BarakGahtan/differentiable-eikonal-wildfire: Differentiable Eikonal Wildfire

Wildfire Simulation with Differentiable Randers-Finsler Eikonal Solvers Barak Gahtan, Jacob Shpund, Alex M. Bronstein Technion – Israel Institute of Technology, The Hebrew University of Jerusalem, ISTA This repository provides the code accompanying the paper Wildfire Simulation with Differentiable Randers-Finsler Eikonal Solvers. The framework solves and differentiates through the Randers-Finsler eikonal equation $$(\nabla T - \mathbf{b})^\top \mathbf{G}^{-1} (\nabla T - \mathbf{b}) = 1,$$ where $\mathbf{G}(\mathbf{x})$ is a spatially-varying metric tensor encoding anisotropic propagation and $\mathbf{b}(\mathbf{x})$ is a drift vector capturing directional bias. The forward solver uses fast sweeping with eight triangular stencils, converging in 2–3 iterations independent of grid size. The backward pass applies implicit differentiation to the converged solution, computing parameter gradients by solving a sparse lower-triangular adjoint system in a single reverse-time pass. A neural network encoder maps environmental covariates (terrain, fuel, weather) to Randers metric parameters, trained end-to-end through the solver. On the Sim2Real-Fire dataset, the framework achieves mean test correlation of 0.824 for within-scene generalization and 0.766 for cross-scene transfer to unseen geographic regions.