**2. Methods**

#### *2.1. Observational Data*

The RxCADRE campaign consisted of 10 operational and 6 small replicate prescribed fires. Collected data are accessible via a US Forest Service online repository, as referenced below. Smoke dispersion and emissions measurements are available for three large fires: L1G and L2G grass fires and L2F sub-forest canopy surface fire. For the purpose of model evaluation, we selected L2G (10 November 2012) for our case study, based on its reported uniformity and consistency of flame propagation [10]. Figure 1 shows a sample snapshot of the burn plot during the ignition. The overall meteorological conditions and instrumental design of the L2G experimental burn are described in detail in [9]. The individual datasets obtained from the US Forest Service online archive used for this study are summarized below.

Georeferencing data, including plot location and burn perimeters, are available from Hudak and Bright [11]. Analysis of fire rate of spread (ROS) and intensity as well as a detailed description of three Highly Instrumented Plots (HIPs) used to produce the estimates can be found in [10]. Locations of HIPs are available from Hudak et al. [12]. HIP1, used for this evaluation, is shown in Figure 1. Near-surface wind and temperature sonic anemometer time series for in-situ and background locations are available from Seto and Clements [13,14]. Ignitions timing and locations were obtained from field-grade GPS units, mounted on-board firing vehicles [15]. Fuel data used for this evaluation study included photographs of pre-burn samples, as well as measurements of size, loading and moisture content of species groups. Data collection methodology is detailed in [16]. Dispersion and emissions measurements included volume-mixing ratio of CO2, CO, CH4, and water vapor at a rate of 2 s, obtained from aircraft-mounted sensors [17]. The georeferenced data consisted of horizontal transects at multiple elevations, as well as "corkscrew" and "parking garage" flight profiles.

**Figure 1.** Long wave infra-red (LWIR) image of L2G lot during ignition (12:32:02 CST) with dashed black lines denoting burn perimeters. Red scatter points correspond to Highly Instrumented Plot (HIP) #1 fire behavior packages (FBP), each containing a system of airflow, temperature and energy sensors.

## *2.2. Numerical Setup*

WRF-SFIRE [3] was configured in idealized LES mode. One of the primary advantages of using this model is that it allows for two-way coupling between the fire and the atmosphere. While WRF-SFIRE does not model combustion directly, the spread and intensity of the fire are parameterized using a semi-empirical approach. The latent heat flux is computed based on the fuel consumption and stoichiometric combustion of cellulose. Heat and moisture fluxes from the simulated burn provide forcing to the atmosphere, which in turn influences fire behavior.

A 10.4 km × 14 km domain with 40 m horizontal grid spacing, 3000 m model top and 51 hyperbolically stretched vertical levels was initialized using the 10:00 CST (16:00 UTC) sounding [9]. While this may appear to be a shallow domain compared to mesoscale ("Real") WRF simulations, the choice is substantially higher than that found in existing published WRF-SFIRE evaluations [1,4,18]. Five lowest model grid centers were located at approximately 8 m, 24 m , 42 m, 60 m and 80 m above ground level (AGL). The simulation was allowed to spin up for 2 h 23 min prior to ignition at ∼12:23 CST (time varied slightly for different fire lines). To aid the formation of buoyancy-driven ambient background turbulence typical for a daytime boundary layer (BL), a lower-boundary surface thermal flux (tke\_heat\_flux) was imposed. The value was estimated from the sonic anemometer time series of vertical wind velocity and temperature over the time period leading up to ignition. As shown in Figure 2, based on the measurements, the ambient background surface heat flux remained fairly constant over the entire spin-up period. Hence, the lower-boundary surface forcing was idealized for the LES simulation as being uniform in space and constant in time. We used full surface initialization (sfc\_full\_init =.true.), with the lower boundary moisture flux and surface roughness characteristics set to standard USGS values for "Grassland" land use category.

**Figure 2.** Five-minute averaged kinematic surface heat flux *T w* derived from 1 Hz wind and temperature sonic anemometer time series of the background ambient environment.

To help trigger convection in a horizontally uniform initial domain a small temperature perturbation "bubble" was added (see namelist.input\_spinup in the Supplementary Materials). With periodic boundary conditions, near-stationary turbulence spectrum was achieved within ∼40 min of run start. The well-mixed modeled BL continued to turn over and warm for a total of 2 h 23 min (10:00:00 CST–12:23:00 CST). Restart file generated at 12:23:00 CST was used as initial conditions for the main burn simulation (12:23:00 CST–13:12:00 CST), ensuring the fire was ignited into a well developed BL. Other key configuration details can be found in Table 1, as well as in the complete namelist initialization files provided as Supplementary Materials.



Following the LES spin up, the northwestern half of the simulated L2G lot was ignited with four roughly parallel fire lines mimicking strip head fire method used during the real-life burn (Figure 1). During the campaign, the prescribed burn was ignited with drip torches attached to moving all-terrain vehicles (ATVs). Using GPS data from these vehicles (available from [15]), we extracted the locations of start and end points of the four fire lines, as well as their individual start and end ignition times. While the real-life ignition process was not perfectly uniform in time, the modeled fire lines were approximated as being ignited at a constant speed, such that the time and location of the start and end points matched those of the real burn (see Animation S1 in the Supplementary Materials). Timing varied slightly for each of the four modeled fire lines (see namelist.input\_main in the Supplementary Materials). We approximated the ignitions as straight lines between observed start and end points, as the ATVs' deflections from a straight path during the real burn remained within a single atmospheric grid in our modeled domain.

Ignited cells in WRF-SFIRE proceeded to spread, while each fire line continued to advance until reaching the opposite end of the L2G lot. Subsequent upwind ignitions of the remaining lot area were excluded to reduce the computational load of the simulation. Taking into account the downwind location and timing of smoke plume observations, this simplification should have no effect on the proposed evaluation. The simulation was allowed to proceed for 49 min, until the emissions reached the downwind end of the domain.

Summary of fire and fuel parameters can be found in Table 2. Based on photographs and average measurements of fuel size, composition and type, we determined Anderson's fuel Category 1 (short grass) [19] to be the best fit for L2G ground cover. Actual burn perimeters were used to mask the remaining domain as containing no fuel to prevent spread of the simulated burn outside of the burn lot. We replaced the standard fuel loading and depth associated with Type 1 fuels with average measured values of 0.267 kg m<sup>−</sup><sup>2</sup> and 0.18 m, respectively. Surface dead fuel moisture content was set to 8.46% based on observations. Heat of combustion of dry fuel was adjusted to 1.64 × 10<sup>7</sup> J kg−<sup>1</sup> as per estimates for grasslands provided by Overholt et al. [20].


**Table 2.** Details of fire and ignition parameters in LES setup.

As the central goal of this work is to evaluate the model's ability to capture wildfire smoke plume dynamics, we did not incorporate chemistry coupling into the simulation. Modeled "smoke plume" was represented by two passive tracers released proportionally to the mass and type of fuel burned. The rate of release for each tracer representing CO and CO2 was controlled by assigned emission factors, based on values for grasslands provided by [21] (see namelist.fire\_emissions in the Supplementary Materials).
