A Real-Time 2D Spatiotemporal Fire Spread Forecasting Artificial Intelligence Agent

During a tunnel fire, the foremost priority is the safe evacuation of passengers. Extreme temperatures and toxic combustion products can quickly lead to mass casualties, so evacuation support systems require fast forecasts of how hazardous conditions will evolve in space and time. This study investigates whether sparse sensor measurements can be used to reconstruct future tunnel-wide fire conditions on two-dimensional sections that are directly relevant to structural assessment and human exposure. To this end, we develop 2D ST-FAM, a data-driven forecasting model that maps time-resolved measurements from 75 tunnel sensors to future temperature, soot, and carbon monoxide (CO) fields derived from 108 computational fluid dynamics (CFD) fire simulations. The study is organized around three questions: whether the model can accurately reconstruct future tunnel fields from sparse measurements, whether this performance is maintained on both the vertical center plane and the horizontal breathing plane, and which physical quantities remain most challenging to predict. Results show high structural agreement with the CFD reference fields over the full 1800 s prediction horizon, with average structural similarity index (SSIM) values of 0.964 for temperature, 0.984 for CO, and 0.937 for soot. These findings indicate that sparse-sensor forecasting is feasible for tunnel-scale temperature and toxic-gas field prediction, while soot prediction remains comparatively more difficult because of its sharper spatial structures.