In this study, two versions of the CMAQ model (https://epa.gov/cmaq, last access: 6 November 2023), the official v5.3.2 (Appel et al., 2021) (available from 10.5281/zenodo.4081737, US EPA Office of Research and Development, 2020) and v5.4 with a community contribution (available from 10.5281/zenodo.7218076, US EPA Office of Research and Development, 2022), were used to perform nested simulations with horizontal resolutions of 36 and 12 km. The outer domain encompassed most of China, while the inner domain focused on eastern China (Fig. S1 in the Supplement). The simulation period covered 17 March–21 April and 29 September–21 November 2018, aligning with the campaigns conducted in Dongying (DY) and Guangzhou (GZ), respectively. Meteorological fields were generated using the Weather Research and Forecasting (WRF) model version 4.2.1. Anthropogenic emissions for China were represented using the 2018 high-resolution (0.25° × 0.25°) Multi-resolution Emission Inventory for China (MEIC) v1.4 (http://www.meicmodel.org, last access: 9 September 2023), while the Regional Emission Inventory in Asia (REAS) v3.2.1 (https://www.nies.go.jp/REAS/, last access: 12 September 2023) was used for the rest of Asia. Emissions from open burning were obtained from the Fire Inventory from NCAR (FINN) v1.5 (https://www2.acom.ucar.edu/modeling/finn-fire-inventory-ncar, last access: 13 January 2021), and biogenic emissions were calculated using the Model of Emissions of Gases and Aerosols from Nature (MEGAN) v2.1 (Guenther et al., 2012). The simulations utilized the SAPRC07tic gas-phase mechanism within CMAQ (Xie et al., 2013). The capabilities of WRF and CMAQ in simulating meteorological factors and major pollutants (NO₂, SO₂, O₃ and PM_2.5) were evaluated (Tables S1–S2 in the Supplement).

All the simulations modeled SOA formation from isoprene, glyoxal, and methylglyoxal, as well as NO₃-initiated oxidation of monoterpenes, using consistent parameterizations from CMAQ's official AERO7i module (see Fig. S2). Specifically, isoprene SOA formed via the aqueous uptake of isoprene epoxydiols (IEPOX)/methacrylic acid epoxide (MAE), and monoterpene-derived organic nitrates (ONs) via NO₃ oxidation were explicitly represented (Pye et al., 2013; Pye et al., 2015). Semi-volatile isoprene SOA, including contributions from non-aqueous pathways, was parameterized with the two-product model (Carlton et al., 2010), which utilized mass yields of two products (αi, i=1, 2) and their effective saturation concentrations (Ci∗, i=1,2) to simulate aerosol growth observed in experiments (Pankow, 1994). All the simulations accounted for multiphase SOA formation from glyoxal and methylglyoxal both in cloud (Carlton et al., 2008) and on wet aerosol surfaces (Pye et al., 2015).

Five simulations were conducted to evaluate model performance in reproducing OA mass concentrations, volatility distributions, and O/C ratios against observational data (Table 1). All simulations retained the same AERO7i heterogeneous chemistry and selected other systems (see above), and focused on different representations of key volatility-based systems. The simulation using the standard AERO7i treatment including SOA from anthropogenic precursors (e.g., aromatics, long-chain alkanes) (Qin et al., 2021), monoterpenes via O3/OH oxidation (Xu et al., 2018), and sesquiterpenes (Carlton et al., 2010) was referred to as a 1-D VBS for simplicity. In another simulation, the 2D-VBS framework (Zhao et al., 2015; Zhao et al., 2016; Chang et al., 2022) was used for anthropogenic VOCs, monoterpenes (excluding NO₃ oxidation pathways), and sesquiterpenes. Additionally, in 2D-VBS, the oxidation of POA (or L/S/IVOCs) was represented differently compared to the 1D-VBS simulations (Murphy et al., 2017).

The 1D-VBS simulation reflected the default CMAQ configuration without modifications to emissions or chemistry. The 1D-VBS_E simulation incorporated additional IVOC emissions, and updated emissions of S/LVOCs (historically classified as POA), based on MEICv1.4 (Sect. S1 in the Supplement). L/SVOC and IVOC emissions in each volatility bin were scaled from POA and total VOC emissions, respectively, using source-specific scaling factors (Table S3). Since the lowest volatility bin for emissions in the standard 1D-VBS parameterization was set to C∗=10-1 µg m⁻³, the difference between original POA emissions and L/SVOC emissions within the model-resolved volatility range (-1≤log⁡C∗(µgm-3)≤2) was reclassified as non-volatile (C∗=10-10 µg m⁻³, see Fig. S3) and treated with heterogeneous aging chemistry (Simon and Bhave, 2012). This adjustment conserved POA mass while addressing the model's limitations in volatility coverage. The estimated nation-wide L/SVOC and IVOC emissions, which were 3.18 and 6.68 Tg yr⁻¹, respectively, were higher than those reported by Zheng et al. (2023), but agreed well with Chen et al. (2024) in magnitude (Table S4, Figs. S4–S5), as both this study and Chen et al. (2024) applied a ratio-based methodology combined with the MEIC emission inventory. In contrast, Zheng et al. (2023) employed a different approach, using emission factors and activity data obtained from ABaCAS (Air Benefit and Cost and Attainment Assessment System) and Chang et al. (2022), which led to larger discrepancies relative to our estimates. The source contributions and volatility distributions of L/S/IVOCs differed slightly between studies. For example, our estimates indicated a lower contribution of solvent use to IVOCs compared with Zheng et al. (2023) (40 % vs. 57 %), and a higher contribution from residential sources to S/LVOCs (49 % vs. 30 %). In addition, the differences in emission magnitudes were primarily in the IVOC volatility range, whereas discrepancies in L/SVOCs were smaller. In the 1D-VBS_E simulation, the aging of L/SVOCs followed the POA treatment established in prior studies (Donahue et al., 2012; Murphy et al., 2017), where a fraction of L/SVOCs was oxidized in the gas phase to form SOA and the volatility distribution continually updated. The IVOC-derived SOA formation adopted the parameterization of Lu et al. (2020). In the 1D-VBS_EY simulation, SOA yields for benzene, toluene, xylene, and naphthalene oxidation were updated (Table S5) to account for vapor wall losses and formation of highly oxygenated organic molecules (HOMs) via autoxidation (Bilsback et al., 2023).

The 2D-VBS scheme (Zhao et al., 2015; Zhao et al., 2016; Chang et al., 2022), was implemented into CMAQv5.4 and evaluated using emission inputs largely consistent with the 1D-VBS_E simulation. The major differences in emissions included (1) reallocating non-volatile POA to the LVOC species with C∗ of 10⁻² µg m⁻³ (i.e., the least volatile primary emission category in 2D-VBS); and (2) mapping SVOC emissions with C∗ of 10⁻¹ µg m⁻³ to the C∗=10-2 µg m⁻³ bin as well (due to the absence of a corresponding volatility bin in the 2D-VBS scheme). L/SVOC aging products were classified as SOA. Initial O/C ratio distributions for L/S/IVOC emissions across volatility bins followed 2D-VBS default profiles. However, these default assumptions led to underestimated POA O/C ratios (see Sect. 3.3). Therefore, we conducted the 2D-VBS_A simulation, wherein sector-specific O/C ratios for POA emissions from gasoline/diesel vehicles, non-road mobile sources, power plants and industrial sources were constrained using observational averages in eastern China (Table S6).

The volatility (expressed in terms of C∗) and O/C ratios for OA surrogate species across all the simulations were summarized in Tables S7–S8. Volatilities for all species, except those in VBS-based parameterizations, were aligned with surrogate data in chamber experiments. Non-volatile POA and SOA formed via pathways such as aromatic oxidation under low-NO_x conditions, aqueous-phase uptake of IEPOX/MAE/glyoxal/methylglyoxal, ON hydrolysis and particle-phase oligomerization were assigned a fixed C∗ of 10⁻¹⁰ µg m⁻³ for simplicity. The volatility of primary L/S/IVOCs in the condensed phase and aged OA, was described in detail by Murphy et al. (2017). In the 1D-VBS_EY simulation, a new SOA species with C∗=0.1µg m⁻³ (unrepresented in the base 1D-VBS in CMAQ) was explicitly incorporated. The volatility distribution of total SOA (POA) was then derived using simulated concentrations of individual SOA (POA) species.

The overall O/C ratios for SOA, POA, and OA were calculated using the mole-weighted averages of the O/C values for individual species. The O/C ratios of OA species were calculated based on mass-based OM/OC ratios provided with the CMAQ code. The OM/OC ratios for SOA were primarily derived from chamber experiments (Pye et al., 2017; Carlton et al., 2010; Xu et al., 2018). For example, SOA species with relatively well-known structures (e.g., IEPOX-/MAE-derived SOA, isoprene dinitrates and monoterpene nitrates), along with seven VBS bins representing monoterpene SOA formed via O₃ and OH oxidation, adopted OM/OC values and other molecular properties from surrogate compounds identified in chamber studies. For other SOA species, once their C∗ values were determined (as described above) and the number of carbons (nC) was assumed, the OM/OC ratios were either inferred based on plausible structures (Pankow et al., 2015) or estimated using the relationship between volatility, nC and the number of oxygens (nO) used in 2D-VBS ((Pye et al., 2017). Additionally, OM/OC ratios for primary L/S/IVOCs and their aging products were constrained by laboratory work and filed observations as documented in previous studies (Murphy et al., 2017; Lu et al., 2020). The non-volatile POA emissions were assumed to have an OM/OC ratio of 1.6 (Turpin and Lim, 2001), which was equivalent to O/C of 0.35.

In the base 2D-VBS, the O/C distributions for L/S/IVOC emissions in each volatility bin followed the default settings in the community-contributed 2D-VBS. These O/C ratios were primarily based on Chang et al. (2022), which used emission test data for specific sectors when available and otherwise adopted values from prior studies. However, in the 2D VBS, the O/C ratios for emissions were capped at 0.4. For SOA products overlapping with AERO7i, such as isoprene SOA, the O/C ratios were set to match those in AERO7i. For other SOA surrogate species, the O/C ratios were determined either by adjusting the O/C distribution of first-generation products to align with experimental data (OH/O₃-intitiated monoterpene and sesquiterpene SOA), or based on explicit chemical mechanisms for initial oxidation (anthropogenic SOA), followed by aging through functionalization and fragmentation within 2D-VBS (Zhao et al., 2015). As a result, the O/C ratios for OH/O₃-intitiated monoterpene and sesquiterpene SOA in 2D-VBS generally ranged from 0 to 1.0 in 0.1 increments. In contrast, anthropogenic SOA (ASOA) exhibited a broader range, with a maximum O/C ratio of 2.0, to account for the high degree of oxygenation observed in toluene SOA in chamber experiments.

Previous measurements have demonstrated a close relationship between volatility and viscosity. For instance, Champion et al. (2019) found that SOA with higher fractions of EL/LVOCs showed increased viscosity. Similarly, an inverse correlation between Tg and vapor pressure was observed for isoprene SOA components (Zhang et al., 2019). Although the dependence of Tg on atomic O/C ratios is generally weaker than on vapor pressure (Koop et al., 2011), strong correlations between Tg and O/C ratios have been observed for oxidation products formed from specific precursors, such as α-pinene (Dette et al., 2014), n-heptadecane, and naphthalene (Saukko et al., 2012). Therefore, the uncertainties in volatility and O/C ratios may impact Tg and viscosity predictions when using the parameterizations that relate Tg and viscosity to volatility and O/C ratios (Li et al., 2020; Zhang et al., 2019).

Here we calculated the Tg of individual OA surrogate species i (Tg,i) using Eqs. (1), (2) (Li et al., 2020) or (3) (Zhang et al., 2019): 1Tg,i=288.70-15.33×log⁡10Ci0-0.33×log⁡10Ci022Tg,i=289.10-16.5×log⁡10Ci0-0.29×log⁡10Ci02+3.23×log⁡10Ci0×O/C3Tg,i=480.07-54395log⁡10RTMiCi0-7.79292+116.49 Where Eq. (1) relies solely on volatility (Ci0 represents the saturation concentration at 298 K, and is equal to Ci∗ assuming ideal thermodynamic mixing), and was developed for coupling into the 1-D VBS framework. Equation (2) incorporates O/C as an additional factor, which is used in 2D-VBS. Both equations yield similar predictions, particularly for compounds with low O/C ratios (Li et al., 2020). Equation (3) is a semi-empirical formula derived from the measurements of isoprene SOA components, which relates Tg to volatility and includes molar mass Mi (Zhang et al., 2019). The overall Tg of OA mixtures under dry conditions (Tg,org) can be calculated using the Gordon-Taylor equation (see Sect. S2 in the Supplement).

OA viscosity depends on aerosol water content, since water significantly influences the phase state of aerosols (Koop et al., 2011). Under humid conditions, the Tg of organic-water mixtures (Tg,ωorg) is calculated based on the mass fraction of organics (ωorg) in the mixtures, along with the Tg values of pure water (136 K) and dry OA (using Eqs. 1, 2 or 3). Notably, ωorg varies with ambient relative humidity (RH) and the effective hygroscopicity parameter of OA (κorg). Further details can be found in previous studies (DeRieux et al., 2018; Shiraiwa et al., 2017) and Sect. S2 in the Supplement. The temperature-dependent viscosity (η) can be estimated using the modified Vogel-Tammann-Fulcher (VTF) equation, as given in Eq. (4) (Angell, 1991), when the ambient temperature T is at or above Tg,ωorg; otherwise η is fixed at 10¹² Pa s. 4η=η∞eT0DT-T0 where η∞ and D denote the viscosity at infinite temperature and the fragility parameter, with the values of 10⁻⁵ Pa s and 10, respectively (DeRieux et al., 2018). T0 represents the Vogel temperature, and can be estimated using Eq. (5) (Angell, 1991; DeRieux et al., 2018): 5T0=39.17Tg,ωorgD+39.17

Two sets of observational data from field campaigns were used to evaluate the simulations. One campaign was conducted in the spring (17 March–21 April) of 2018 in DY, at a site located in the Yellow River Delta National Nature Reserve and strongly influenced by urban outflows from the North China Plain (NCP) (Feng et al., 2023). The observational data include the mass concentration and elemental ratios (e.g., O/C, H/C) of OA, measured using a high-resolution time-of-flight aerosol mass spectrometer (HR-AMS). The volatility distribution of OA (expressed as VBS) was estimated using an empirical method based on data from a thermodenuder (TD) combined with AMS. The sources of OA components were characterized using the positive matrix factorization (PMF) method, which resolved OA factors, including hydrocarbon-like OA (HOA), biomass-burning OA (BBOA), transported OOA and background OOA. The other campaign, led by the same team, was conducted in the autumn (29 September–21 November) of 2018 in urban GZ surrounded by transportation and residential districts, with significant impacts from local emissions (Chen et al., 2021). The resolved OA factors in GZ identified three POA groups (cooking OA (COA), HOA, and nitrogen-containing OA (NOA)) along with two types of SOA (low-volatility oxygenated OA, LV-OOA, and semi-volatile oxygenated OA, SV-OOA).

Figure 1 compared the observed and simulated diurnal variations of OA and its components, POA and SOA, in DY and GZ, respectively. The simulation with updated O/C ratios of POA emissions in the 2D-VBS framework (i.e., 2D-VBS_A) did not alter the emission magnitudes or volatility, and therefore had a minimal impact on the predicted POA mass concentration or volatility distribution. As a result, this case is not discussed in Sect. 3.1 and 3.2. The observations indicated that SOA dominated OA at both sites, contributing 72 % and 64 % at DY and GZ, respectively. The model generally underestimated the SOA contribution in DY (except for the 2D-VBS case), while slightly overestimating it in GZ (Table S9). In general, the 2D-VBS case predicted higher contributions of SOA to total OA (> 80 %) than in the 1D-VBS simulations.

The base 1D-VBS significantly underestimated SOA production, with NMBs of -72 % in DY and -71 % in GZ (Table S10). Adding L/S/IVOC emissions and their contributions to SOA in the 1D-VBS_E case resulted in approximately a two-fold enhancement in SOA mass concentration. This finding was consistent with earlier studies showing that 64 % to 100 % of observed SOA in GZ can be explained when both conventional precursors and S/IVOCs are considered (Hu et al., 2022), and that simulated S/IVOC contributions to SOA exceeded 50 % across most of China (Li et al., 2021a; Miao et al., 2021). The 1D-VBS_EY case substantially increased SOA production from aromatics and PAHs with updated yields (Fig. S6), and the diurnal variation was similar to that in the 2D-VBS simulation in DY. In GZ, the notable daytime differences in SOA between the 2D-VBS and 1D-VBS_EY cases were driven by variations in ASOA representations (Fig. S7). Mass increases in ASOA as a result of gas-phase aging were likely overestimated with the 4 km grid spacing in the 1D-VBS_EY simulation (Bilsback et al., 2023). While both simulations significantly improved SOA predictions, they still underestimated SOA concentrations compared to observations by 11 %–14 % (2D-VBS) and 30 %–33 % (1D-VBS_EY), respectively. Diagnosing the causes of these underestimations remains challenging due to the lack of chemically resolved SOA measurements. Although POA was underestimated (as discussed below), the sensitivity simulations indicated that its impact on SOA was limited (Fig. S8). At the GZ site, for example, increasing POA emissions by 70 % in the 1D-VBS_EY simulation to match the observed POA concentrations led to only 4 % increase in SOA, suggesting that missing or incomplete SOA formation pathways are likely the primary drivers of the SOA low biases. We infer that the nighttime low biases in both cases were likely attributed to insufficient formation of organic nitrates in the presence of NO₃ radicals. This might result from missing anthropogenic terpene and phenolic emissions from sources such as biomass burning and volatile chemical products (Wang et al., 2022; Coggon et al., 2021; Xie et al., 2025; Liu et al., 2024). The underestimated nocturnal SOA formation could also be attributed to the underprediction of aqueous-phase formation pathways, which are enhanced under high RH conditions at night (Wang et al., 2019; Gu et al., 2023).

Simulated POA also exhibited underestimation at both locations. In the comparison, BBOA and COA were excluded from the observations, due to the complexity and considerable uncertainties associated with cooking and biomass burning (particularly open burning) emission estimation (Li et al., 2023; Zhou et al., 2017). Therefore, the POA underestimation was most likely attributable to uncertainties in mobile emissions. Observed POA concentrations at both sites showed peaks during traffic rush hours, which were not well captured by the simulations. At the DY site, the resolved HOA in observations was partially aged, as noted in Feng et al. (2023) and included some SOA transported over long distances, which may also explain the POA underestimation. At the GZ site, which is more influenced by local emissions, the simulated peak concentration exhibited an earlier shift compared to the observed peak, suggesting that the diurnal patterns of POA emissions may differ from those specified in the model. Additionally, the POA volatility distributions (see Methods) and aging schemes in the 1D-VBS and 2D-VBS simulations were different: (1) The 1D-VBS involved heterogeneous aging of non-volatile POA. (2) In 2D-VBS, POA aging led to its complete transformation into SOA, whereas in the 1D-VBS, some oxidation products remained as POA. These differences led to distinct POA levels between the two cases.

Overall, total mass concentrations of OA were underestimated. The 1D-VBS_EY case, which had the smallest negative biases among the 1D-VBS simulations, underestimated OA by approximately 27 % in DY and 34 % in GZ. These discrepancies were dominated by SOA underestimation at both sites (Fig. S9). The 2D-VBS simulation demonstrated better performance in OA predictions than 1D-VBS_EY, with NMBs of -24 % in both DY and GZ, primarily due to higher predicted SOA mass concentrations.

The mass distributions of OA, POA, and SOA across different volatility bins (i.e., volatility distribution) were examined. This study primarily focused on compounds with C∗ equal or below 10⁻¹ µg m⁻³ (logC∗≤-1, referred to as low-volatility OA hereafter), and on SVOCs (i.e., -1<log⁡C∗≤2), which exist in both gas and condensed phases. Despite variations in OA mass concentrations across different cases, CMAQ generally reproduced the observed volatility distributions of OA, SOA, and POA in all the simulations (Fig. 2). Specifically, the observations indicated that more than 70 % of OA was low-volatility OA, consistent across the simulations. The base 1D-VBS significantly underestimated low-volatility OA concentrations, leading to underestimation in total OA mass (Fig. S10). The 1D-VBS_E simulation, which included L/S/IVOC emissions, increased low-volatility SOA formation by 1.1 µg m⁻³ in DY and 1.5 µg m⁻³ in GZ, although the contributions of low-volatility SOA to total SOA decreased. It also replaced the default volatility distribution of POA in CMAQv5.3 with source-specific gas/particle partitioning, resulting in a significant amount of non-volatile POA (Fig. S11). Collectively, these adjustments led to slightly higher fractions of low-volatility OA with the inclusion of L/S/IVOC emissions, as proposed by Xu et al. (2019), albeit for a different reason (mainly due to changes in POA rather than SOA volatility distributions). The updated SOA yields for aromatics and PAHs had minor effects on POA but increased SOA mass concentrations across three of the four volatility bins in Fig. S10, slightly changing the SOA volatility distributions.

The simulated OA volatility distributions, particularly for POA, differed between the 2D-VBS and 1D-VBS_EY simulations. The observations presented a higher fraction of low-volatility POA in GZ than in DY. The spatial variation in POA volatility was captured by the 1D-VBS_EY case. However, both 1D-VBS_EY and 2D-VBS overestimated the contributions of low-volatility POA. In 1D-VBS_EY, non-volatile POA was substantial in mass, and increased with aging (Fig. 3); semi-volatile POA underwent aging, with a portion remaining as POA (rather than fully converting into SOA as in 2D-VBS) but becoming more volatile. This ultimately led to a higher low-volatility POA mass and a lower contribution compared to 2D-VBS. The SOA volatility distributions in the 1D-VBS_EY and 2D-VBS cases both agreed well with the observations, which were similar at the two locations. However, the 2D-VBS simulation predicted a lower fraction of low-volatility BSOA in GZ compared to the 1D-VBS simulations (over 73 % vs. less than 86 %), primarily due to differences in the treatment of monoterpene-derived SOA through O3/OH oxidation pathways. Additionally, the contribution of low-volatility ASOA was also lower in the 2D-VBS simulation (over 62 % vs. over 69 %; see Fig. S12).

It is important to note that LVOCs (-4<log⁡C∗ (µg m⁻³) ≤-1) and ELVOCs (extremely low-volatility VOCs with logC∗ (µg m⁻³) ≤-4) were not explicitly resolved in this study (Fig. 3). In all simulations, SOA formation was dominated by species within C∗ bins ranging from 10⁻² to 10⁴ µg m⁻³. In addition, certain assumed “non-volatile” products (IEPOX/MAE-derived SOA, oligomers, etc.) were arbitrarily assigned to the volatility bin of C∗=10-10 µg m⁻³ (see Methods). The abundance of mass in the C∗=10-2 µg m⁻³ bin was particularly high for the 2D-VBS and largely due to aging. The absence of predicted SOA between the C∗ bins at 10⁻¹⁰ µg m⁻³ and C∗=10-2 µg m⁻³ was in contrary to the observed more uniform distribution across C∗ bins spanning from 10⁻⁹ to 10⁻¹ µg m⁻³. The 1D-VBS_EY case included formation of mass for the volatility bin at C∗=10-6 µg m⁻³ for HOMs formed from aromatics and PAHs under low-NO_x conditions, but the predicted mass concentration was negligible and biased low relative to the observations. Major ELVOC formation pathways, including autoxidation and bimolecular peroxy radical reactions for monoterpenes, have been incorporated into the Community Regional Atmospheric Chemistry Multiphase Mechanism (CRACMM) (Pye et al., 2023). However, the volatility distribution of the resulting products remained insufficiently resolved. Furthermore, anthropogenic ELVOC formation is not yet well understood (Shrivastava et al., 2024; Yin et al., 2024). As ambient observations consistently reported significant amounts of LVOCs and ELVOCs (Chen et al., 2024; Huang et al., 2024), future work should refine and expand the representation of autoxidation and other chemical processes contributing to LVOC and ELVOC formation from both anthropogenic and biogenic precursors. These improvements will enhance model accuracy in predicting SOA volatility and new particle formation.

Both observed and simulated OA showed higher O/C ratios during the day and lower ratios at night (Fig. S13), aligning with the temporal patterns of SOA mass concentrations. The higher O/C ratios in the afternoon were linked to elevated oxidant levels, which facilitated SOA formation (typically characterized by higher O/C ratios compared to POA, see Table 2) and the photochemical aging of OA. In contrast, during the night and early morning, the increased contributions of fresh POA led to lower O/C ratios.

Most simulations underestimated the O/C ratios of OA, with a more pronounced underestimation at the DY site, where both POA and SOA O/C ratios were lower than observed (Table 2). In DY, OA underwent prolonged aging during transport, resulting in a higher O/C ratio (0.84) than the national average (0.3–0.65) (Feng et al., 2023). Although the 1D-VBS_EY case predicted the highest O/C ratio of OA (due to potentially overestimated multi-generational oxidation as discussed in Sect. 3.1), it was still lower than the observed value (0.61 vs. 0.83). This discrepancy suggests that accurately simulating the evolution of OA O/C ratios with aging remains a significant challenge for current CTMs (despite satisfactory mass simulations). The observed POA contained aged HOA, which, although essentially SOA, could not be separated from HOA in DY (Feng et al., 2023), partly explaining the underestimation of POA O/C ratios. In the 1D-VBS_E case, the inclusion of L/S/IVOC emissions increased the O/C ratio of POA but decreased that of SOA, consistent with the impacts on volatility, i.e., increased (decreased) the contribution of low-volatility POA (SOA) (see Figs. 2 and S11). The 1D-VBS_EY case promoted SOA formation at C∗=0.1 µg m⁻³ through the newly added volatility bin for ASOA, which was highly oxygenated (O/C = 2.53, see Table S7), thereby increasing the O/C ratios of SOA. However, the impacts of HOMs were negligible due to their minimal contribution to mass.

At the GZ site, the overestimation of POA O/C ratios in certain cases (i.e., 1D-VBS_E and 1D-VBS_EY) partially offset the underestimation of SOA O/C ratios. As a result, the O/C ratio of total OA in 1D-VBS_E was closest to the observed value (both were 0.59), while the 1D-VBS_EY case overpredicted the O/C ratio (0.66 vs. 0.59). The more oxygenated POA in both simulations than observed could be attributed to overestimations in non-volatile POA mass concentrations. Additionally, COA, with an observed O/C ratio of 0.19 in GZ, were likely underestimated in the simulations given uncertainties in cooking emissions. If COA were well represented, it could slightly lower the simulated POA O/C ratios and reduce the positive biases in the 1D-VBS_E and 1D-VBS_EY simulations in GZ. It was also found that wind speeds were overestimated at GZ (Fig. S14), which could lead to excessive dilution and consequently reduce the simulated concentrations of locally emitted POA and freshly formed SOA (both of which typically exhibit lower O/C ratios than aged POA or SOA). This effect would be expected to increase the simulated O/C. However, the low biases in O/C at GZ indicate that meteorological biases are unlikely to be the dominant factor, and that limitations in the SOA representations may play a more important role.

The O/C ratios predicted by the 2D-VBS case were lower than those in other simulations, primarily due to the underestimation in SOA O/C ratios. In particular, the distribution of ASOA O/C ratios differed significantly from those in the 1D-VBS simulations (see Fig. S15), highlighting the constraint of ASOA O/C ratios in future work. Simulated POA O/C ratios were mainly influenced by how emissions were specified in the model. As the default settings in 2D-VBS underestimated POA O/C ratios, an additional simulation, 2D-VBS_A, was conducted with updated O/C ratios for emissions, constrained by those of POA factors in observations from prior studies (Table S6 and Fig. S16). For instance, in the default configuration, mobile sources were assumed to emit primarily hydrocarbons and low-oxygenated compounds, with over 90 % of emissions distributed in the O/C bins of 0 and 0.1. The optimized parameters in the 2D-VBS_A case reflected a higher degree of oxygenation, with more than 60 % of emissions allocated to the O/C bins of 0.2 and 0.4. As a result, the POA O/C increased from 0.26 (DY) and 0.20 (GZ) in 2D-VBS to 0.29 and 0.25 in 2D-VBS_A, respectively, and showed good agreement with the observation in GZ. However, the persistent low bias in POA O/C ratios in DY was likely attributable to uncertainties in emissions from sources with higher O/C, such as biomass burning (with an observed O/C ratio of 0.37 for BBOA) in the simulations, as well as potential overestimation of observed POA O/C ratios due to the influence of aged HOA.

Apart from underestimations in OA O/C ratios in most cases, the simulations struggled to capture the spatiotemporal variability on an hourly basis (Fig. 4). The observations revealed significant differences in the O/C ratios between DY and GZ, reflecting distinct aging processes associated with site characteristics (i.e., regional transport at DY versus dominant local emissions at GZ), rather than differences in oxidant levels, as comparable O₃ concentrations were observed at both sites. As discussed above, current SOA parameterizations inadequately represent chemical aging processes and thereby fail to reproduce the observed spatial contrasts in oxidation state. The observations also exhibited considerable temporal variability, particularly at the GZ site with abundant fresh emissions. Specifically, the interquartile range (IQR) of hourly O/C ratios was 0.77–0.89 in DY and 0.52–0.68 in GZ in the observations. However, the simulated O/C ratios of OA showed less variability than observed in GZ, with narrower IQRs and Standard Deviation Ratios (SDR) lower than 1.0.

The responses of daily OA O/C ratios to ambient O₃ concentrations were also examined (Fig. 5). All the simulations demonstrated an increased in O/C ratios as O₃ levels rose (indicating a higher atmospheric oxidative capacity), with the rates in the range of 0.0023–0.0045 per ppb O₃. The 2D-VBS simulations exhibited a steeper slope, suggesting stronger sensitivity to O₃ levels. This implied that oxidant abundance in the gas phase was an important driver of the increase in O/C ratios in the simulations. However, the observations did not show a clear correlation between O/C ratios and O₃. In DY, this could be due to the decoupling of O/C ratios from local O₃ in aged air masses. Other factors, such as fresh emissions in GZ (see Fig. S17) or aqueous formation pathways, may also contribute to the observed variations in O/C ratios.

Several widely used CTMs, including WRF-Chem and GEOS-Chem, have been utilized to predict OA viscosity and phase state based on volatility and O/C (Zhang et al., 2024; Luu et al., 2025). In this study, three parameterizations of Tg were employed to investigate how model capability in representing volatility and O/C influences the predictions of Tg (and the resulting viscosity using the VTF equation, see Methods): (1) one incorporating both O/C and volatility (Eq. 2), (2) one accounting for both volatility and molecular mass (Eq. 3), and (3) one based solely on volatility (Eq. 1). Given that Eq. (3) was developed from laboratory experiments focused on isoprene SOA, its applicability to POA and ASOA warrants further evaluation (which is beyond the scope of this study). Nevertheless, it was included here as a comparative parameterization to evaluate impacts of volatility representations. The simulation at the DY site was selected for detailed analysis as a representative case.

In DY, the mean Tg,org values for POA and SOA showed differences across parameterizations, ranging from 15.5 to 24.4 K for POA and 19.3 to 23.0 K for SOA (Fig. 6). The parameterization based on Eq. (1) generally predicted the highest Tg,org values, closely aligning with those from Eq. (2) as has been reported in Li et al. (2020). This suggests that simulated volatility distributions exerted a stronger influence on Tg,org predictions than O/C ratios, as further evidenced by the Tg,org values of individual SOA surrogates (Fig. S18). Equation (3) predicted the lowest Tg,org values, including for nearly all SOA components except a few isoprene SOA surrogates. For isoprene SOA components, Tg estimates remained constant using either Eqs. (1) or (3), due to the use of a fixed C* of 10⁻¹⁰ µg m⁻³ in CMAQ (see Methods), which resulted in unrealistically high Tg values, e.g., ∼ 400 K with Eq. (1). However, 2-methyltetrol (AIETET), despite its unexpectedly high ambient concentrations, might be semi-volatile, with predicted C* of 10² µg m⁻³, while 2-methyltetrol sulfate (AIEOS) has a relatively low volatility, with C* of 10⁻¹ µg m⁻³ (Budisulistiorini et al., 2017). These values resulted in much lowerTg estimates of approximately 230 and 276 K, respectively (Zhang et al., 2019). Therefore, the C* values for these compounds should be revisited when applied in Tg predictions.

The discrepancies in POA Tg,org derived from observed versus simulated volatility and O/C ratios using Eq. (2) ranged from 12.4 to 33.7 K across the five simulations. Compared to observation-based Tg,org, the 2D-VBS, 2D-VBS_A, and 1D-VBS simulations underestimated POA Tg,org, whereas the 1D-VBS_E and 1D-VBS_EY simulations overpredicted it, primarily due to a substantial contribution from non-volatile POA (Fig. S11). Although the 2D-VBS simulation exhibited the highest fraction of low-volatility POA (Fig. 2), these species were mainly allocated to the volatility bin of C∗=10-2 µg m⁻³, which led to lower Tg,org values than those inferred from observations. The biases in SOA Tg,org were smaller, ranging from 2.5 to 17.2 K. Most simulations exhibited a slight underestimations of SOA Tg,org relative to observation-derived values. Conversely, the 1D-VBS case overestimated SOA Tg,org, with an excessive proportion of SOA at C∗=10-10 µg m⁻³ (Fig. S11). Since Eqs. (1) and (2) were developed for use within the 1D-VBS and 2D-VBS frameworks, respectively, the evaluation presented here suggests that the differences in Tg,org estimates were driven by the volatility representations in the 1D-VBS versus 2D-VBS, rather than by the specific Tg parameterizations. Notably, the 2D-VBS implementation in CMAQv5.4 generally produced lower SOA Tg,org values.

The values of OA viscosity η, which determine their phase state, were calculated using the VTF equation, with Tg parameterized as a function of volatility and O/C ratios (Eq. 2), and incorporating additional variables including κorg, RH, and T (see Methods). Figure 7a and e compared viscosity estimated from simulations and observations, with all input variables (i.e., Tg,org, κorg, RH and T) obtained consistently from either model output or observations. In the simulations, κorg for each OA species was parameterized as a function of OM/OC (Eq. S7), assuming a constant density of 1.4 g cm⁻³. Bulk κorg for SOA and POA was calculated as a mass-weighted average across species. For observations, κorg was inferred from f44 (i.e., the fraction of m/z 44 signal in total organic signals) following Feng et al. (2023). The simulations generally reproduced the diurnal variability in OA viscosity, predicting higher η values during the daytime than at night, in line with the observed diurnal patterns. Field measurements indicated that POA was predominantly semi-solid (102≤η≤1012 Pa s) with a transition to the solid phase (η>1012 Pa s) between 12:00 and 17:00 LT. However, the model overestimated viscosity. Consequently, the 2D-VBS and 2D-VBS_A simulations predicted prolonged solid phase for POA, while 1D-VBS_E and 1D-VBS_EY predicted POA to remain solid throughout the day. For SOA, observations showed that it was semi-solid, while the simulations reproduced this phase state during nighttime hours only.

To identify the dominant source of overestimation in predicted viscosity, three sensitivity experiments were conducted (Table 3), each isolating the influence of a specific input variable. The results from sensitivity experiment A (Fig. 7b and f) revealed that the evident model-observation discrepancies in POA and SOA viscosity were positively correlated with the biases in Tg,org. For instance, in the 1D-VBS case, POA viscosity was significantly underestimated (by up to five orders of magnitude) due to a ∼ 40 K negative bias in Tg,org (Fig. S19a), while an overestimation of Tg,org for SOA (∼ 20 K, Fig. S19d) led to viscosity exceeding observation-based values by up to two orders of magnitude. In contrast, simulations such as 1D-VBS_E and 1D-VBS_EY, which better captured SOA Tg,org, yielded viscosity values in good agreement with observations. Thus, by comparing Fig. 7e and f, we conclude that the considerable overestimation of SOA viscosity in Fig. 7e, except in 1D-VBS, cannot be attributed to Tg,org bias. O'Brien et al. (2021) also reported a moderate sensitivity of predicted OA viscosity to Tg,org, e.g., under conditions of RH = 60 %, a reduction in Tg,org by ∼ 200 K resulted in a decrease in predicted viscosity from 10¹² to 10 Pa s (with κorg of 0.14).

On the other hand, the OA hygroscopicity parameter (κorg) was an important parameter in viscosity estimation. Sensitivity experiment B predicted viscosity comparable to those in Fig. 7a and e, indicating that the discrepancies between simulation- and observation-based viscosity values was due to the variation in κorg. Specifically, the κorg values used in the simulations underestimated the hygroscopicity of both POA and SOA relative to the values derived from m/z 44 (Fig. S20), e.g. 0.16 versus 0.048–0.069 (1D-VBS)/0.064–0.068 (2D-VBS) for POA, and 0.35 verus 0.013–0.014(1D-VBS)/0.12 (2D-VBS) for SOA. Earlier sensitivity simulations showed that a perturbation of ±0.05 in κorg resulted in changes of Tg,ωorg by ∼ 5 %–15 % over high RH areas, supporting a critical role of κorg in viscosity estimation (Shiraiwa et al., 2017). Consistently, the underestimation in κorg led to an overestimation in simulated viscosity in this study (given the inverse relationship between viscosity and hygroscopicity), and even offset the underestimations of viscosity that were caused by low biases in Tg,org (see Sensitivity experiment A). In CMAQ, κorg was parameterized as a function of OM/OC. In light of the notable underestimation of O/C ratios at the DY site (Table 2), this could be a key driver of the bias in viscosity predictions. While the impact of relative humidity on viscosity has also been emphasized in other studies (Maclean et al., 2021; Rasool et al., 2021; Li et al., 2021b), its effect is minimal in this study due to the strong agreement between simulated and observed meteorological variables (Fig. S21). This was further supported by the similarity in viscosity estimates derived using identical κorg and Tg,org but different RH and T inputs (Fig. 7d and h).