Large wildfires are notoriously unpredictable and devastating, with the smoke emitted creating unhealthy air across vast regions and warming the globe in in a similar way to volcanic eruptions. Despite their increasing prevalence and intensity, the drivers of these extreme events are often difficult to understand and model. A distinction is often made between wind driven and fuels driven fires, but a spectrum of regimes exist between the two categories, and environmental factors such as bark beetle infestation and human fire suppression have the potential to influence these events in ways not fully understood. Modeling wildfires thus has the potential to alert communities to danger and address fundamental questions about the drivers of large wildfires, although the process can be deeply challenging, as numerical methods, fuel models, resolution, meteorology, turbulence, ignition method, topography, and containment modeling can all play a role in how a fire evolves in simulation and how much smoke is predicted to be lofted into the atmosphere. Because of this complication, different communities often use different models deepening on the question their trying to answer. Wildland firefighters often use simple fire-spread models which can take local measurements of conditions easily and produce accurate results over short periods of time, while air-quality forecasters use even simpler methods, but over longer periods of time, assuming the fire behaves the same as the day before. Researchers using the most detailed and computationally expensive models can account for many time scales and phenomena such as boundary layer turbulence, local meteorology, fire-spread, fire-weather feedback, and fuel dynamics, but the range of interrelated factors involved often leads to complicated sources of error and uncertainty in forecasted variables. With that in mind, this dissertation summarizes the execution and analysis of simulations from Weather Research and Forecasting Model with Fire Code (WRF-Fire) of the 2019 Williams Flats Fire and 2020 Creek Fire to better understand the drivers of large wildfires and the associated uncertainty in their modeled output. Of particular curiosity was the question of whether detailed models such as WRF-Fire can outperform simpler models in the prediction of important variables for air-quality modelers. Improving air-quality forecasting from wildfire events has the potential to improve health outcomes for millions of people and better inform organizations where operations are dependent on outdoor visibility or clean air. This work also addresses the questions of how impactful firefighting containment efforts can be, how bark beetle infestations enhance large wildfires, why pyrocumulonimbus clouds form over certain fires, what is the role of fuel moisture in modulating wildfire dynamics, how canopy burning influences the depth of smoke injection, how fuel loads can influence the depth of the fire front, and what modeling choices lead to the most accurate simulation outputs. The breadth of this knowledge can be used to inform modeling communities in continued development and use of wildfire simulations and to better understand the behavior of wildfire dynamics under changing environmental conditions. I find that WRF-Fire can be a valuable tool for air quality and emergency response communities when careful consideration is paid to the inputs and model configuration. When accounting for containment lines, fuel moisture maps, and high fuel loads, simulations had ~30% less error on daily burned area compared with persistence forecasting over a 5-day forecast for the Williams Flats Fire. To capture accurate smoke injection heights, the inclusion of an explicit canopy model and an increase in fuel depths to a 50 cm average in forested regions was needed. For the 2020 Creek Fire, the fuel depth in forested regions needed to be increased to 1 m to capture ~16 km injection heights and intense pyrocumulonimbus activity, likely due to bark beetle infestation and subsequent modification of the fuel structure. The circularity of the fire front, partially aided by high fuel loads, was also found to play an important role in creating deep plumes as plume buoyancy was better persevered during assent.

Wildfires are increasing in size and frequency in the Western US due to a complex interplay between climate change and landscape-scale fire exclusion practices. The smoke from these fires is degrading air quality across much of the Continental US. Chemical transport models are vital for warning the public about smoky periods, but uncertainties related to fresh smoke plumes can propagate through these models and cause errors in the resulting air quality forecasts.

We address model uncertainty related to smoke plume vertical extent and total emissions. First, we use aircraft observations obtained during the 2019 Western US wildfires (FIREX-AQ) to evaluate and constrain a commonly used smoke plume rise parameterization in two smoke models (WRF-Chem and HRRR-Smoke). Observations show that free tropospheric smoke layers occur in 35% of observed plumes and up to 95% of modeled plumes. False free tropospheric smoke injections were primarily associated with models overestimating fire heat flux by up to a factor of 25. Next, we present data-driven methods for predicting day-to-day changes in smoke emissions. Our top-performing model (random forest) explains 48% of the variance in observed daily emissions and outperforms the current operational assumption that emissions will remain constant over a forecast period (persistence, R2=0.02). This model primarily relies on fire weather data to inform its predictions. Finally, we show preliminary results from WRF-Chem simulations which include random forest-derived emissions and updated heat flux values. We find that in the vicinity of large wildfires in 2020 under less severe fire weather, the random forest-derived emissions can produce better predictions of aerosol optical depth (AOD) and fine particulate matter (PM2.5) than the persistence fire emissions. However, in most cases, persistence and random forest-derived emissions yield very similar AOD and PM2.5 predictions, and that the random forest-derived emissions can both improve and degrade AOD and PM2.5 forecasts. Overall, this work demonstrates the utility of incorporating fire observations to quantify and address uncertainties in our state-of-the-art air quality modeling systems.

Background

Methods

Findings

Conclusion