Computational Performance of Physics-Informed Neural Networks Versus the Rothermel Model in Wind-Topography-Fuel Interaction-Based Wildfire Spread Prediction
Compares the accuracy of fire spread prediction models in wildfire scenarios, suggesting improvements for future modeling.
Key Points
The study aims to compare the prediction accuracy and computational performance of physics-informed neural networks and the Rothermel model in wildfire spread dynamics.
Systematic comparison of PINN-based model and Rothermel model.
Evaluation using a hybrid dataset of synthetic and real-world wildfire data.
Assessment based on accuracy metrics including RMSE, MAE, and R².
Measurement of inference time and scalability under various conditions.
PINN approach shows 18–35% lower RMSE values compared to the Rothermel model.
Achieves a 40–120× speedup in inference time.
Highlights limitations in training cost and generalizability of the PINN model.