Wildfires pose an escalating global threat exacerbated by climate change and shifting land-use patterns. Accurate and rapid modeling of fire spread dynamics is critical for early warning systems and emergency response strategies. Among conventional approaches, the semi-empirical Rothermel fire spread model remains one of the most widely employed frameworks; however, it exhibits limitations in capturing the nonlinear nature of wind-topography-fuel interactions. Physics-informed neural networks (PINNs), which integrate governing physical equations into the loss function of artificial neural networks, have recently emerged as a novel paradigm that unifies data-driven learning with physical consistency. This study systematically compares a PINN-based fire spread model against the classical Rothermel model in terms of prediction accuracy and computational speed when modeling multivariate interactions involving wind speed and direction, terrain slope, fuel moisture, and fuel type. A hybrid dataset comprising synthetic FARSITE-generated scenarios and real-world wildfire observations was used to evaluate both models across accuracy metrics (RMSE, MAE, R²), inference time, and scalability. Results demonstrate that the PINN approach achieves 18–35% lower RMSE values compared to the Rothermel model under complex terrain conditions and provides a 40–120× speedup during inference. Nevertheless, training cost and generalizability limitations of the PINN model are discussed, and hybrid approaches are proposed as a promising future research direction.
Kaan Alper (Tue,) studied this question.