Wildfire Simulation with Differentiable Randers-Finsler Eikonal Solvers

Barak Gahtan, Jacob Shpund, Alex M. BronsteinarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces a novel differentiable framework for simulating wildfire spread using a Randers-Finsler eikonal equation. This allows for learning complex fire behavior directly from observational data, enabling more accurate and data-driven wildfire prediction models.

What Problem Does This Paper Address?

Accurately simulating wildfire spread is crucial for effective disaster management. Traditional models often rely on simplified physics or require extensive manual parameter tuning. The challenge lies in capturing the complex interplay of environmental factors (terrain, fuel, weather) that influence fire behavior, especially its anisotropic and asymmetric spread. Furthermore, learning these complex relationships from real-world data is difficult because standard wildfire simulation models are not easily differentiable, hindering the use of gradient-based optimization techniques common in machine learning.

Methods

The authors propose a differentiable framework built around a Randers-Finsler eikonal equation. This equation generalizes standard eikonal equations to model both anisotropic propagation (due to varying terrain and fuel properties) and directional drift (due to wind).

Key methodological contributions include:

  • Differentiable Fast Sweeping Solver: They adapt the Fast Sweeping Method (FSM), an efficient algorithm for solving eikonal equations, to be differentiable. This is achieved by using implicit differentiation on the converged solution, avoiding the need to unroll the entire iterative process, which is computationally prohibitive.
  • Handling Non-Differentiability: A core challenge is the discrete stencil selection in FSM, which introduces non-differentiable points. The authors analyze this and show that these points form a "measure-zero" set and that gradients remain stable even when crossing these boundaries.
  • Randers-Finsler Metric: The framework models fire spread using a metric tensor $G$ (for anisotropy) and a drift vector $b$ (for directional bias). These parameters are learned from data.
  • Neural Network Encoder: A neural network takes environmental covariates (terrain, fuel, weather) as input and predicts the spatially varying Randers-Finsler metric parameters ($G$ and $b$).
  • End-to-End Training: The entire system, from covariate encoding to eikonal solving, is trained end-to-end using observed fire perimeters.

Datasets

The paper utilizes the Sim2Real-Fire dataset, a large-scale synthetic dataset containing over one million simulated wildfire scenarios. This dataset includes five aligned data modalities: topography, fuel, vegetation, weather, and satellite-derived fire masks, with spatial resolution of 30 meters and temporal resolution of one hour. The dataset is used for both within-scene and cross-scene generalization experiments.

Key Insights and Ideas

  • Differentiable Eikonal Solvers are Powerful: The ability to backpropagate gradients through a physics-based solver unlocks data-driven learning of complex physical phenomena.
  • Randers-Finsler Captures Realism: This metric formulation is surprisingly effective at capturing anisotropic and asymmetric fire spread, which are crucial for realistic simulations.
  • Implicit Differentiation is Key: For complex iterative solvers like FSM, implicit differentiation is essential for efficient gradient computation, avoiding memory and computational bottlenecks.
  • Data-Driven Models Can Generalize: The framework demonstrates successful generalization both within a landscape (new ignition points) and across different landscapes, suggesting learned covariate-to-metric mappings capture universal physics.
  • Covariate-to-Metric Mapping is the Bottleneck: The study suggests that learning generalizable mappings from environmental covariates to fire spread parameters is the primary challenge, rather than the expressiveness of the physics model itself.

Key Results

  • Forward Solver Accuracy: The fast sweeping solver achieves sub-1% relative error for isotropic, anisotropic, and Randers-Finsler configurations.
  • Gradient Verification: Implicit differentiation matches finite differences with high accuracy away from stencil boundaries.
  • Synthetic Inverse Problems: The framework can recover isotropic metrics to ~6% error and drift fields to <3% error. Multi-source learning improves accuracy by up to 45.6%.
  • Wildfire Simulation:
    • Within-scene generalization: Mean test correlation of 0.824.
  • Cross-scene generalization: Mean test correlation of 0.766, demonstrating geographic transfer.
  • Computational Efficiency: Fast sweeping offers significant speedups (up to 1400x) over Jacobi iteration, and implicit differentiation is ~755x faster than finite differences for gradient computation.

Limitations and Future Work

  • Cartesian Grids: The current implementation is limited to Cartesian grids. Extending to curved surfaces or unstructured meshes is a future direction.
  • Smoothness Approximation: While gradients are stable, the discrete stencil selection introduces non-differentiability. Future work could explore soft-min relaxations or gradient averaging to achieve smoother gradients.
  • Wildfire Model Realism: The learned models produce smooth wavefronts but struggle with fine-grained spatial variations and complex phenomena like spotting or suppression effects. Future work could focus on improving encoder architecture or hybrid physics-learned approaches.
  • Uncertainty Quantification: Bayesian methods for uncertainty quantification are suggested.
  • 3D Extension: Extending the framework to three dimensions is a potential future research avenue.

Summary updated

BibTeX
@misc{gahtan2026wildfire,
  author = {Barak Gahtan and Jacob Shpund and Alex M. Bronstein},
  title = {{Wildfire Simulation with Differentiable Randers-Finsler Eikonal Solvers}},
  year = {2026},
  eprint = {2603.00035},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.00035},
}
Figure