Learning fire spread dynamics with physics-constrained machine learning

Recent years have seen a dramatic increase in the extent and intensity of area burned by large wildfires in different parts of the world. However, a complete theoretical understanding of the dynamics driving the spread of fires across a landscape is still elusive, leading to significant uncertainty in predicting the spread of active wildfires as well as mitigating their impact. In this talk, I will present preliminary results from a physics-constrained machine learning (ML) model of fire spread dynamics. Our ML model consists of an Ensemble Kalman Filter (EnKF) data assimilation (DA) algorithm applied to the latent space of a conditional Variational Autoencoder (cVAE). The cVAE is trained on simulation data from an Hamilton-Jacobi PDE advected by a fire spread rate parameterized by the Rothermel and Balbi models, whereas for the DA step we incorporate fire geometries observed at half-day time steps by the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument. We show that including the surface fire spread dynamics simulated based on an idealized representation of the interaction between ignition patterns, fuel characteristics, and meteorological conditions dramatically improves the performance of the DA-cVAE model when compared to a purely data driven approach. Using the ML model as a simulation-based inference technique allows us to robustly quantify the uncertainty in the parameters of a fire spread rate model. Altogether, our model provides quick forecasts of wildland fire spread to facilitate risk minimization, while serving as a diagnostic framework for the limitations of the current theoretical paradigm.