Knowledge-Guided Large Language Models for Enhancing Agent-Based Wildfire Spatial Simulation
Wildfire spatial simulation is essential for understanding the spatial spread pattern of fire across a landscape over time, and Agent-Based Modeling (ABM) is one viable approach. This study evaluates how domain knowledge, integrated via Retrieval-Augmented Generation (RAG), can enhance the spatial reasoning of Large Language Models (LLMs) in wildfire propagation modeling. We propose a wildfire spatial simulation framework where LLMs integrate both simulated and real-world data (e.g., land cover, slope, wind, and human activities) and generate code based on structured prompts. Using Claude 3.7 Sonnet, we compare Claude-only and RAG-enhanced simulation outputs across various scenarios. Results show that while the baseline model often has inaccurate fire spread behavior, RAG-enhanced models significantly improve the simulation results and more similar to the simulation results from the well-known wildfire simulation and modeling software–FARSITE, as well as with real data from the 2024 Corral Fire in California. This work demonstrates the great potential of combining LLMs with domain knowledge to enable wildfire spatial simulation and risk assessment.