PhysicsGatedUNet: Physics-Guided Feature Gating for Next-Day Wildfire Spread Prediction

The increasing frequency and severity of wildfires have made accurate wildfire spread prediction increasingly critical for emergency response, risk mitigation, and resource allocation. Despite substantial progress in data-driven wildfire modeling, predicting next-day fire propagation remains challenging because fire growth is governed by complex interactions among atmospheric conditions, terrain, and fuel characteristics. Existing deep learning approaches typically incorporate environmental variables by concatenating them with image-derived features, without explicitly modeling how these physical factors influence spatial fire dynamics. Moreover, compared with purely data-driven approaches, physics-guided learning for wildfire spread prediction remains relatively underexplored, presenting an opportunity to integrate domain knowledge with deep learning to improve physically consistent and generalizable predictions. In this work, we introduce PhysicsGatedUNet, a physics-guided encoder-decoder model that embeds wildfire domain knowledge directly into feature propagation. At each decoder scale, a Physics-Guided Feature Gating (PGFG) module conditions the U-Net skip connection on eight physics channels derived from wind, terrain, prior fire extent, and fire-front geometry. This design modulates which encoder features are transmitted to the decoder according to physically relevant spread conditions, rather than relying only on late-stage feature fusion. Two geometry channels—fire perimeter and fire proximity—are derived from the previous-day burn mask. The final prediction is refined by a fuel gate which suppresses predictions in physically implausible no-burn regions. We evaluate PhysicsGatedUNet on the western contiguous United States subset of the Next Day Wildfire Spread (NDWS) dataset. The model achieves an AUC-PR of 0.368, AUC-ROC of 0.879, F1 score of 0.441, precision of 0.384, and recall of 0.517. Compared with FireSenseNet, the current state-of-the-art on this benchmark (AUC-PR = 0.344; F1 = 0.418), PhysicsGatedUNet improves AUC-PR by 7.0\% and F1 by 5.5\% while maintaining comparable model complexity. These results show that injecting physical inductive bias into multiscale feature gating improves next-day fire-spread forecasting and provides a more interpretable mechanism for integrating remote-sensing and environmental drivers. Applying physics conditioning at multiple decoder resolutions allows the network to connect coarse environmental context with fine fire-boundary reconstruction. Although the present evaluation is limited to the western U.S. NDWS subset and a one-day prediction horizon, the framework is modular and can be extended to additional environmental drivers, uncertainty-aware forecasting, and cross-region evaluation. PhysicsGatedUNet therefore offers a practical bridge between data-driven semantic segmentation and domain-informed wildfire modeling, with potential to support faster and more interpretable operational forecasting.