Simulated fires trained a model to forecast fire spread.
Simulated fires on real terrain let a model learn how wind, moisture and ignition change what happens next.
Generate varied fires on real terrain. Learn to forecast their spread from a noisy drone perimeter, then use several possible forecasts to explore evacuation routes and the next places to observe.
Edited demonstration briefThe result
Unseen fires exposed forecast uncertainty and cases where simpler physics performed better.
The model forecast fire spread from imperfect observations. Comparing those forecasts with the simulator exposes uncertainty and where a simpler physics model still performs better.
Synthetic surface fires on real terrain; no operational wildfire or evacuation validation.
Recorded simulation
The recorded fire provides a reference for checking the forecast.
Move through the recorded fire. Compare the physics model with the learned forecast.
Recorded outputs, replayed in your browser. No new simulation is running.
Changing wind and moisture can put a route in the fire’s path.
The world we created
A fire-spread model runs over terrain, fuel maps and roads in the Santa Monica Mountains. Changing wind, moisture and ignition creates synthetic futures, with the simulated arrival time recorded across the landscape.
What this enables
A learned model forecasts fire arrival from a current perimeter and environmental conditions. Planning algorithms use multiple forecasts to explore routes and drone observation points.
Wind, moisture and ignition varied across real terrain.
Vary wind, moisture and ignition on real terrain. Record how the simulated fire spreads.
A learned model forecast spread from incomplete observations.
Train on simulated growth, then test predictions from noisy observations with gaps.
Reference fires revealed errors in unseen forecasts.
Compare unseen predictions with the reference simulation and explore the consequences for planning.
The simulator provides the future an observation cannot reveal. Held-out tests show what the learned model captures and where simpler physics still wins.
Original research artifacts
The original files retain the recorded measurements and model details.