Background
Climate change is driving an increase in wildfire extent and severity. Recently, the 2018 Camp Fire burned 150,000+ acres, destroyed nearly 19,000 buildings (costing $16.5 billion), killed 85 people, and contributed to hundreds of excess deaths from poor air quality. Consequently, the U.S. spends $1+ billion annually to combat wildfires. Increases in fire activity, wildland urban interfaces, suppression costs, and controlled burn hazards suggest that current fire management practices are unsustainable and a fundamental shift from reactive to proactive responses is needed. More advanced modeling and decision-making tools are key to directing this paradigm shift. However, wildfires entail highly complex characteristics such as intricate combustion dynamics and external space-time uncertainties (e.g., weather, vegetation composition). Research is underway for high-fidelity wildfire models that accurately capture behavior and widespread impacts (e.g., air pollution). Challenges include modeling large terrains, procuring adequate data, and embedding these models in decision-making strategies.
Proposed Solution
My group will engineer scalable surrogate wildfire models that guide data collection (monitoring) and readily embed into decision-making tools. Inspired by the success of neural operator models in emulating chaotic Navier-Stokes flow systems, we plan to develop physics-informed neural operator models to create high-fidelity models that rapidly predict wildfire behavior. With these we can conduct scalable uncertainty propagation to guide data collection (monitoring). Moreover, the computational efficiency of these surrogates facilitates the integration of high-fidelity models (e.g., fire and weather simulators) into decision-making tools, promoting better informed fire-management strategies. For instance, these can consider the impact of smoke from controlled burns on surrounding population areas. This work will be done in close collaboration with experts in wildfire research.