PyTorchFire: A Differentiable Cellular Automata-Based Wildfire Simulator with GPU Acceleration

Accurate and rapid prediction of wildfire behavior is essential for effective management and mitigation efforts. However, the unpredictable nature of fire spread poses significant challenges to developing reliable simulators. Moreover, these models typically require parameter identification and adjustments based on real-time observations. Current physics-based simulations are mainly CPU-based, which can be computationally intensive and non-differentiable, making direct parameter calibration difficult. While deep learning surrogate models can enhance prediction efficiency, their generalizability to different ecoregions and climate conditions remains limited. This paper introduces PyTorchFire, an open-source Python library built on PyTorch that harnesses GPU acceleration. By utilizing a newly designed differentiable wildfire Cellular Automata (CA) model, the system achieves computational efficiency at the millisecond scale, outperforming conventional CPU-based wildfire simulators when applied to high-resolution, real-world fire scenarios. More importantly, real-time parameter calibration is enabled through gradient descent, allowing simulations to closely align with observed wildfire dynamics both spatially and temporally, thereby improving the realism of the results. By integrating real-world environmental data, PyTorchFire demonstrates enhanced generalizability compared to traditional supervised learning surrogate models. Its ability to simulate and adjust wildfire behavior in real time ensures a high level of accuracy, stability, and efficiency. Numerical tests have been conducted using simplified data from real wildfire events in California, specifically the Pier Fire in 2017 and the Bear Fire in 2020.