no photochemical exposure. The partitioning- and IVOC-related choices for the base case are discussed in Sects. 3.2 and 3.3, respectively.

In Fig. 1, we compare predictions of OA from the VBS and SOM models using the base case to the measurements for the idle–diesel–none experiment performed on 5 June. Figure 1a and b compare predictions to the measurements in units of µg m-3 and g kg-fuel-1, respectively; hereafter we present all mass predictions in units of g kg-fuel-1. For this experiment, the VBS and SOM models overpredicted the OA mass by a factor of 2.0 and 2.2 at the lowest photochemical exposure (0.06 OH days) and a factor of 1.6 and 1.8 at the next highest photochemical exposure (0.17 OH days), respectively. The overprediction was because the models significantly overpredicted the SOA formation at these two photochemical exposures. For higher photochemical exposures (> 0.5 OH days), both models slightly underpredicted the OA mass but predictions were still within the measurement uncertainty. Our base case seemed to offer a mixed model–measurement comparison for this specific experiment (i.e., overprediction at lower photochemical ages and a slight underprediction at higher photochemical ages) because the 30 % IVOC mass fraction used in the base case was optimized to achieve a favorable model–measurement comparison across all experiments at all photochemical exposures. In other words, the overprediction in this experiment at lower photochemical exposures was probably offset by an underprediction at similar photochemical exposures for some of the other experiments. It is important to note that the model performance varied across the suite of experiments and this overall model performance is discussed in more detail in Section 3.3. The VBS and SOM models predicted that the OA at the maximum photochemical exposure was dominated by SOA produced from VOC and IVOC oxidation (92–93 %), which agreed well with the measured composition (see Fig. 1c). For the measurements, POA was defined as fresh OA while SOA was defined as OA formed in addition to the POA. Furthermore, both models suggested that most of the SOA emanated from the oxidation of IVOCs with only 8.6–14 % resulting from the oxidation of aromatic VOCs and less than 0.6–4 % resulting from alkane VOCs smaller than a C12. This dominance of IVOCs in explaining the photochemically produced SOA is in line with previous OFR and chamber studies that have modeled SOA formation from diesel exhaust (Tkacik et al., 2014; Zhao et al., 2015; Jathar et al., 2014b).

In Fig. 2, we plot predictions from the VBS and SOM models for the idle–diesel–none and idle–diesel–DPF + DOC experiments assuming instantaneous and kinetic gas–particle partitioning. The two different experiments were deliberately chosen to highlight the role instantaneous partitioning plays at the extremities. We found that for the idle–diesel–none experiment, the use of instantaneous partitioning roughly produced the same result as kinetic partitioning with α values of 0.1 and 1 and that all these predictions resulted in roughly the same model–measurement comparison. The instantaneous partitioning predictions were slightly higher than the kinetic partitioning predictions for the VBS simulations. The kinetic partitioning simulations (except for that with an α of 0.01) produced the same result as the instantaneous partitioning simulation most likely because the initial condensational sink was large enough (1.12 min-1) in this experiment that there were no kinetic limitations to partitioning. The increase in the condensational sink through condensation of SOA (10 min-1 at the highest photochemical exposure) tended to further reduce any differences in the predictions between the kinetic and instantaneous partitioning simulations. However, for the idle–diesel–DPF + DOC experiment, the instantaneous partitioning simulations predicted substantial OA mass at the lower photochemical exposures (0.04 and 0.12 OH days) compared to the kinetic partitioning simulations, specifically a factor of 9.8–29 larger at 0.04 OH days and a factor of 9.7–75 larger at 0.12 OH days for the VBS model and a factor of 3.9–5.8 larger at 0.04 OH days and a factor of 6.4–9.1 larger for the SOM. The instantaneous partitioning simulations consequently overpredicted the measurements while the kinetic partitioning simulations were more in line with the measurements. The instantaneous partitioning simulations predicted a lot more SOA because all condensable products of organic precursor oxidation were allowed to condense instantaneously (according to their respective volatilities) while the kinetic partitioning simulations predicted little SOA production because the initial condensational sink was quite small (0.002 min-1). Predictions from the instantaneous and kinetic partitioning simulations were much closer at the higher photochemical exposures because the SOA formed had grown the condensational sink enough to reduce limitations to partitioning (1 min-1 at the highest photochemical exposure). These results imply that the condensation of SOA in OFRs, in some instances, could be kinetically limited and that instantaneous partitioning may result in models overpredicting the condensation and formation of SOA.

We make two additional observations based on the results in Fig. 2. First, the initial condensational sink for the idle–diesel–none experiment was large (1.12 min-1) compared to condensational sinks one would encounter in the real atmosphere. For example, 5 µg m-3 of aerosol in a representative rural or remote environment will have a condensational sink < 0.05 min-1 (Seinfeld and Pandis, 2006). Therefore, modeling ambient applications of the OFR or OFR use with sources that use emissions control devices will need to be even more mindful of the instantaneous partitioning assumption while predicting SOA formation. Second, for the kinetic partitioning results, predictions from both models were relatively less sensitive to α values between 0.1 and 1 but were dramatically lower for an α value of 0.01 – more than a factor of 2 for the idle–diesel–none experiment and more than an order of magnitude for the idle–diesel–DPF + DOC experiment. Given the low sensitivity to α values greater than 0.1 and the reasonable model–measurement comparison at an α value of 0.1 and 1 at least for the idle–diesel–none experiment, we argue that the SOA condensation can be represented by an α value larger than 0.1 for the OFR experiments in this work. This α value for diesel exhaust SOA was consistent with prior estimates of the α value for biogenic SOA estimated from chamber, OFR, and aerosol heating experiments (Lee et al., 2011; Saleh et al., 2013; Karnezi et al., 2014; Palm et al., 2016) and direct measurements of α for alkanol SOA (Krechmer et al., 2017). However, an α of 0.1 was an order of magnitude higher than that observed recently for toluene SOA under dry conditions (Zhang et al., 2014). Model results presented hereafter include a kinetic treatment of gas–particle partitioning and assumed a mass accommodation coefficient of 0.1.

Results from model simulations performed using different initial condensational sink inputs, some of which captured the influence of new particle formation, are plotted in Fig. 3. We found that the initial condensational sink had no influence on the OA predictions from both models for the idle–diesel–none experiment, despite substantial differences in the initial condensational sink between the different cases. This was because the amount of SOA formed (920 µg m-3 at the highest photochemical exposure) was sufficient to grow the condensational sink enough that the initial condensational sink did not matter. In contrast, for both models we found large differences between the model predictions of OA for the idle–diesel–DPF + DOC experiment. The use of inputs based on the measurements at no OH exposure, where the aftertreatment system significantly reduced number concentrations (910 cm-3) and hence the available condensational sink (0.002 min-1), produced much less SOA (an order of magnitude lower or more) and poorer agreement with the measurements – see curve (i) in Fig. 3b. Initial condensational sinks that captured the influence of new particle formation resulted in higher model predictions but were still about a factor of ∼ 2 lower for the VBS simulations and a factor of ∼ 2.7 lower for the SOM simulations when compared against the measurements. The DPF + DOC results also suggest that calculating an initial condensational sink using data from before and after the photochemical exposure, as done by Palm et al. (2016), could be used as an input to model OFR data. Slight differences between the different curves for the idle–diesel–none experiment and curves (ii)–(iv) for the idle–diesel–DPF + DOC experiment can be attributed to the interaction of multigenerational aging and kinetic gas–particle partitioning.

In Fig. 4a, we compare predictions of SOA concentrations from the SOM against measurements for all the experiments listed in Table 1 and at all photochemical exposures. For visual clarity, we do not present results from the VBS model as both models had nearly identical predictions with a few exceptions. The panels in Fig. 4 show model–measurement comparisons assuming four different fractions of IVOCs: 0, 13.76, 30, and 60 % (from left to right); statistical metrics of fractional bias, fractional error, and R2 for the comparison for both models are listed in Table S4 (fractional bias = 1N∑i=1nM-OM+O2 and fractional error = 1N∑i=1n|M-O|M+O2, where M is the predicted value, O is the observed value, and N is the sample size). The model–measurement comparison and the model skill was very poor when no IVOCs were included (fractional bias = -109 %, fractional error = 125 %, and R2 = 0.52); this model reflects the treatment of diesel-powered sources in most traditional emissions inventories and large-scale models. The model–measurement comparison was reasonable with 13.76 % IVOCs (fractional bias = -46 %, fractional error = 101 %, and R2 = 0.95) but model predictions were overpredicted with 60 % IVOCs (fractional bias = 72 %, fractional error = 97 %, and R2 = 0.99). The optimal model performance that produced the lowest fractional bias and fractional error was realized at an IVOC mass fraction of 0.3 (fractional bias = 6 %, fractional error = 86 %, and R2 = 0.88). For predictions with 30 % IVOCs, 66 % and 70 % of the model predictions were within a factor of 1.5 and 2 of the measurements and IVOCs on average accounted for 67 and 72 % (VBS and SOM, respectively) of the SOA at the highest photochemical exposure across all experiments. Given the optimal performance, the base case used in this work assumed 30 % IVOCs. These comparisons indicate that it is critical that IVOCs be included when modeling the SOA formation from diesel exhaust and also validate the IVOC composition estimates made by Zhao et al. (2015). We note that the model of Peng and Jimenez (2017) suggested that the organic compounds in the OFR experiments performed by Jathar et al. (2017a) may have been subjected to non-tropospheric photolysis at 185 and 254 nm. Accounting for the photolysis of the key SOA precursors (IVOCs and aromatics) could affect the optimal IVOC fraction identified above and hence needs to be considered in future work.

We further investigated the IVOC species that contributed the most to SOA formation. For 30 % IVOCs, cyclic alkane IVOCs accounted for 23 % of the THC emissions and on average accounted for 59 and 53 % (VBS and SOM, respectively) of the SOA formation across the different experiments. We should note that the speciation of cyclic alkane IVOCs in Zhao et al. (2015), while robust in quantifying the carbon number, did not include any specificity in terms of the molecular structure; i.e., their methods would not be able to distinguish between a pure C10 cyclic alkane and a cyclohexane with a 4-carbon branch. Further, the parameterizations to model SOA formation from cyclic alkane IVOCs for both models were based on the behavior of particular compounds. In the VBS model when using the high NOx parameterizations, the surrogate for a cyclic alkane IVOC was determined through equivalence with a straight alkane IVOC, while in the VBS model when using the low NOx parameterizations or the SOM the cyclic alkane IVOCs were tied to parameterizations for hexylcyclohexane. (The observed SOA yield and derived SOM parameterization for hexylcyclohexane are actually quite similar to that for cyclododecane for low NOx conditions, but not for high NOx conditions; Cappa et al., 2013.) This lack of specificity in the speciation and the SOA parameterizations made the SOA predictions from the oxidation of cyclic alkane IVOCs relatively uncertain. To examine the sensitivity of the model predictions to uncertainties in the model treatment of cyclic alkane IVOCs, we performed simulations with both models for one of the idle–diesel–none experiments where the cyclic alkane IVOCs were treated as branched alkane IVOCs; results from these simulations are shown in Fig. 5a. The use of branched alkane IVOCs to model cyclic alkane IVOCs only marginally reduced OA predictions for both the VBS and SOM models, suggesting that the model predictions were not sensitive to the SOA parameterization used for cyclic alkane IVOCs. Regardless, we recommend that future work focus on a more detailed speciation of cyclic alkane IVOCs in combustion emissions as well as on chamber and OFR experiments on those speciated compounds to improve quantification of their SOA mass yields.

As there were no direct measurements of any SOA precursors in the study of Jathar et al. (2017a), we have used previously published emissions profiles for diesel exhaust to determine initial concentrations of the SOA precursors. We examined the sensitivity of model predictions to two different emissions profiles from the EPA SPECIATE (version 4.3) database: Profile 3161 (included in the base case) and Profile 8774 (emissions from heavy duty diesel exhaust); the speciation for both profiles is provided in Tables S2 and S3. Both profiles only included speciation for VOC emissions and in these simulations we assumed an IVOC mass fraction of 0.3. The results captured in Fig. 5b for one of the idle–diesel–none experiments show that the choice in the emissions profile had no influence on the OA evolution for the VBS model but had a small influence on the OA evolution for the SOM. This relatively small influence was expected given that most of the SOA was formed from IVOC, rather than VOC, oxidation. This further demonstrates that IVOCs, not VOCs, play an important role in controlling the SOA formation from diesel exhaust emissions and it is important that future studies work towards better understanding the IVOC speciation.

The IVOC speciation of Zhao et al. (2015) included 37 unique species, each of which required a unique surrogate to model the SOA formation from that species. Tracking these many IVOC species in an atmospheric model (e.g., global climate model) may be intractable, and hence there is a need to develop simplified parameterizations to efficiently model SOA formation from IVOCs. We note that species using the same surrogate in the VBS model (e.g., a C15 linear alkane, C17 branched alkane, and C13 cyclic alkane are all parameterized using n-pentadecane when using the high NOx parameterizations) could be lumped together to reduce the number of precursors and products tracked and that there are no penalties for a precursor type (e.g., n-alkanes) to include additional precursor and product species once a SOM grid is setup. Nonetheless, to investigate the possibility of developing a simplified parameterization, we modeled SOA from IVOCs assuming that all the IVOCs could be modeled together as a single linear C13, C15, C17, or C19 alkane; a similar strategy was employed by Jathar et al. (2014b) to model SOA formation from unspeciated organic compounds in combustion emissions. Results from these simulations are shown in Fig. 5c for one of the idle–diesel–none experiments. For the VBS model, the use of a larger carbon number alkane to model IVOC SOA produced increasingly more OA, with the C19 alkane providing the best comparison against measurements. For the SOM, the use of a C13 and C15 alkane produced good agreement with measurements with a C13 alkane slightly underpredicting the OA at 0.5 OH days and the C15 alkane slightly overpredicting the OA at lower photochemical exposures (0.06 and 0.17 OH days). It was interesting to observe that for the SOM, in contrast to the VBS, the use of different linear alkanes produced different OA masses at lower photochemical exposures but converged at the highest photochemical exposure, suggesting that the effective SOA mass yield in the SOM varied dynamically with photochemical age. Differences in the VBS and SOM predictions with different alkane parameterizations point to inherent differences in the coupled representation of multigenerational aging and gas–particle partitioning. Results from these simulations indicate that in cases where computational efficiency is demanded, the SOA formation from IVOCs in diesel exhaust could be modeled using a surrogate linear alkane, possibly a C19 linear alkane with the VBS and a C13 or C15 linear alkane for the SOM.

The SOM tracks both the carbon and oxygen number of the oxidation products, which allowed us to predict the O : C ratio of the OA. The O : C of the OA was calculated by combining the measured O : C of the POA with the modeled O : C of the SOA. We compare predictions of the O : C of OA from the SOM against measurements for all the experiments listed in Table 1 and at all photochemical exposures in Fig. 4; statistical metrics of fractional bias, fractional error, and R2 for the comparison are listed in Table S5. Model predictions for the no IVOC case, where the O : C of the OA was dominated by the O : C of the aromatic SOA, compared well with measurements (fractional bias = -4.2 %, fractional error = 28 %, and R2 = 0.77). However, the poor OA mass predictions with no IVOCs suggests that the good O : C performance was purely coincidental. The 13.76, 30, and 60 % IVOC cases underpredicted the OA O : C, where the underprediction appeared to increase as the IVOC influence increased (fractional bias = -32 %, fractional error = 38 %, and R2 = 0.72 for the 13.76 % IVOC case; fractional bias = -37 %, fractional error = 42 %, and R2 = 0.70 for the 30 % IVOC case; and fractional bias = -60 %, fractional error = -62 %, and R2 = 0.46 for the 60 % IVOC case). A higher IVOC fraction resulted in a lower O : C ratio because the IVOCs were primarily composed of higher carbon number species that on oxidation produced low O : C SOA compared to SOA formed from precursors such as aromatics. On average, the 30 % IVOC case predicted an O : C ratio that was 28 % lower than the measurements. For the three non-zero IVOC cases (13.76, 30, and 60 %), the model skill in predicting the O : C was much better for the non-DPF + DOC experiments (R2 = 0.82, 0.83, and 0.80, respectively) than for the DPF + DOC experiments (R2 = 0.02, 0.02, and 0.29, respectively). Measurements and model predictions of the OA O : C ratio from the 30 % IVOC case as a function of photochemical age are presented in Fig. S2.

The underprediction in O : C ratios was confounding when compared to earlier applications of the SOM and in light of the reasonable model–measurement comparison found in this work in predicting OA mass. We note that the low O : C in the 13.76, 30, and 60 % IVOC cases stems from the dominance of product species that have high carbon numbers and low oxygen numbers. We explored several lines of reasoning for this underprediction. First, Cappa et al. (2013) found good agreement between the SOM-predicted and observed O : C for chamber experiments conducted using individual linear, branched and cyclic C12 alkanes. Also, general predictions of the dependence of O : C on the carbon number of the parent hydrocarbon (cf. Fig. 2b in Cappa and Wilson, 2012) show good agreement with observations (cf. Fig. 2a in Tkacik et al., 2012), in terms of both absolute values and shape. This suggests that uncertainties in the SOM parameters may not be the dominant reason for the underprediction. A possible reason for the underprediction then is that the compounds identified by Zhao et al. (2015) as IVOCs are structurally different than the alkanes used to model them in this work. Second, the gas-phase chemistry in the OFR might be inherently different than that in a chamber. For example, kinetic limitations to gas–particle partitioning may result in gas-phase oxidation of low-volatility products having high O : C that typically would have partitioned to the particle phase in a chamber experiment but instead are fragmented (Palm et al., 2016). Why the chamber-based SOM parameters then offer good model performance on OA mass remains unclear. One way in which this issue could be addressed in the future is by developing SOM parameters exclusively based on OFR data, as and when they become available. Third, the SOM used here did not include heterogeneous and particle-phase reactions that might influence the OA composition and O : C ratio. When heterogeneous reactions of OA were included assuming an OH uptake coefficient of 1 (the product distribution from the oxidation reaction was kept the same as the gas-phase reactions), SOA production at the highest photochemical exposure for all the experiments was reduced, on average, by 7 % from fragmentation reactions within the particle phase, but the O : C ratio was only marginally increased (average of 2 %).

To understand the O : C underprediction better, we compared model predictions of normalized gas-phase species concentrations from the SOM to normalized gas-phase measurements made by Friedman et al. (2017) using a CIMS. The CIMS detects an array of oxygenated organic species and the high resolution of the time-of-flight mass spectrometer enables identification of the elemental composition of each detected peak. The CIMS data were aggregated by carbon and oxygen number to facilitate comparison with the SOM data. The comparison was performed on a normalized basis because the CIMS did not provide absolute concentrations for every detected peak. The SOM–CIMS comparisons for the idle–diesel–none and load–diesel–none experiments at the highest photochemical exposure are shown in Fig. 6, which highlight four findings of note. First, the CIMS measured species larger than a carbon number of 12 that are presumably products from oxidation of higher molecular-weight organic compounds, although the possibility of dimer formation in the instrument cannot be entirely ruled out. Nonetheless, this provides additional evidence for the presence of IVOC oxidation products in diesel exhaust emissions. Second, the CIMS measured organic compounds with high O : C ratios (e.g., C6O6, C7O7). This implies that the reaction chemistry in OFRs rapidly adds functional groups to the carbon backbone, although larger, less oxidized compounds could be simultaneously functionalized and fragmented in the CIMS leading to the appearance of highly oxidized species. Third, the SOM offered a reasonable correlation against the CIMS measurements for both experiments across a majority of the carbon–oxygen combinations that spanned more than 4 orders of magnitude. Qualitatively, this finding validates the statistical evolution of organic compounds tracked through the generalized SOM mechanism, although certainly some differences are evident. Finally, for the mid-carbon number species (∼ C10), the SOM seemed to produce higher fractions of species with low oxygen numbers (O0 to O3) but lower fractions of species with high oxygen numbers (O5 to O7). This underprediction of the high oxygen number species might potentially explain why the SOM may be underpredicting the OA O : C ratio. The SOM–CIMS comparison is preliminary and we intend to explore the implications of this comparison in future work.

We performed sensitivity analyses to examine the influence of other key processes on predictions from both the VBS and SOM models. When examining the sensitivity to each process, all the other inputs were kept the same as those listed in the base case. We only present sensitivity results for the idle–diesel–none experiment performed on 5 June, as the results for this experiment were generally representative of all experiments (Fig. 7). For completeness, we performed simulations for all the experiments at the highest photochemical exposure since each of the processes explored below manifested the strongest response at the highest photochemical exposure. The results from these simulations are presented as a change in the model predictions relative to that offered by the base case.