An Exploratory Study of Transformer-Based Generative AI Surrogate Modeling for 3D Wildfire Spread Simulation in Voxelized City

Accurate and efficient simulation of wildfire spread in complex wildland–urban interface (WUI) environments remains a significant challenge due to the high computational cost of physics-based models and the limited fidelity of simplified approaches. This study presents an exploratory generative AI surrogate model for 3D wildfire simulation in voxelized urban environments. The proposed framework leverages a patch-based Transformer architecture with convolutional embedding and autoregressive modeling to learn spatiotemporal fire dynamics from data generated by a physics-based simulation model. Environmental and physical drivers, including fuel properties, wind conditions, and terrain characteristics, are encoded as aligned multi-channel voxel features to provide structured inputs for learning. The model predicts voxel-wise combustion states over time, enabling efficient temporal rollout of fire spread without explicitly simulating the full physical process. A pilot study using over 800 simulated scenarios demonstrates that the proposed approach achieves strong predictive performance, with voxel-wise accuracy ranging from 76% to 100%, while maintaining inference times on the order of 90–220 milliseconds per timestep. The results indicate that the model can effectively capture both spatial propagation and temporal evolution of wildfire dynamics, closely approximating the behavior of the underlying physics-based simulations. This work demonstrates the feasibility of using generative AI as an efficient surrogate for high-resolution 3D wildfire modeling, providing a foundation for scalable, real-time fire simulation and potential integration into urban-scale digital twin systems.