The GFSv16 is the current operational global forecast system in NOAA/NCEP using the FV3 dynamical core. Its detailed configuration can be found in Campbell et al. (2022) and Yang et al. (2020). Compared with the previous version (v15), GFSv16 updated many physical schemes (Table 1) and added the parameterization for sub-grid-scale nonstationary gravity-wave drag. To use the GFS's meteorology to drive CMAQ, a meteorological coupler, NACC, is developed (Campbell et al., 2022). Differing from the original MCIP, which was developed to process WRF/ARW meteorology for CMAQ, the NACC coupler interpolates non-native-grid meteorology to a user-defined grid and has parallel processing capability, which drastically reduces its run time for operational forecasts (Campbell et al., 2022). Currently, NACC employs two horizontal interpolation methods: bilinear and nearest neighbor. In this study, we use the nearest-neighbor method for categorical (discontinuous) variables that include land use types, vegetation fraction, terrain elevation, Monin–Obukhov length, friction velocity and soil temperatures, whereas the bilinear interpolation is used for mainly smoothly varying (continuous) meteorological variables that include wind fields, temperature, pressure and specific humidity. The CMAQ model is defined in the Arakawa C-grid (Arakawa and Lamb, 1977); thus, the GFSv16 horizontal wind components (U, V) need to be interpolated to the perpendicular cell faces instead of the cell center (Otte and Pleim, 2010) after rotation to the defined map projection. The scalar variables are defined in the cell centers of the targeted grid; thus, their interpolations are more straightforward: from GFSv16 A-grid to CMAQ A-grid. The NACC coupler can either use the native layers or interpolate inputs to a set number of user-defined CMAQ vertical layers. The GFSv16 has 127 vertical layers with global coverage at roughly a 13 km horizontal resolution, where the targeted CMAQ domain is in 12 km × 12 km grids over the contiguous United States (CONUS) with 35 vertical layers (Campbell et al., 2022). Here, we use 24 h GFSv16 forecasts starting at 12:00 UTC each day.

Most variables needed by CMAQ are directly interpolated from the GFSv16 hourly outputs. The NACC processor has options to calculate diagnostic variables, such as the planetary boundary layer (PBL) height, if they are needed. In this study, we use the interpolated GFSv16 PBL height instead of the diagnostic one. It also has an option to import the externally provided land surface variables. Here, we import the updated the 2018–2020 climatological averaged leaf area index (LAI) and NOAA near-real-time (NRT) greenness vegetation fraction (GVF) from satellite-based Visible Infrared Imaging Radiometer Suite (VIIRS) retrievals (Campbell et al., 2022). The updated satellite-based LAI and GVF impact CMAQ's biogenic emissions and dry-deposition processes, which were described in detail in Campbell et al. (2022).

For comparison with GFSv16 meteorology processed by NACC, a corresponding WRF version 4.0.3 (Skamarock et al., 2021) simulation is run covering the NAQFC's native grid, which is a 12 km horizontal resolution with Lambert conformal map projection over the CONUS. Table 1 shows the WRF configuration, which is commonly employed in CONUS meteorological and air quality studies in the community, versus the current NOAA/NWS operational global model GFSv16. GFSv16 uses the NOAA/NCEP's Global Data Assimilation System (GDAS) (https://www.emc.ncep.noaa.gov/data_assimilation/data.html, last access: 24 October 2022) for its initial conditions and runs on its own global dynamics and physics without any other constraints. The regional WRF simulation uses GFSv16 for its initial conditions. In this study, GFSv16 was re-initialized with GDAS every 24 h, and WRF conducted the continuous run after its spin-up period. Furthermore, WRF also takes its lateral boundary conditions from GFSv16 every 6 h. For the WRF run, we have enabled 4D data assimilation (FDDA) for the horizontal winds, temperature and humidity (Table 1) every 6 h, thereby nudging towards GFSv16. This nudging method used in WRF runs can help reduce the difference between two meteorological models, although its effect may vary depending on events because WRF and GFS use different physics.

WRF and GFSv16 have similar settings for the land surface model, surface layer and radiation schemes; however, their microphysics and PBL schemes are different (Table 1). Compared with the 35-layer WRF model with a 100 hPa domain top, GFSv16 has a much higher domain top (0.2 hPa) and 127 vertical layers, which are interpolated by NACC to 35σ layers up to 14 km for CMAQ. We use NACC (inherited from MCIP version 5.0) to process WRF hourly meteorology while also maintaining the 35-vertical-layer structure. Thus, in contrast to GFS-CMAQ, the WRF-CMAQ system uses the native grid without interpolation.

Here, CMAQ version 5.3.1 (Appel et al., 2021) is used with the Carbon Bond 6 revision 3 (CB6r3; Yarwood et al., 2010, 2014; Luecken et al., 2019) chemical mechanism and AERO7 (version 7 of the aerosol module) treatment of secondary organic aerosols (CB6r3_AE7_AQ). CMAQv5.3.1 includes a series of scientific updates from the previous version (Appel et al., 2021), such as the updated air–surface exchange and deposition modules, which showed a significant impact on ozone (O3) prediction compared with the previous NAQFC (Campbell et al., 2022). We also include the bidirectional NH3 (BIDI-NH3) exchange model for NH3 surface fluxes. An updated Biogenic Emissions Landuse Dataset v5 (BELD5) is used in this study to drive the in-line Biogenic Emissions Inventory System (BEIS) version 3.61. The anthropogenic emissions are provided by the National Emissions Inventory Collaborative (NEIC) with base year 2016 version 1 (NEIC, 2019). We replace the US EPA default CMAQ dust emissions model with an in-line windblown dust model known as “FENGSHA” (Fu et al., 2014; Huang et al., 2015; Dong et al., 2016). The FENGSHA dust scheme uses the sediment supply map and magnitude of the friction velocity (USTAR) compared with a threshold friction velocity (UTHR) to calculate the potential of dust emission flux. The UTHR depends on the land cover, soil type (clay fraction) and soil moisture (Campbell et al., 2022).

We have updated the wildfire emissions system in CMAQv5.3.1 based on the Blended Global Biomass Burning Emissions Product (GBBEPx) (Zhang and Kondragunta, 2006; Zhang et al., 2011). The GBBEPx uses satellite-detected fire radiative power (FRP) to estimate wildfire smoke emissions for a number of species: CO (carbon monoxide), NOx (nitrogen oxides), SO2 (sulfur dioxide), elemental carbon, primarily emitted organic aerosols and fine particulate matter (PM2.5). We derive the wildfire volatile organic compound (VOC) emissions from GBBEPx CO emissions based on the emission ratios of Fire INventory from NCAR (FINN) data (Wiedinmyer et al., 2011), and the splitting factors from the SMOKE model (Baker et al., 2016) are used to further split the total fire VOC emissions to speciated hydrocarbon emissions. The satellite FRP is estimated from the satellite brightness temperature anomaly, and the GBBEPx processor assumes that the wildfire emissions are proportional to the FRP over certain land use types in certain regions. The GBBEPx emissions are based on polar orbiting satellites: MODIS (Aqua and Terra satellites) and VIIRS (Suomi NPP and NOAA-20 satellites) instruments, which are updated once per day. A wildfire emission preprocessor converts the GBBEPx emissions to CMAQ-ready input files using emission speciation and diurnal profiles (high during daytime and low at night), which are adopted from US EPA-based profiles (Baker et al., 2016), and a daily scaling factor. Here, we classify the wildfire into either a long-lasting fire (longer than 24 h) or a short-term fire (shorter than 24 h) based on land use types and regions. As historic statistics show that most fires (> 95 %) in the region east of 110∘ W last less than 24 h, only fires west of 110∘ W that have a model grid cell total forest fraction > 0.4 are assumed to be long-lasting fires. All other short-term GBBEPx fires are assumed to have smoke emissions for 24 h (i.e., day 1 only). Burning area could be highly uncertain, as GBBEPx data do not include this information. One grid cell could have multiple fires, and some big fires could appear in several grids. Here, we carry the previous NAQFC's method and apply a constant ratio, 10 % of the grid cell, as the burning area (Pan et al., 2020), according to Rolph et al. (2009). CMAQ treats wildfire emissions as point sources that undergo in-line plume rise to distribute the smoke vertically. The default CMAQ plume rise used here is based on Briggs (1965) and is driven by fire heat flux (converted from FRP with a ratio of 1) and fixed burning area (assumed to be 10 % of the 0.1∘ × 0.1∘ grid cell).

Figure 1 shows the mean bias (MB) of interpolated GFS- and WRF-predicted surface meteorological variables compared to METAR data during August 2019. Both meteorological models have a cool bias over the western and northeastern US and a warm bias over the western Rocky Mountains region and southeastern US (Fig. 1a, b). Similar temperature predictions are expected because WRF uses the FDDA method, nudging toward GFS data. However, GFS tends to be cooler than WRF over the Rocky Mountains and in the central and northeastern USA due to their different dynamics and physics. The GFSv16 cold bias in the lower troposphere is impacted by excessive evaporative cooling from rainfall (personal communication with NOAA/NCEP, 2021). Campbell et al. (2022) provides detailed discussions about GFSv16 biases.

Both the GFSv16 and WRF models have similar and rather significant dry biases for specific humidity (SH) predictions across the CONUS domain (Fig. 1c, d). Qian et al. (2020) investigated this common dry bias in many models and found that neglecting an irrigation contribution could cause this dry bias. Besides this issue, WRF's dry bias could also be affected by its nudging toward GFS, as GFS has widespread dry biases (Campbell et al., 2022). Their biases also have some noticeable differences over certain regions. For instance, WRF has less dry bias over southern Texas than GFS.

Both models underestimate the mean 10 m wind speeds compared with METAR stations over the western US: WRF has stronger underpredictions over the Rocky Mountains and overpredictions over northeastern US, whereas GFS has stronger underpredictions over the Appalachian Mountains and overpredictions over Texas and Oklahoma. GFSv16's operational verification (https://www.emc.ncep.noaa.gov/gmb/emc.glopara/vsdb/v16rt2/g2o/g2o_00Z/index.html, last access: 24 October 2022) also shows that it tends to underpredict the 10 m wind speed over the western US during both daytime and nighttime, but it shows overpredictions over the eastern US. Besides the difference in physical schemes (Table 1), for example, other possible reasons causing this surface wind difference could be effect of gravity-wave drag (GFSv16 includes it, but the WRF run here does not) and vertical resolutions (GFS's 127 layers versus WRF's 35 layers, although they have similar vertical layers below 1 km) (Campbell et al., 2022). Some studies (Skamarock et al., 2019) have revealed the necessity for a fine vertical resolution for atmospheric simulations, especially within the PBL, near the tropospheric top and during convective events. Insufficient vertical resolution could also cause plume dilution in chemical transport modeling (Zhuang et al., 2018). The gravity-wave drag is also known to influence the synoptic-scale dynamics of the atmospheric flow over irregularities at the Earth's surface, such as mountains and valleys, and the uneven distribution of diabatic heat sources associated with convective systems (Kim et al., 2003). Its parameterization is needed for large-scale models.

There are strong regional variabilities in the monthly mean PBL height differences between GFS and WRF during normal daytime (represented by 18:00 UTC) and nighttime (represented by 06:00 UTC) (Fig. 2). During daytime, GFS has a higher PBL height compared with WRF over the US Pacific coast, northern Rocky Mountains, and northeastern and southeastern US, but it becomes lower over the central US (e.g., Texas, Oklahoma and Kansas). At night, however, most of these regional differences between GFS and WRF are reversed. This diurnal difference is mainly driven by the different PBL schemes employed in GFS (Han and Bretherton, 2019) and WRF (i.e., YSU) and the associated other physical suites, including the land surface data. Hence, this PBL difference has strong regional variations depending on geographic differences. The GFS's PBL height has a strong diurnal variation over these regions, including the western and northeastern US, and it shows a sharp rise and collapse after sunrise and sunset, respectively (Campbell et al., 2022). These two selected times (18:00 and 06:00 UTC) are not in the transition periods for fast diurnal changes in the PBL.

The US EPA AIRNow network provides hourly observations of near-surface O3, PM2.5 and meteorology. Campbell et al. (2022) showed detailed verification of GFS-CMAQ with the surface AIRNow data. Here, we focus on the difference between the interpolation-based GFSv16 and WRF downscaling as well as the impacts on meteorological and chemical model performance. Figure 3 shows a comparison of these two models over two specific regions, the western US (CA, OR and WA) and the northeastern states (CT, DE, MA, MD, ME, NH, NJ, NY, PA, RI, VT and the District of Columbia) (Fig. S1 in the Supplement), where the two models have relatively large differences for some meteorological variables. GFS and WRF predict very similar 2 m temperatures over the Pacific coast states: WA, OR and CA, and both of them had a similar cool bias (around 1 K), R value and RMSE (Fig. 3a). However, these two models show significant differences with respect to the 10 m wind speed prediction over the Pacific coast (Fig. 3c), where WRF overpredicts the wind speed, especially at night and in later August. Most AIRNow stations are located near urban or suburban areas, which generally have a weaker 10 m wind speed than those at the METAR aviation weather stations near airports. For this reason, although Fig. 1e and f show that GFS and WRF underpredict the monthly mean wind speed over the METAR stations in the west, they still tend to overpredict wind (Fig. 3c) over AIRNow stations, especially for the WRF 10 m wind speed at night. Considering that the model grid cells represent 12 × 12 km2 averages, the true model–observation comparisons likely fall somewhere between the urban/suburban AIRNow stations and METAR stations, depending on the land use fractions of each grid. Obviously the representation characteristics of observations could affect the verification results. Compared with AIRNow observations, GFSv16 has overall better scores for surface wind speed predictions over the western US, where the WRF's higher surface wind speed overprediction is associated with its PBL height predictions (Fig. 3e, f). During the nighttime, GFS has a lower PBL height (10 %–50 % lower than WRF) and weaker vertical mixing, which tends to bring less momentum flux from the upper layers to the surface, leading to lower nighttime wind and better agreements with the AIRNow wind speed observation.

Over the northeastern US, the mean bias (MB) of GFS temperature is about -1 K, whereas the WRF model has a slightly warm MB of about 1.53 K (Fig. 3b). However, the GFS's temperature prediction has a better correlation coefficient (R) and RMSE, implying that it better captures some events, such as the storm on 28–29 August. Both models overpredict 10 m wind speeds in the northeast, but the GFS model yields better results due to a slightly lower PBL height at night (Fig. 3f) compared with WRF, which had significant overpredictions, especially during 25–29 August (Fig. 3d) when tropical storm Erin approached this region. Especially on 28 August, when the storm was centered near the east coast of North Carolina, WRF significantly underpredicted the 2 m temperature (Fig. 3b) and overpredicted the 10 m wind speed (Fig. 3d). In the west around the same period, tropical storm Ivo appeared southwest of the Baja California Peninsula, bringing heavy rainfall to Mexico. Associated with this storm, a low-pressure system expanded over most of the western US. Differing from GFSv16, which is designed for the operational meteorological forecast, the WRF configuration used in this study is normal for driving CMAQ, but it is not tuned for storm weather prediction.

Figure 4a and b show the O3 predictions of the two models over these two regions, and GFS-CMAQ yields predominantly lower O3 than WRF-CMAQ, especially at night. Over the west, the lower O3 in GFS-CMAQ is associated with their PBL height difference. First, with a certain dry-deposition velocity between the models, it is easier to deplete O3 given the smaller volume of a shallower PBL. Second, the shallower PBL results in higher surface NOx concentrations (not shown) and O3 titration rates near NOx source regions, consequently resulting in lower O3 in these areas at night. Last, the lower PBL could decouple from the residual layer and result in weaker or no vertical O3 exchange with the residual layer at night (Caputi et al., 2019). All of these factors contribute to the lower nighttime O3 of GFS-CMAQ compared with WRF-CMAQ. As GFS-CMAQ already underpredicts O3 due to combined meteorological factors, such as the temperature underprediction (Fig. 4a), the GFS-CMAQ's further O3 reduction (possibly due to its lower PBL height at night) exacerbates its low bias. However, over the northeastern US, the similar impacts help the GFS-CMAQ yield a much better MB due to its better agreement with the observed nighttime low O3 over this region. Over the entire CONUS domain, the situation is similar, and the GFS-CMAQ has a lower ozone MB (1.1 ppb) compared with WRF-CMAQ (4.7 ppb). Figure 5 shows that both models have similar daytime O3 prediction over the CONUS. However, GFS-CMAQ better captures low nighttime O3 over the eastern US than WRF-CMAQ (Fig. 5c, d).

GFS-CMAQ has substantially higher PM2.5 mean concentrations over the western US but lower mean concentrations over the northeastern US compared with WRF-CMAQ (Fig. 4c, d). These model differences are also related to their interpolated GFSv16 versus downscaled WRF meteorological drivers. Because both models use the same emissions under relatively clean background conditions in the west (i.e., prevailing westerly flow from the Pacific Ocean), the PBL and wind speed differences have significant impacts on their near-surface pollutant concentrations, especially at night. Both models show strong PM2.5 diurnal variations (high at night and low during daytime), driven by the meteorological diurnal variation (e.g., PBL), which overcomes the emission diurnal variation (usually high during daytime and low at night). Compared with WRF-CMAQ, GFS-CMAQ has a lower nighttime PBL height and a weaker wind speed at night, which lead to weaker vertical mixing and venting, increasing the pollutant concentrations near the surface and yielding higher surface PM2.5 over the western US (Fig. 4c). Its higher surface PM2.5 could also result in stronger local dry deposition. In contrast to the local vertical mixing and venting effects on PM2.5 discussed above, there are strong (and potentially counterbalancing) impacts of model PBL and horizontal wind speed differences on downstream PM2.5 concentrations at night. WRF-CMAQ's deeper PBL and stronger wind speeds at night (Fig. 3c, d, e, f) tend to transport aerosols and their precursors more efficiently downstream via the dominant advection pathway. Figure 6 shows that these monthly mean background PM2.5 differences appear east of the Rocky Mountains (WRF-CMAQ is about 2 µg m-3 higher) during both daytime and nighttime. This effect is very prominent in the northeastern region. Although both models predicted a similar PM2.5 magnitude over the northeastern US, GFS-CMAQ yields an overall PM2.5 underprediction, and its monthly mean PM2.5 is 2.6 µg m-3 lower than the WRF-CMAQ prediction (Fig. 4d). Especially during the 1–9 August period, WRF-CMAQ had about a 4 µg m-3 higher surface PM2.5 background than that of GFS-CMAQ. In this case, the WRF-CMAQ model shows better agreement with observations (Fig. 4d). It is possible that the GFS-CMAQ's nighttime PBL heights (wind speeds) are too shallow (weak) in this case, which does not allow enough transport of pollutants downstream (to the eastern USA). Overall, GFS-CMAQ and WRF-CMAQ show mixed performance with respect to PM2.5 predictions during the August 2019 period: GFS-CMAQ has better PM2.5 prediction over the western US, and WRF-CMAQ yields better results over the region east of the Rocky Mountains (Fig. 6).

On 22 July, the DC-8 aircraft flew from California to Boise, ID, while maintaining a relatively low altitude (< 1 km a.s.l., above sea level) over the California Central Valley (Fig. 7). This flight was not impacted by any major wildfire event and was mainly controlled by anthropogenic emissions and local meteorological conditions. Figure 7 shows that the GFSv16 and WRF models had similar meteorological temperature and humidity predictions and that both models have dry and warm biases over the Central Valley at lower altitudes (Fig. 7d, e) (Qian et al., 2022). GFS's horizontal wind speeds tend to have a stronger variability than WRF (Fig. 7b), especially in high altitudes. With respect to wind direction, WRF shows a better prediction than GFS around 20:00 and 24:00 UTC (Fig. 7c).

Both GFS-CMAQ and WRF-CMAQ underestimate the vertical wind (W) variability by at least 1 order of magnitude, and WRF-CMAQ has weaker W variability than that of GFS-CMAQ, especially at high altitudes (Fig. 7f). The model vertical velocities are not directly from the GFS nor the WRF model; rather, they are re-diagnosed in CMAQ to conserve mass (Otte and Pleim, 2010) and, thus, represent the whole layer's vertical movement across the 12 km × 12 km grid cell. With its flight speed of around 80 to 240 m s-1, the DC-8 aircraft's 1 min average sampling frequency results in an approximate 4.8 to 14 km horizontal scale, respectively, which is comparable with the 12 km CMAQ model resolution. The aircraft observations, however, include turbulence effects during its 1 min averages, which may not be temporally resolved by the models at this resolution. Thus, both the GFS-CMAQ and WRF-CMAQ vertical velocities are much lower and have almost no correlation with the aircraft observations.

Although both GFS-CMAQ and WRF-CMAQ have reasonable comparisons for most meteorological variables, including the horizontal winds, it continues to be a challenge to compare them with the observed vertical velocities. Thus to further elucidate the model–observation differences in vertical motion, Fig. 8 shows a curtain plot of vertical velocities along the flight path from the two models. As WRF-CMAQ remains on a native grid, its wind fields tend to be more balanced and have lower variability compared with the GFS-CMAQ wind fields. The stronger variability in W for GFS-CMAQ could be caused by GFS's non-hydrostatic dynamics or CMAQ's effort to counteract mass-inconsistency effects from the interpolated horizontal wind fields (Byun, 1999b). Our comparison shows that the first factor should be the major one (Fig. S2) for this event, as the GFS-CMAQ-diagnosed W is very similar to that from the original GFSv16 around 1 km a.g.l. (above ground level). As the original GFSv16 also has similar stronger vertical velocities compared with the original WRF, the W difference between GFS-CMAQ and WRF-CMAQ is unlikely to be due to interpolation error in the horizontal winds.

GFS-CMAQ and WRF-CMAQ yield similar results overall for specific chemical species during this DC-8 flight (Fig. 9). Both models underestimate CO, O3 and ethane (C2H6) concentrations over the lower altitudes in the California Central Valley. Over the same flight segment, they have better NOx (NO + NO2) and ethene (C2H4) predictions, implying that the emissions of these two species have better accuracy than those of CO and C2H6. Figure 9f shows that the two models also underestimate NOz (NOy–NOx), or the oxidized nitrogen species besides NOx, indicating that photochemical O3 production may also be underestimated. NOz is a good indicator of the O3 photochemical formation (Sillman et al., 1997), where the O3/NOz ratio represents the O3 photochemical efficiency per NOx oxidation product. Thus, NOz and O3 are typically highly correlated over regions with active photochemical production. Our later analysis shows that the models tend to underestimate certain hydrocarbons, such as C2H6, which is likely linked to O3 and NOz underestimations, as the hydrocarbons are photochemical precursors of O3 and NOz.

The two models show slight differences in peak values of CO, C2H4 and NOx around 23:30 UTC: the GFS-CMAQ-predicted concentrations are slightly higher and closer to observations (Fig. 9). These differences are due to their PBL predictions (both from the corresponding meteorological model outputs): GFS-CMAQ has a lower PBL height and weaker emission vertical dilution compared with WRF-CMAQ (Fig. 8). However, GFS-CMAQ tends to underpredict O3 more (Fig. 9b) due to its higher NOx titration. This implies that the effects of the transport and nonlocal transformation of O3 could be stronger than that of local precursor emissions. WRF-CMAQ has higher NOz (Fig. 9f) but lower NOx compared with GFS-CMAQ due to the time lag of O3 and NOz photochemical formation. Consequently, the peak O3 values may not be well correlated with the emitted precursors, such as NOx and VOCs. Furthermore, the modeled peak C2H6 and C2H4 concentrations do not occur at the same time around 23:30 UTC, whereas observations indicate that these two species should be highly correlated in this region. This model mismatch implies that the VOC speciation factors for a certain area or emission sector need to be improved over Southern California.

On 6 August, the DC-8 observed a cluster of three wildfires: the Williams Flats fire (47.98∘ N, 118.624∘ W; 80 km to the northwest of Spokane, WA), the Snow Creek fire (47.703∘ N, 113.4∘ W; 32 km northeast of Condon, MT) and the Horsefly fire (46.963∘ N, 112.441∘ W; 24 km east of Lincoln, MT). Figure 10a shows the flight path on that date: the DC-8 aircraft departed from Boise, ID; flew over the Williams Flats fire region; flew to Montana to sample the Snow Creek and Horsefly fires (i.e., Montana fires); and finally returned to the Boise base. The aircraft flew below 8 km for most flight segments near the fire plumes. Figure S3 shows the corresponding GOES-16 satellite true-color image, where these 6 August fires and the associated smoke plumes are visible and can be distinguished from the cloud bands to the south that move northward later that day (Fig. S3). The Williams Flats fire was ignited by lightning and was the largest fire event sampled during the FIREX-AQ campaign, burning from about 2–8 August 2019.

Both models significantly underpredicted CO (Fig. 10c), submicron (aerosol diameter < 1 µm) organic aerosol (OA) (Fig. 10e) and the aerosol optical extinction coefficient (AOE) (Fig. 10f), suggesting an issue with the GBBEPx gas and aerosol emissions. The models performed well for NO2 during the Williams Flats and Montana fires below 6 km a.s.l., but there were prominent underestimations for the high-altitude flight segments (Fig. 10d). However, as the NO2 instrument (the NOAA NOyO3 four-channel chemiluminescence instrument) had an NO2 detection limit of around 0.01 ppb (https://airbornescience.nasa.gov/sites/default/files/documents/NOAA NOyO3_SEAC4RS.pdf, last access: 24 October 2022), the models might not truly underestimate NO2 for these flight segments with extremely low NO2. WRF-CMAQ predicted higher O3 values than the GFS-CMAQ, which generally agreed better with observations for the Williams Flats fire (Fig. 10b). However, for the Montana fires (∼ 23:00–24:00 UTC), WRF-CMAQ has higher O3 biases, and GFS-CMAQ yields better results. The difference in O3 is largely driven by the regional background concentration difference between the two models: WRF-CMAQ tends to have higher domain-wide O3 concentrations than GFS-CMAQ due to the meteorological effects discussed in Sect. 3, even though they used the same lateral boundary conditions.

Figure S4 shows the spatial overlay comparison of vertically averaged GFS-CMAQ predictions at 21:00 UTC and the DC-8 flight observations for the altitude of 1–3 km a.g.l. on 6 August 2019. The peak NO2 observation around 48∘ N, 118.5∘ W indicates the general location of the Williams Flats fire. The GBBEPx emissions and GFS-CMAQ prediction showed shifted peak-value locations driven by the westerly modeled winds. For this flight, the GBBEPx had stronger NOx fire emissions over two Montana locations than that over Williams Flats. The model overpredicts the column-averaged NO2 concentrations, especially over the Montana fires, which can not be reflected by the point-by-point NO2 comparison result in Fig. 10d. For this flight, the mean GFS-CMAQ NO2 along the flight path for 1–3 km a.g.l. is about 0.125 ppbv compared with the observed mean NO2 of 0.169 ppbv, and the model indeed showed an NO2 underprediction along the flight path. However, in this case, the flight path did not encounter the locations with modeled peak NO2 values, as the model misplaced the plumes, especially over the Montana fires, leading to this inconsistency. With respect to the O3 comparison (Fig. S4b), this inconsistency could also exist, although it may not be as significant as the inconsistency for the high-gradient NO2 concentrations. In the GFS-CMAQ prediction, the high O3 concentrations are almost co-located with high NO2 concentrations (Fig. S4b), but the observations did not show this feature. Instead, some high-O3 flight segments had relatively low NO2 concentrations, such as those circled in the black rectangle in Fig. S4b. The observed NOx titration was not able to be produced by the 12 km models. Wang et al. (2021) used a 100 m horizontal resolution large-eddy simulation and demonstrated the capability of using such techniques to capture some high-resolution features of fire plumes and the associated chemical behavior. While such high-resolution techniques are not currently feasible for the operational NAQFC, they demonstrate the limitation of using regional-scale (12 km × 12 km) models to capture such fine-scale features of the fire plume.

GFS-CMAQ has higher wildfire-related CO, NO2, OA and AOE values that are closer to observations than WRF-CMAQ for the Montana fires between 23:00 and 24:00 UTC at flight altitudes of ∼ 4–5 km (Fig. 10c, d, e, f). As these two models use the same GBBEPx emissions and wildfire plume rise algorithm (Briggs, 1965), the differences should be due to other reasons. To help explain these model differences, Fig. 11a and b show the aerosol backscatter coefficients (ABCs) retrieved by the differential absorption high-spectral-resolution lidar (DIAL-HSRL) aboard the DC-8 aircraft without and with cloud screening, respectively. It shows that the major fire plumes of the William Flats fire were below 4 km (∼ 19:00–22:00 UTC), but the Montana fires (∼ 23:00–24:00 UTC) extended from the surface up to 6 km, with some detached plumes reaching 10 km. The model-predicted AOEs have an overall similar pattern, with major plumes below 4 km for the Williams Flats fire (Fig. 11c, d). Over the Montana fires, the GFS-CMAQ predicts a slightly higher PBL height, thereby allowing the fire plume to reach a higher height near the DC-8 cruising altitude. In contrast, the WRF-CMAQ wildfire plumes are slightly lower than the aircraft flight path around 23:00–24:00 UTC, which leads to underpredictions of the fire-emitted species (Fig. 11d).

An interesting feature in the DIAL observations is the detached plume from 8 to 10 km altitude (Fig. 11a): some cirrus clouds existed in this region, and the DIAL retrieval could not distinguish whether they were pure clouds or clouds mixed with elevated aerosols above 8 km. The cloud-screened image (Fig. 11b) mainly showed the enhanced aerosols below 7 km and some scattered signals near the high cloud edges. Cloud mixing with aerosols was usual for fire-induced clouds, or pyrocumulonimbus (Peterson et al., 2021). Although, in this event, the middle-sized fires did not show evidence of inducing high-altitude clouds, the indicators of mixed clouds and aerosols at high altitudes still existed: both OA measured in situ (Fig. 10e) and the AOE (Figs. 10f; 11c, d) showed the enhanced aerosols around 01:00 UTC of the next day above 8 km. This elevated plume was generally captured by the GFS-CMAQ simulation, although its strength was underestimated (Fig. 11c); however, this feature was completely missed in WRF-CMAQ (Fig. 11d). Considering the altitude range of the detached plume, the major model disparities are likely due to model convection differences in the free troposphere. To further investigate this impact, Fig. 11e and f show curtain plots of relative humidity (RH) predicted by the two models. GFS-CMAQ yields higher RH at such altitudes (10 km) compared with WRF-CMAQ around 23:00–24:00 UTC, indicating that GFS-CMAQ has a stronger convection. The CMAQ model uses input meteorology to diagnose convection activity and drive its Asymmetric Convective Model, version 2 (ACM2) convection scheme. This convective activity is apparent in GOES-16 satellite images (Fig. S3), as more fractional clouds appeared ahead of the northward-moving frontal band. Both the GFSv16 and WRF models used here do not consider the fire heat feedback effect; thus, their predicted convection and clouds are only driven by the synoptic weather conditions. If such synoptic-to-mesoscale weather models consider wildfire heat feedback effects, their predictions may result in stronger convection and help correct their underpredictions of PBL heights.