PyroPredict 0.1.0

Reference implementation accompanying the paper.

The January 2025 LA fires exposed three separate failures: spread models underestimated the fires by an order of magnitude, evacuation routing ignored spread forecasts entirely, and crew placement ran on intuition. PyroPredict is one pipeline that addresses all three, in pure NumPy on a single CPU.

The number that matters is in the evacuation experiment: routing residents along the shortest path traps 27.2% of them (about 3,627 people per scenario) when roads burn over mid-evacuation. Feeding the same Dijkstra router the spread forecast, so it avoids roads predicted to be cut, traps nobody. The routing decision, not the routing algorithm, is the life-safety variable.

Also included: Rothermel 8-neighbour CA with wind and slope, a hybrid ridge bias correction (honest result: only a marginal IoU gain under heavy spotting, the residual needs spatial structure), and greedy submodular crew placement where 3 forecast-aware crews protect what 10 forecast-blind crews do.

make reproduce runs all three experiments in about 30 seconds. Limitations are listed plainly in the README, including that v0.1 has not yet been validated against the real Palisades/Eaton perimeters.