Wildfire managers make decisions in hours, yet most learned spread models forecast on daily timescales, often targeting total burned area. That mismatch can leave operational teams with predictions that are too coarse for the moment-to-moment choices they face.

FireGraphNet, proposed in a new arXiv paper, instead formulates 12-hour fire growth as a node classification problem on directed H3 graphs. The H3 hexagonal grid and directed edges are designed to represent spatial spread in a way that matches the granularity of real-world response planning.

The paper does not report performance results in the abstract, so it is not yet clear how well FireGraphNet works in practice. But the core idea—matching prediction cadence to decision cadence—addresses a real gap in current wildfire modeling.