Deep learning models can approximate physics-based wildfire simulators at a fraction of the cost, according to new research.
Researchers trained four architectures - a patch-based U-Net, a transfer-learned ResNet-50, a physics-informed network built around the wind-driven advection equation, and a Swin-Unet transformer - on 10,584 fire-spread simulations run at 2-meter resolution for the Rectoret region in Catalonia, Spain. Testing which terrain and vegetation inputs actually mattered, they found only surface fuel load carried any real predictive signal, with a correlation of 0.27, and adding it cut prediction error by 21 percent. Every other variable tested correlated weakly and mostly duplicated information already captured elsewhere. Saliency, occlusion, and rotation tests then showed the convolutional models lean mainly on distance from the current fire front, while the Swin-Unet transformer weighed fuel and terrain more heavily, a pattern also seen in an unrelated wildfire dataset.
Physics-based fire simulators are trusted because their logic is transparent, but running enough of them to cover a range of possible outcomes gets expensive fast, which caps how many scenarios agencies can model before a fire starts. Cheaper surrogates that hold up would ease that bottleneck, echoing a shift already underway in AI weather forecasting, where models trained on physics simulations now run far faster than the originals. The catch: when the researchers applied the convolutional models, without retraining, to a second region called Pedriza, accuracy dropped by a small but consistent margin, suggesting some of what looked like learned physics was really just fitting to one landscape.
Two test regions in Spain is a thin basis for concluding fuel load matters more than everything else fire touches, and no one has yet checked whether these surrogates hold up against an actual, unpredictable fire rather than a simulation of one.