**4. Discussion**

The aim of our WRF-SFIRE evaluation was to assess its ability to capture fire-generated emissions in the context of air quality. Hence, we examined the implications of the above results based on their potential applications for wildfire smoke plume rise and dispersion modeling. The following sections discuss model performance and accuracy from the perspective of atmospheric dynamics, as well as address potential implications of uncertainty in fire behavior and the associated input parameters.

#### *4.1. Vertical Plume Rise in the Boundary Layer*

As demonstrated in our results summary in Section 3.2, initially WRF-SFIRE produced a fairly accurate near-source emissions distribution and plume top with a slight under-prediction of concentrations (Figure 5a).

Over time model performance appears to deteriorate. Given that the fire thermal forcing compares relatively well with observations (Section 3.1), a more likely cause for the increasing difference between model and observations is background boundary layer dynamics. The atmosphere was initiated with 10:00 CST sounding, and continually forced with an observations-based constant surface heat flux. However, the cyclic lateral boundary conditions maintained the same vertical wind profile as initially supplied by the sounding at 10:00 CST, irrespective of potentially changing mesoscale conditions in the real atmosphere. Over the course of more than three hours between spin up start and the final minutes of the fire simulation, from which the corkscrew emissions distribution was obtained (Figure 5b), the real atmospheric wind profile likely evolved.

With time and further downwind the effects of any small changes in mesoscale conditions become more pronounced, which is why initially encouraging model performance deteriorated towards the end of the simulation. The markedly narrow band of emissions in Figure 5b suggests that the "corkscrew" location in the LES domain corresponded to the very edge of the plume rather than the center, indicating a shift in mesoscale wind conditions.

Indeed, analysis of observed background 30 m wind direction leading up to and during the burn shows a significant shift to the west, resulting in the LES "corkscrew" profile being extracted from the edge of the plume, rather then the intended center (Figure 6). Accounting for this observed wind

rotation, it is possible to extract a wind-corrected profile, such as shown with a red dotted line in Figure 5b. Assuming an average 20 degree rotation over the course of available wind observations (based on the slope of linear regression shown Figure 6a), the corrected location of the corkscrew maneuver indeed corresponds to the center of the plume (Figure 6b). The wind-corrected profile shown in Figure 5b is a notable improvement from the original non-rotated estimate. Note that this adjustment is extremely crude, as it is based on an estimated wind rotation at one point on a single vertical level and does not take into account potential changes in vertical wind shear.

**Figure 6.** The effects of changing mesoscale wind conditions on plume observations (**a**) Observed change in 30 m wind direction prior to and during the burn. Significant linear trend is shown with a red dashed line. (**b**) Top view of modeled smoke plume during the "corkscrew" maneuver by the instrumented aircraft. Black dot and red star indicate the average location of the "corkscrew" profile from flight with and without wind-correction, respectively.

Unfortunately, unlike the Real-mode WRF simulations, there is no easy way to account for changing lateral boundary conditions in WRF-SFIRE large-eddy mode. Hence, we can expect the ability of the model to accurately capture dispersion to depend strongly on the variability of real background conditions as well as the simulation length and spatial extent of the modeled domain. Namely, an LES will provide better simulations for situations where that actual atmosphere is horizontally uniform and temporally steady. While this presents a limitation for smoke plume rise and dispersion modelers, it is important to consider it in the context of existing alternative sources of field data. Given a typical uncertainty of ∼500 m associated with the most accurate widely available plume height dataset from Multi-angle Image SpectroRadiometer (MISR) [22], WRF-SFIRE provides a valuable alternative source for generating comparatively accurate "synthetic plume height data".

Moreover, unlike instantaneous observational point measurements or overpass-limited derived satellite data, the LES allows us to examine the domain-wide temporal evolution of the plume and identify key features, which are likely to be of interest to dispersion modelers. As shown in Figure 7 and Animation S2, the vertical distribution of emissions in the domain changes throughout the simulation. Following an initial overshoot and a period of active smoke production near the ground, most of the emissions rise and end up near the top of the BL, accumulating just under the inversion level in a wide span of heights. While this vertical distribution may contain modeling and initial condition biases, it is likely to offer dispersion modelers an advantage over the common current approach of using a single empirically derived injection height.

**Figure 7.** Evolution of total column CO2 anomaly.

#### *4.2. Importance of Fire Input Parameters*

As noted in the Introduction, our evaluation work focused on assessing the relationship between coupled surface forcing and the atmosphere in WRF-SFIRE rather than on fire behavior. However, as we attempt to move forward from simple uncoupled burner-type experiments with prescribed constant surface heat flux to more realistic dynamic simulations, we must address the challenges in selecting proper fire input parameters.

Similar to Kochanski et al. [5], we found that the fire behavior model is particularly sensitive to the choice of fuel moisture. This parameter in WRF-SFIRE does not depend on the selected fuel category and was based entirely on measurements in our simulation. We also modified the standard fuel depth and loading parameters associated with Category 1 fuels to match observations, which resulted in very accurate surface heat flux forcing but substantially lower ROS values than observed or those obtained with standard settings.

Notably, similar thermal forcing to the atmosphere can be produced using a range of combinations of fuel categories and parameters in the model. We have not carried out a formal sensitivity analysis as it was beyond our scope and computational abilities, however, future modelers may find the following information helpful. As preliminary tests for our study, we have used Category 1 and Category 3 fuels (short and tall grass) with various combinations of both standard and measurement-based fuel depth and loading parameters to achieve similar surface forcing. The relationships between these parameters are highly non-linear, which makes determining the "correct" choice (in the absence of detailed observational data) difficult. What we found to be encouraging is that while the absolute value of modeled concentrations and ROS changes dramatically depending on the chosen fuel category for a given fire intensity, the relative distribution of emissions does not. The simulated atmosphere is forced solely by the parameterized heat and moisture fluxes, so WRF-SFIRE does not discriminate which combination of fuel characteristics produced a given heat flux that drives the buoyant plume rise.

Given any thermal forcing, the atmospheric response appears to be fairly robust, irrespective of the particular combination of fuel parameters or ROS with which it was achieved. While this study does not aim to establish whether the model sensitivity to fuel conditions is physical, it does sugges<sup>t</sup> that the LES produces realistic plume rise for the given fire intensity.

#### *4.3. ROS and Biases in Modeled Emissions*

The model's poor performance for ROS in our case study likely resulted in reduced simulated emissions concentrations due to lower parameterized fuel consumption rate. This is consistent with the notable negative bias in our modeled CO2 profiles.

As mentioned above, the low ROS values on our simulation are largely a result of our use of non-standard fuel depth and loading parameters. To eliminate alternative causes for slow fire line advance, we compared horizontal winds at the first and second model levels (at ∼8 m and ∼25 m AGL) with data obtained from 2D sonic anemometers mounted at multiple heights of the CSU-MAPS meteorological tower. As shown in Figure 8, the near-surface winds are generally accurately captured by the model. At the lowest vertical level, there tends to be a slight positive bias, which one would expect to contribute to higher rather than lower ROS values.

**Figure 8.** Modeled (red) and observed (black) near-surface horizontal wind.

Apart from their dependency on ROS and fuel consumption, the absolute values of WRF-SFIRE emissions are also controlled by user-prescribed emission factors. In our case study, these factors were not derived from measurements, but were rather based on standard values typical for the Grassland fuel category (see Section 2.2). Hence, the negative bias in our modeled smoke distribution could potentially be reduced, should observations-based emissions factors become available.

#### *4.4. Experimental Design Considerations*

One of the shortcomings of the RxCADRE dataset and this experiment is the substantial (nearly 2.5 h) difference in timing between the sounding balloon launch and the fire ignition. Availability of an additional vertical profile for model evaluation just prior to ignition would have been extremely helpful in mitigating some of the sources of error mentioned in the above sections. A similar recommendation was offered by Kochanski et al. [4], who suggested that an on-site sounding just prior to the burn rather than a few hours earlier would be most useful.

While we recognize the challenges of coordinating balloon launches in the presence of aircraft over the fire, a potential alternative would be to include on-board temperature and wind sensor data from flight with the smoke dispersion measurements.
