Modified Fourier Neural Operators for Direct Wildfire Spread Prediction: A FourCastNet Adaptation
CurrentStructural Adaptation of FourCastNet for Direct Wildfire Spread Modeling This research re-engineers the FourCastNet architecture—originally designed for global weather emulation—to serve as a direct solver for wildfire propagation. Unlike traditional approaches that use FourCastNet solely for atmospheric inputs, this project modifies the internal tokenization and latent space to integrate static and dynamic fire drivers, including fuel moisture, vegetation density (LANDFIRE), and high-resolution topography. By expanding the model's input channels and retraining the Adaptive Fourier Neural Operator (AFNO) on historical fire perimeters, the system learns to predict the evolution of the fire front as a discrete state-space variable. This approach leverages FourCastNet’s characteristic speed to generate subseasonal fire-spread ensembles up to 80,000 times faster than coupled physics models like WRF-SFIRE.
