A Hybrid Neural Physical Framework for Wildfire Propagation Modeling and Dynamic Evacuation Routing

Forest fires represent an escalating global crisis demanding precise forecasting and immediate operational response. Traditional navigation systems rely on static road closure reports, inadvertently routing civilians into active hazards due to latency. Furthermore, modern predictive frameworks utilizing end-to-end deep learning often succumb to persistence bias, wherein models achieve high statistical accuracy by propagating current fire states forward while ignoring fluid dynamics. This paper introduces an integrated forecasting and routing architecture to address these failures. We engineered a multimodal dataset targeting the topographical variations of Uttarakhand, India, and use it to show empirically that neither a recurrent nor a non-recurrent network learns to advect fire past the persistence baseline. We therefore decouple the predictive workload: a lightweight U-Net supplies a per-cell flammability prior, and an explicit Cellular Automaton physics engine executing within PyTorch performs the temporal spread that the networks could not learn. The engine enforces mathematically rigorous wind advection and convective heating, processing full scale grids in microseconds. Concurrently, we present a dual-layer pathfinding architecture. The system orchestrates OpenRouteService vector routing using spatial polygon clustering to bypass API area limitations. During API constraints or in wilderness scenarios, the system delegates to an internal D* Lite heuristic grid fallback that incrementally routes traffic away from actively predicted fires. This integration provides an interactive simulation environment capable of realistic forecasting and guaranteed safe evacuation.