Biomass burning (BB), including wild and prescribed fires, is estimated to be the largest primary source of fine particulate matter (PM) and the second-largest source of volatile organic compounds (VOCs) globally (Yokelson et al., 2008), impacting air quality, public health, and climate. In fire-prone areas such as the western United States of America (USA), the relative importance of BB emissions as a source of air pollution has been growing due to increased wildfire activity (Westerling, 2016; Higuera et al., 2021) and decreased anthropogenic emissions (Warneke et al., 2012; Simon et al., 2015). Wildfires have been suggested to account for up to half of the overall PM2.5 burden since 2012 and contribute to its increasing trend in the last 3 decades in the western USA (McClure and Jaffe, 2018; O'Dell et al., 2019; Burke et al., 2021). Wildfire impacts on VOC burdens are highly uncertain, in part due to the limited observational constraints on BB VOC emissions. Here we apply comprehensive VOC observations from two recent aircraft campaigns targeting fires, along with the GEOS-Chem chemical transport model (CTM), to examine our understanding of BB emissions in the western USA.

Current CTMs often poorly simulate the impact of wildfire smoke partly because of an incomplete description of the quantity and speciation of VOC emissions, along with poor representation of their spatial, temporal, and vertical distributions (Alvarado and Prinn, 2009; Jaffe and Wigder, 2012; Jaffe et al., 2018; Baker et al., 2016, 2018; Wolfe et al., 2022). BB emission estimates are typically derived from the product of a compound-and-biome-specific emission factor (EF, expressed as mass of species in grams per dry biomass burned in kilograms) and an effective amount of dry matter burned (effective DM burned, kg). Both EF and DM burned are subject to large uncertainties (Akagi et al., 2011; Andreae, 2019; Carter et al., 2020). EFs are either measured in laboratory burning experiments that attempt to simulate real-world fires or quantified from near-field measurements on the ground or air that may be influenced by atmospheric aging processes before sampling (e.g., Burling et al., 2010; Warneke et al., 2010; Wooster et al., 2011; Permar et al., 2021; Majluf et al., 2022). Recent efforts to reconcile the difference between laboratory and field measurements support the need to adjust lab EFs to the typical field combustion efficiency (Permar et al., 2021; Selimovic et al., 2018). However, the burn conditions throughout the course of a fire are currently not considered in inventories. In addition, commonly used global BB emission inventories often consider only three to six biome groups (Andreae and Merlet, 2001; Wiedinmyer et al., 2011; Akagi et al., 2011; Randerson et al., 2012; Kaiser et al., 2012; Koster et al., 2015; Andreae, 2019). For example, the Quick Fire Emissions Dataset version 2.4 (QFED2.4) inventory has three biome groups to represent all global biomass: tropical forest, extratropical forest, and savanna/grass (Koster et al., 2015). In the Global Fire Emissions Database version 4 with small fires (GFED4s), the extratropical forest biome is subdivided into the boreal forest and temperate forest, and additionally two biomes for peatland and agriculture/waste burning are considered, thus giving a total of six (van der Werf et al., 2017). Differences (and errors) in vegetation classifications among inventories can also lead to diverse assigned EFs, even though those EFs may come from the same experimental studies, thus resulting in different emission estimates.

DM burned in commonly used emission inventories is estimated by two different satellite remote sensing approaches. The “fire-detection-based and/or burned-area-based (FD-BA)” method estimates DM burned from the product of fire burn areas (BAs) and fuel consumption (i.e., loading, type, timing, and rate). Global BB emission inventories using this method include GFED4s (van der Werf et al., 2017) and the Fire INventory from NCAR version 1.5 (FINNv1.5; Wiedinmyer et al., 2011). Another approach uses fire radiative power (FRP, radiant energy released per time by burning fuel) and its empirical relationship with biomass burned. Some widely used BB emission inventories using this approach include QFED2.4 (Koster et al., 2015) and the Global Fire Assimilation System version 1.2 (GFASv1.2; Kaiser et al., 2012). Both FD-BA-based and FRP-based inventories share common sources of uncertainties, such as missing fire detections and/or FRP observations used to initialize DM burned estimates. Additionally, those fire products are mostly from polar-orbiting satellites with a low temporal coverage (i.e., once or twice daily at a fixed local time) and can be obscured by clouds and smoke, resulting in assumptions often having to be made to fill both temporal and spatial gaps in the observations (Wang et al., 2018; Wiggins et al., 2020; Stockwell et al., 2022). Current operational BB emission inventories can produce monthly CO and aerosol fluxes that vary by a factor of 5 or even 20 for a specific region (Al-Saadi et al., 2008; Zhang et al., 2014; Koster et al., 2015). These differences in global total emissions averaged over longer periods are smaller but still on the order of a factor of 2–4 (Stroppiana et al., 2010; Granier et al., 2011; Carter et al., 2020; Liu et al., 2020; Pan et al., 2020). The discrepancy could be even larger in VOC emission estimates due to different speciation among inventories (i.e., GFED4s has 21 VOCs, while QFED2.4 has 9 VOCs). Different input data used to drive BB emissions, such as EFs, fire detections, fire burned area, and the amount of biomass burned, are all thought to contribute to the divergent estimates among emission inventories. Recently, Carter et al. (2020) suggested that, at least for aerosol, the BB emission uncertainties are mostly from DM burned at both regional and global scales and that differences in EFs across inventories are smaller than differences in DM burned. These errors in estimating DM burned will also affect VOC emission estimates; thus their uncertainty is thought to be at least on a similar order to that of aerosol and CO.

When compared to observations, previous model evaluation studies (again mostly focusing on CO and aerosol) often point to a general underestimation of BB emissions in the commonly used inventories and a factor of 2 as the global BB emission uncertainty (Kopacz et al., 2010; Wang et al., 2018; Carter et al., 2020; Pan et al., 2020; Bela et al., 2022). For example, various degrees of negative model bias are found in aerosol optical depth and CO near BB source regions when compared to corresponding satellite and ground observations, though the FRP-based BB inventories often provide higher emissions than the FD-BA-based estimates (Yurganov et al., 2011; Petrenko et al., 2012, 2017; Zhang et al., 2014; Reddington et al., 2016; Pan et al., 2020; Liu et al., 2020; Bela et al., 2022). For the western USA, Pfister et al. (2011) suggested that an early version of FINN (version 1) underestimated BB emissions by a factor of 4 over California, as revealed by constraints from aircraft and satellite measurements. More recently, Carter et al. (2021) found that the GEOS-Chem model driven by GFED4s is biased low for CO but captures the carbonaceous BB aerosol when compared to the 2018 WE-CAN (Western Wildfire Experiment for Cloud Chemistry, Aerosol Absorption and Nitrogen) airborne observations. Another recent study by Bela et al. (2022) found that the daily mean emission estimates from seven existing inventories for a case study of a western US wildfire varied by a factor of 83, despite bracketing the observed BB CO fluxes. Even with observational constraints on certain input parameters (e.g., for relating FRP to the quantity of biomass burned or emissions released), their uncertainty range is still a factor of ∼ 2 compared to the direct CO flux measurements in fire plumes (Bela et al., 2022). A similar case study also suggested a wide spread of the hourly emission estimates (spanning a factor of > 33) from nine satellite-based inventories in the FIREX-AQ (Fire Influence on Regional to Global Environments and Air Quality) airborne observations (Stockwell et al., 2022).

Here we aim to improve current understanding of VOC emissions from wildfires in the western USA. Leveraging the comprehensive VOC observations from the WE-CAN airborne campaign, we evaluate a 0.25∘ × 0.3125∘ nested version of the GEOS-Chem CTM driven by three commonly used global BB emissions inventories (Sect. 4). We assess the potential reasons for model and observation discrepancies including the fire detections, emission ratios, and plume injection heights used in the emission inventory/CTM (Sects. 5 and 6). We further apply independent measurements from ground sites and the FIREX-AQ airborne campaign to test the regional representativeness and interannual variability in our findings (Sects. 7 and 8). Finally, we discuss the potential implications of our findings, taking into account the uncertainties associated with the model results (Sect. 9).

Figure 1 shows the VOC primary emissions over the western USA in the base simulation during the 2018 fire season (June–September, or JJAS). This 4-month period typically accounts for 70 %–90 % of the annual acreage burned in this region (Jaffe et al., 2008). In GFAS, the JJAS contributes 85 % of the 2018 annual BB emissions in the western USA. According to emission inventories chosen in GEOS-Chem, biogenic emissions are thought to be the dominant VOC source in summer in this region (2200 GgC or ∼ 75 % of the total VOC emissions), followed by anthropogenic emissions (405 GgC or ∼ 15 %) and BB emissions. In JJAS 2018, the total BB emissions from 14 VOCs in the model range from 220 to 340 GgC (or 10 % of the total). As we will show later, our model–observation comparisons suggest that the role of BB is significantly underestimated in the current CTMs.

In the 2019 fire season, the BB VOC emissions (40–75 GgC or 1 %–3 % of total primary VOC emissions) are 10 %–30 % of those in 2018 for this region and about 40 %–50 % of the 2019 annual BB emissions depending on which BB inventory is used. This shows the WE-CAN and FIREX-AQ aircraft campaigns sampled two distinct fire seasons which may reflect upper and lower bounds of wildfire activity in this region (https://www.nifc.gov/fire-information/statistics/wildfires, last access: 21 May 2023). Despite the large interannual variability in wildfire emissions, the western USA accounted for ∼ 90 % of BB VOC emissions in the contiguous United States (CONUS) in 2018 and ∼ 60 % in 2019 according to GFAS, confirming a significant fire influence exists in the western USA, which could also affect the rest of CONUS downwind (O'Dell et al., 2021).

The total BB VOC emission estimates in the western USA differ by 20 %–40 % across the three global inventories examined for the 2018 fire season. All emission inventories show similar spatial distributions as they all use MODIS satellite products such as active fire, burned areas, or FRP as inputs. However, larger differences between inventories occur for emission estimates for individual fires on specific days (more than a factor of 20), also shown in Bela et al. (2022) and Stockwell et al. (2022). These differences likely reflect the various assumptions or adjustments made for fire persistence, small fires, or fires obscured by clouds and haze in the inventories (Liu et al., 2020).

All three global BB inventories suggest aldehydes, alkanes, and alkenes are the most abundantly emitted VOCs from western US wildfires, largely consistent with recent field measurements (Permar et al., 2021). However, emission estimates for individual VOCs disagree by a factor of 1–5 in the western US fire season (Figs. 2 and S3). Emission estimates for xylenes show the largest difference (5 GgC in GFED4 versus 1 GgC in GFAS), while propane emissions agree within ± 20 % across the three inventories (8–10 GgC). Despite the same FRP products being used, QFED is lower than GFAS by a factor of 2–3 for emission estimates of acetaldehyde, lumped ≥ C4 alkanes, and lumped ≥ C3 alkenes. This discrepancy can likely be explained by different emission ratios and speciation used for lumped compounds in these two inventories, in which QFED tends to have simpler speciation (Kaiser et al., 2012; Koster et al., 2015). For instance, C3–C6 and C8 alkenes are incorporated in the GFAS, while only C3 alkene is considered in the QFED. Besides, the emission ratios in QFED are primarily sourced from Andreae and Merlet (2001), whereas GFAS has incorporated updates from literature through 2009.

A recent global study comparing BB aerosol emissions from inventories suggests that the effective DM burned is the biggest contribution to divergent emission estimates across inventories (Carter et al., 2020). In contrast, we find that the regionally averaged ERs dominate disagreement in emission estimates for most VOCs across the three inventories (Figs. 2 and S4). These ERs are regionally averaged from each inventory and thus are functions of both assigned ERs for specific biome and vegetation classifications and are calculated from the regression of daily mean VOC and CO BB emission fluxes at each grid cell for the region from inventories.

We also find that, at least for the western USA, these three inventories agree on the amount of effective DM burned to within 40 % (47–67 Tg in 2018, as calculated by dividing VOC and CO emission estimates by the corresponding regionally averaged EF). Even though the amount of effective DM does not drive the inventory–inventory discrepancy of emission estimates in the region, our model–observation evaluations in the following sections infer that it is likely too low. Indeed, using the National Interagency Fire Center burned area report (∼2.42×106 ha for 2018 in the west), back-of-the-envelope calculation suggests that these global inventories' effective DM burned per area is 19–28 Mgha-1. These values of biomass burned per area are at the low end of the range of estimates from the United States Forest Service (USFS) fuel consumption models, such as FOFEM and Consume (25–193 Mgha-1 for western US wildfires; Reinhardt et al., 1997; Drury et al., 2014). Besides, limited field fuel consumption measurements of western mixed conifer forest wildfires, ranging from 32–44 Mgha-1 (Campbell et al., 2007; Hyde et al., 2015), are also higher than values from the global BB emission inventories. Taken together, these calculations suggest that the three global BB inventories underestimate the amount of DM burned in the western USA for the fire seasons examined here, a conclusion that will also be inferred from our model–observation comparisons in the following sections.

Figures 3 and 4 show the vertical distribution of CO and VOCs sampled by the C-130 during WE-CAN, as well as comparisons to various simulations. The observed abundance of all species is elevated by 50 %–300 % within the planetary boundary layer (> 850 hPa), indicating influences from anthropogenic, biogenic, and/or BB emissions near the surface during takeoff and landing time. The higher abundance in the middle troposphere (750–500 hPa) than typical background conditions (i.e., < 500 hPa) is mostly due to BB, as the C-130 targeted sampling both wildfire smoke in the near field and aged smoke whenever feasible while in transit during WE-CAN.

Simulations driven by different BB emission inventories show remarkably similar abundance (mostly within ± 10 %, except for surface toluene and CO within ± 30 %–40 %). All the inventories capture the enhancement patterns observed by the C-130, both elevated altitudes and timing with high correlations with observations (r = 0.7 to 1.0 in 5 min averaged data). The sensitivity run with no BB emissions (noBB) indicates that wildfire is a significant source for CO and primary VOCs including propane, benzene, and toluene during WE-CAN (enhanced by 2–3 times compared to noBB) but a lesser source for oxygenated VOCs (OVOCs), especially for formaldehyde (Figs. 4 and S5). The model driven by GFAS (GEOS-Chem + GFAS) tends to simulate slightly higher and better VOCs than GFED4 and QFED, possibly reflecting that GFAS has more accurate ERs as discussed later in Sect. 6.

All three inventory experiments significantly underestimate observed CO and VOCs, except for MEK. In the middle to lower troposphere (> 500 hPa), simulations reproduce 40 %–70 % of the observed abundance of CO, benzene, toluene, and acetone and 30 %–40 % of the observed propane, formaldehyde, acetaldehyde, and lumped ≥ C3 aldehydes but only 0 %–10 % of the observed organic acids. The model suggests mixed performance for xylenes, i.e., a high bias of 0 %–100 % in the lower troposphere and a low bias of 50 %–100 % in middle troposphere. In a relatively clean environment (< 500 hPa), the simulations show relatively small negative biases for all compounds and tend to match observations in generally clean or well-mixed environments during WE-CAN.

Unlike other VOCs and CO, MEK is systematically overestimated by 50 %–300 % throughout the middle to lower troposphere in all simulations including noBB but is reproduced in a relatively clean environment (< 500 hPa). Similar positive model bias has been reported in a recent study comparing GEOS-Chem to a comprehensive suite of airborne datasets over North America (Chen et al., 2019). This is likely due to the overestimation of MEK or its precursors in the EPA NEI and/or MEGAN inventories (Yáñez-Serrano et al., 2016), as such large high model bias exists even when the BB influence is removed (Fig. 5). Thus, further evaluation is needed for the sources of MEK and its precursors in anthropogenic and biogenic emission inventories.

We further refine the analysis in low-/no-smoke conditions by filtering out data when either the observed acetonitrile (CH3CN) mixing ratio, a known BB tracer, is > 159 ppt (25th quantile of CH3CN) or the enhancement ratio of CH3CN relative to CO is > 2.01 ppbppm-1 (Huangfu et al., 2021). The vertical profiles after applying this filter are shown in Figs. 5 and S6 and represent about one-third of the sampling time during WE-CAN, allowing us to examine the non-BB-related processes/emissions. Compared to the full campaign data, the observations of CO and all VOCs in low-/no-smoke conditions are lower by a factor of 2 or more, confirming the important influence from BB in the western USA during WE-CAN. The simulations capture the observed CO, benzene, and toluene in this clean environment but still underestimate the rest of the VOCs (especially OVOCs) by 10 %–90 %, except for MEK. The low model bias for formaldehyde in the free troposphere can be partly due to underestimated oxidation of CH4 or other precursors (Zhao et al., 2022). The negative model bias for acetaldehyde, formic acid, and acetic acid in the PBL may be related to reasons including missing or underestimated precursors from biogenic emissions (Millet et al., 2010, 2015; Paulot et al., 2011). The negative bias for acetone in the middle–upper troposphere may reflect a poorly constrained global background from ocean sources in GEOS-Chem (Wang et al., 2020). Nevertheless, the negative model bias in the low-/no-smoke conditions sampled during WE-CAN (Fig. 5) is much smaller than the BB-influenced environment. Thus, underestimation in the low-/no-smoke conditions does not explain model underestimation across compounds in the full campaign dataset (Fig. 4).

We calculate the average model biases that are due to BB processes for each species using the enhancements between the full campaign dataset and the low-/no-smoke conditions. Given the calculation of primary trace gases (CO, propane, benzene, and toluene), we conclude that the model potentially underestimates BB emissions or related processes by a factor of 3–7 in the GFAS while the bias can slightly vary in the GFED4 or QFED. Thus, we further carry out a sensitivity run by tripling the GFAS emissions in the model (GEOS-Chem + 3 × GFAS) as a test of the BB impact in the western USA. Figures 3 and S5 show that tripling BB primary emissions results in evident improvements and reproduces the observed levels for CO and most primary VOCs (propane, benzene, and toluene). The improvement for xylenes is moderate due to other model errors in the averaged OH reaction rate constant and ER (Sect. 6).

GEOS-Chem + 3 × GFAS has elevated simulated abundance for OVOCs to various degrees compared to the base run. For acetaldehyde and acetone, we find that 3 × GFAS brings the model close to the measurement uncertainty. For formaldehyde, formic acid, acetic acid, and lumped ≥ C3 aldehydes, tripling the primary BB emissions of these species (and their precursors that are included in the model) does not significantly improve the model–observation discrepancy (the improvement is within 5 %). Since 3 × GFAS mostly corrects the model error in primary BB emissions, this underestimation suggests the current model likely misses large secondary sources of these compounds in BB plumes due to insufficient VOCs representation or errors in the chemical mechanism, which is supported by a recent box-modeling study (Wolfe et al., 2022).

Eulerian models are known to have trouble preserving sub-grid features such as concentrated fire plumes over time due to rapid dissipation by numerical diffusion (Eastham and Jacob, 2017; Rastigejev et al., 2010). Campaigns targeting plumes like WE-CAN can get particularly intense and thus deviate from the climatologically diurnal distribution of BB emissions used in the model, resulting in low model bias when compared to aircraft measurements. In addition, any wind direction or plume height errors in the model would result in the model's aircraft diagnostics missing the fire plume when the real aircraft sampled it, contributing to some amount of low bias. Finally, if the plumes are narrower than ∼ 25 km (and the aircraft transect lengths are also narrower than ∼ 25 km), then the plume will be diluted in the model grid box more than the plane observed (even when including the transect portions outside of the plumes), also contributing to a low model bias. In addition to the BB emissions, those above factors due to fire sub-grid features may all have contributed to the low model bias in the aircraft analysis, but it is difficult to fully tease them out if at all possible. We thus consider the model bias revealed here as the upper limit of BB emission negative bias in the global inventories (Sects. 3 and 8).

To explore causes for the underestimation of BB emissions for these three emission inventories, we first determine if the inventories have detected the 27 individual fire plumes sampled in WE-CAN. A fire is considered to be detected if the inventory has registered any CO emissions in the model surface grid box at its location when the C-130 arrives. Table S1 shows that all the BB inventories (including FINNv1.5) capture all the sampled fires. BB emission inventories typically rely on space-based observations of burned area or FRP (i.e., MODIS Terra and Aqua fire products) for fire detections. For example, MCD64A1 burned area products are applied in GFED4 and MOD14–MYD14 FRP products are used in GFAS and QFED. During WE-CAN, wildfires were mostly sampled in the late afternoon when fires were the most active. The fires sampled by the C-130 tended to have developed well-defined plumes that were visible from geostationary GOES-16 or GOES-17 GeoColor images in the morning of the same day when flight planning was finalized. Our finding suggests that the fire detection products from low-orbit satellites commonly used in global BB emission inventories are efficient at detecting large fires in the western USA that tend to burn for several days if not weeks or months.

We further examine the impact of the assumed injection altitude of BB emissions by conducting sensitivity tests using five different BB injection height schemes (Table 2; Sect. 2.3). Figure S7 shows almost identical model vertical profiles in the five plume injection experiments, particularly in the free troposphere. In the middle troposphere, the simulations with higher plume injection heights tend to show larger enhancements; in the PBL, releasing BB emissions at the surface tends to result in the highest surface mixing ratios among the experiments. But the differences across simulations are within ± 10 % except for benzene and toluene (about ± 40 % near the surface). The model does not appear to be highly sensitive to assumptions regarding BB injection heights in the western USA at ∼ 25 km resolution. This insensitivity is likely because the trace gas emissions from large wildfires are efficiently lifted into the free troposphere by strong vertical mixing in the summer (Chen et al., 2009; Jian and Fu, 2014). However, the choice of plume injection heights can still be important for secondary production and downwind areas (Tang et al., 2022). For example, daily mean ozone concentrations vary by up to 14 % or 4 ppb at the surface in our injection experiments. Thus, the impact of various BB emission injection schemes on surface air quality needs further investigation, especially for populated downwind regions.

Emission ratios (ERs; often interchangeable with emission factors or EFs) can be a source of uncertainty in BB emissions estimates if they are poorly characterized or unmeasured (e.g., Akagi et al., 2011; Urbanski et al., 2011). We calculate ERs from the slope of the reduced major axis regression of VOCs and CO measured (and simulated) in emission samples. In order to calculate ERs, plume samples with physical ages less than 2 h in the WE-CAN campaign and less than 1 h in the FIREX-AQ campaign are used; these are deemed to be relatively fresh, with minimal or no secondary production. We note the observed ERs derived here using the 5 min averaged data tend to agree with what Permar et al. (2021) reported within 20 %, despite Permar et al. (2021) calculating ERs from 1 s observations and using the integration approach. Also, calculating observed and simulated ERs in a consistent way and according to the temporal and spatial resolution of the model can provide a valuable constraint on the overall model processes in terms of BB emission locations, timing, transport, and chemistry in fire-influenced environments.

Figure 6 illustrates this approach with scatterplots of a subset of observed VOCs and CO in emission transects and their comparison to the simulated relationship in GEOS-Chem + GFAS. The model shows the strong correlations between VOCs and CO (r = 0.7 to 1.0), suggesting GFAS captures the regional BB locations and timing sampled by the C-130 (Sect. 5). We find GFAS ERs agree with observed ERs within 30 % or better for formaldehyde, acetaldehyde, benzene, toluene, and lumped ≥ C4 alkanes. GFAS is either too high or too low by 50 %–70 % for ethane, propane, and acetone. Overall, GEOS-Chem + GFAS tends to produce higher and more accurate ERs than the other two inventories (Figs. 7 and S8). Some notably large errors in simulated ERs (≥ a factor of 2) include acetaldehyde in QFED and acetone, MEK, benzene, and toluene in GFED4.

The modeled abundance and ERs of xylenes and lumped ≥ C3 alkenes are significantly underestimated across all inventory experiments. These two lumped VOC groups are highly reactive, with lifetimes of ∼ 1 h (assuming an average in-plume OH concentration of 1 × 107 molec.cm-3 and an OH reaction rate constant kOH of 23.1 × 10-12–25.0 × 10-12 cm3molec.-1s-1). Errors in their loss via OH reactions due to incorrect OH concentration or kOH could distort their simulated abundance and ERs. Model bias in OH concentration would affect all primary species in the same direction, and reactive VOCs would be particularly sensitive to such error. Thus, we use aromatic hydrocarbon–hydrocarbon relationships to diagnose if there are any major model OH biases in the current version. Figure S9 shows that the base model can capture the observed toluene–benzene relationship, in terms of both emission ratios and their relative decay rates. This agreement indicates the good reproduction of the OH level in the model, and future analysis is needed for evaluating current kOH in the model.

Further, we find that kOH for xylenes in recent GEOS-Chem versions has been updated based on new assumptions. The GEOS-Chem version 12.5.0 used in this analysis assigns 23.1 × 10-12 cm3molec.-1s-1 as kOH for xylenes, based on the assumption that m-xylene is the dominant isomer (Fischer et al., 2014). Other studies using the fractions of xylene isomers observed in urban atmospheres for a weighted kOH have suggested values of 13.2×10-12–17.0 × 10-12 cm3molec.-1s-1, about 25 %–40 % lower than used here, which, if updated, would result in higher simulated xylenes (Atkinson and Arey, 2003; Hu et al., 2015b; Bates et al., 2021). Therefore, correcting kOH could partly reconcile the negative model bias for xylene ERs (i.e., 0.15 in corrected simulations vs. 0.32 ppbppm-1 in observations). The isomer fractional information for other lumped species and their chemistry in various environments is less known; thus future investigation is needed to refine and assess the chemical impact of these lumped species.

The national wildland fire burned area in 2019 was only about half that in 2018 (https://www.nifc.gov/fire-information/statistics/wildfires, last access: 21 May 2023). This is also reflected in the different acetonitrile distributions measured between the two aircraft campaigns (median 295 ppt during WE-CAN versus 205 ppt during FIREX-AQ; Fig. S2). To examine the year-to-year variability and regional representativeness of findings inferred from the WE-CAN C-130 measurements, we expand the analysis to observations from nine ground-based sites in 2018 and the FIREX-AQ DC-8 aircraft in 2019. The ground stations span several urban areas that are regularly affected by wildfire smoke. More importantly, the longer-term stationary measurements are further downwind in a better-mixed environment and physically unable to target plumes, and they can thus provide a counter-test to the contribution of the other factors from fire sub-grid features to model bias relative to aircraft observations that target the plumes (Sect. 4).

Figure 8 shows that most of the nine ground sites were heavily impacted by wildfire smoke in the 2018 summer, as indicated by elevated CO mixing ratios up to 250 ppb or higher lasting for a few days at times, while the general urban background CO is about 150–200 ppb (Pfister et al., 2011; Kim et al., 2013; Lopez-Coto et al., 2020; Gonzalez et al., 2021). Using the noBB and the base simulations, we define “BB-impacted days” as days when the modeled CO daily mean is increased by more than 20 % relative to the noBB run, and the rest of the days are termed low-/no-smoke days. By this definition, Seattle and Denver were least affected by BB in 2018 among the nine sites but still experienced 7–8 BB-impacted days out of 55 d. The rest of the sites all experienced ≥ 25 BB-impacted days, according to GEOS-Chem + GFAS. In general, the base model captures the daily variation in the observed CO (r > 0.40 at all sites, with six sites having r> 0.65). In Seattle and Denver, anthropogenic emissions dominated local CO abundance and variability in 2018. The US EPA NEI appears to have spatial biases as the base simulation captures observed CO in Denver but overpredicts CO in Seattle.

Tables S6–S8 summarize the mean bias, root mean square error (RMSE), and observation–model correlations for the entire data period, BB-impacted days, and low-/no-smoke days. Results show that GEOS-Chem + GFAS underpredicts observed CO at the other seven sites by 95–140 ppb on average for the entire period. The negative model mean biases are larger on BB-impacted days, pointing to model errors in BB-related processes. The base model does overpredict a few BB-impacted events, i.e., 4 and 17 August in California (Chico, Stockton, or Fresno), likely because local meteorological processes affecting smoke transport or the timing of BB emissions of certain individual fires are not captured in the model (O'Neill and Raffuse, 2021). Even so, the simulated CO abundance is underpredicted by > 100 ppb on 40 %–60 % BB-impacted days for all seven sites, while the model background bias (loosely calculated by the 5th-percentile CO mixing ratio) tends to be less than 70 ppb. Thus, similar to the findings in Sect. 4, correcting the model background CO bias (due to anthropogenic emissions or global background) is not enough to reconcile the large model–observation discrepancy. We find that the 3 × GFAS simulation systematically improves the model mean bias to various degrees across the western USA for the seven fire-influenced sites without degraded correlation coefficients with observations.

Figures 9 and S10 show the model evaluation with FIREX-AQ DC-8 VOC observations for the western USA. Observed VOC mixing ratios during FIREX-AQ are lower than in WE-CAN for this region partly due to fewer BB emissions in 2019 (Sect. 3). Overall, our findings for 2019 FIREX-AQ are consistent with the 2018 WE-CAN evaluation: the base simulation tends to underestimate all observed VOCs but MEK by a factor of 2–12 in the middle to lower troposphere. When we restrict the analysis to the low-/no-smoke environment, the base model also underestimates OVOCs and these negative model biases tend to be 40 %–100 % for the entire campaign average (Fig. S11). The model improvement for primary VOCs from tripling BB emissions is significant across the troposphere but not as obvious as during WE-CAN due to smaller BB emissions in 2019 (Sect. 3). Both WE-CAN and FIREX-AQ observations imply that the model misses substantial sources for OVOCs, particularly formaldehyde, formic acid, acetic acid, and lumped ≥ C3 aldehydes.

We do not attempt to evaluate the modeled ERs for FIREX-AQ because the inventories do not update ERs for different years. Figure 7 shows the observed ERs in WE-CAN and FIREX-AQ are consistent within the combined instrument uncertainty (± 40 %) for a majority of VOCs in western fuel types. Given the observational constraints in ERs and primary BB emissions, we infer that the above missing OVOC sources in the model are most likely from photochemical reactions in smoke plumes.

We infer the systematic underestimation of simulated CO and individual VOCs in the western USA is mostly driven by the low bias of effective dry matter burned in fire-detected areas across three global BB emission inventories. This finding is also supported by the low bias of inventories' DM burned per area (Sect. 3), the analysis of fire detections, injection heights (Sect. 5), ERs from airborne measurements (Sect. 6), additional model evaluations with long-term stationary ground measurements (Sect. 7), and aircraft observations in a different year (Sect. 8). Nevertheless, the 3-times underestimation of effective dry matter burned can be recognized as the upper limit as the negative model bias could also be attributed to the Eulerian models not being able to resolve sub-grid features such as fire plumes (Sects. 2.3 and 4). It is impossible to rule out and quantify these sub-grid uncertainties in the 0.25∘ × 0.3125∘ GEOS-Chem nested simulation (Rastigejev et al., 2010; Eastham and Jacob, 2017), though our evaluation using ground measurements helps support the argument of the dry matter burned underestimation. Novel methods such as adaptive grids or embedded Lagrangian plumes are needed to fully resolve local conditions of the plume in future studies.

Sensitivity tests with tripled BB emissions result in better agreement between observations and model outputs, particularly for primary VOCs. Thus, our best estimate of the BB primary emissions of the 14 modeled VOCs for the western US 2018–2019 fire seasons is 120–1020 GgC, which is 3 times the default emission estimates in three BB inventories. This is also ∼ 5 %–30 % of the total VOC emissions from primary sources for the western US fire seasons. However, the model still underpredicts OVOCs, even with tripled BB primary emissions; we are thus unable to constrain secondary production of BB VOCs in this work.

The above BB emission estimates are derived from 14 modeled VOCs with BB representation in three BB inventories (Table 1). However, the total ER of these 14 BB VOCs only accounts for half of the total measured VOC ERs from 161 species observed during WE-CAN (75 versus 150 ppbppm-1; Permar et al., 2021). The uncharacterized BB VOCs in the model mean that there is a significant quantity of missing reactive organic carbon fluxes in many major BB emission inventories and CTMs. Their chemical and health impacts on the regional and global scale remain largely unexplored (Carter et al., 2022; Permar et al., 2023). Considering both underpredicted dry matter burned and uncharacterized VOCs, we infer that BB contributed ∼ 10 %–45 % (or 240–2040 GgC) of the total VOC primary emissions in the western USA during the 2018–2019 fire seasons, which is far more significant than common model representations as in Fig. 1.

We performed nested GEOS-Chem simulations and compared them with observations from two recent airborne campaigns and nine surface sites to constrain the BB CO and VOC emissions in the western USA. We evaluated three widely used global BB emission inventories including potentially significant errors in their dry matter burned, fire detection efficiency, injection heights, and emission ratios. Based on the model–observation comparison, we provided updated emission estimates of BB VOCs for both modeled and uncharacterized VOCs during two different fire seasons in the western USA.

In the standard GEOS-Chem, BB VOC emissions in the western USA rank as third in the rankings of total VOC primary sources (including biogenic and anthropogenic emissions). Despite large interannual variability, the western USA accounted for 60 %–90 % of BB VOC emissions over the CONUS in 2018 and 2019. Across three global BB inventories, total BB VOC emission estimates in the western USA agreed with each other within 30 %–40 %. However, estimates for individual VOCs can differ by a factor of 1–5, mostly driven by regionally averaged emission ratios (reflecting a combination of assigned ERs for specific biome and vegetation classifications) rather than effective biomass burned.

We found that simulations driven by three different BB inventories produce similar CO and VOC abundances. The model reproduced the plume enhancements in the locations observed in WE-CAN but showed negative biases for CO and VOCs (except MEK). Better model performance was found in relatively clean environments. By comparing BB-impacted abundance enhancements between no-/low-smoke times and the entire campaign, we found that the model, regardless of which BB inventory was used, underestimated the BB emissions for primary compounds by a factor of 3–7; these include CO, benzene, toluene, and propane. For OVOCs that have both primary and secondary sources including formaldehyde, formic acid, acetic acid, and lumped ≥ C3 aldehydes, tripling the BB emissions cannot fully explain the negative model bias; the model–observation comparison likely pointed to a large amount of missing secondary production in BB-impacted conditions in GEOS-Chem, which could account for the remaining bias. Unlike other VOCs, MEK was overestimated by a factor of 2–4 throughout the middle to lower troposphere, due to the overestimation of MEK itself or its precursors in the EPA NEI and MEGAN emission inventories. Tripling the BB emissions in GFAS reproduced observed mixing ratios for primary compounds but showed no or less significant improvement for OVOCs.

We found that the fire detection products in all the inventories detected the large fires sampled in the WE-CAN campaign. GEOS-Chem vertical profiles were not strongly sensitive to the various tested BB injection height schemes, as constrained by the observed VOC vertical profiles during WE-CAN. This is likely because strong and efficient vertical mixing during hot and dry summers in the western USA dominates the vertical transport processes. However, different injection height assumptions influenced the modeled downwind surface ozone mixing ratios (i.e., daily mean ozone differed by up to 14 % or 4 ppb); thus, the influence of injection heights on surface air quality requires further investigations.

We evaluated modeled ERs with WE-CAN (and FIREX-AQ) observations and found that GFAS performs slightly better than the QFED or GFED4 inventories for both VOC-CO correlations and ER values. GEOS-Chem + GFAS captured the observed ERs in aircraft emission transects within 30 % for formaldehyde, acetaldehyde, benzene, toluene, and lumped ≥ C4 alkanes and within 50 %–70 % for ethane, propane, and acetone. We also found the modeled abundance and ERs of xylenes and lumped ≥ C3 alkenes are significantly underestimated across all inventory experiments, likely reflecting the overestimation of the OH reaction rate constant kOH used in the model.

Given that the errors in fire detection, plume injection, and ERs are relatively small, we infer that the underestimation of BB emissions in these inventories (a factor of 3–7) is likely due to underpredicted dry matter burned, which is also supported by our back-of-the-envelope calculation of effective DM burned. However, we cannot rule out the uncertainties in the nested GEOS-Chem (0.25∘ × 0.3125∘) not being able to fully resolve the sub-grid features of BB emissions. Therefore, the above findings revealed by 2018 WE-CAN observational constraints are further tested for their regional representativeness and interannual variability with observations from nine western US ground sites and the 2019 FIREX-AQ airborne campaign. Compared to the ground-based “downwind” CO measurements, GEOS-Chem + GFAS captures the observed BB smoke events but underpredicts the mixing ratios in most cases. Tripling the BB emissions reduces the model's negative bias across the western USA without degrading the correlation coefficients with observations. Repeating the analyses with FIREX-AQ observations also confirms the above conclusions.

Constrained by 2018 and 2019 airborne and ground measurements, the 14 BB VOCs included in the model contributed to 120–1020 GgC of primary emissions in the western US 2018–2019 fire seasons. However, the total emission ratio relative to CO for these 14 VOCs in GEOS-Chem only accounted for half of that from the 161 measured VOCs in wildfire smoke, pointing to a significant quantity of uncharacterized reactive organic carbon fluxes that were missing in many current BB emission inventories and CTMs. Thus, accounting for both these missing species and underestimated DM burned, the total BB VOC emission estimates can reach 240–2040 GgC or 10 %–45 % of the total primary VOC emissions in the western US fire seasons, highlighting a significant role of wildfires in US air quality.