Wildfire Prediction and Simulation

Wildfires are becoming more severe and frequent, driven in part by anthropogenic climate change. These devastating events not only result in tragic loss of life and property but also impose enormous economic costs. Accurately predicting the spread of wildfires remains a critical challenge, as current fire behavior models are largely empirical. These models, often calibrated for specific conditions, lack the generality needed to perform reliably when actual conditions deviate from their assumptions. The problem is especially pronounced in regions of complex terrain, where local flow patterns interact with the fire, altering its behavior in unpredictable and dangerous ways.

To address these challenges, our research at the FxLab focuses on developing a next-generation, open-source fire/atmosphere solver, Swirl-LM/Swirl-FIRE, capable of simulating wildfire spread in nearly real-time (Wang et al., 2023). Collaborating with Google Research, we leverage Tensor Processing Units (TPUs)–highly specialized processors designed for machine learning and scientific computing. Using TensorFlow and just-in-time compilation, this advanced solver captures the critical physics of coupled fire/atmosphere interactions, enabling large-scale simulations that could transform our ability to predict and manage wildfires.