Mila builds AI wildfire-mapping tool, reporting up to 3,400-fold speed gains

Canada’s wildfire analysts have a faster way to explore where land could burn. In an October 22 announcement, Mila described an open-source AI model developed with Natural Resources Canada that imitates its wildfire hazard simulator. Mila reports speeds 806 to 3,400 times faster, while keeping the original simulator in place for precise calculations.

A national map that took more than a year

The bottleneck was national-scale computation. Natural Resources Canada, or NRCan, uses BurnP3+, a simulator based on physical equations of fire behavior. To map Canada at very high resolution, it must simulate millions of individual ignitions under different weather conditions. Producing the 2026 national hazard map took NRCan’s teams more than a year; the previous map dated to 2017.

These maps describe how likely an area is to burn, along with potential fire intensity and speed of spread. They draw on weather, terrain and vegetation. Hazard is not the same as risk: risk arises when that potential fire threatens people, property or economic activity. Authorities use hazard maps for long-term prevention and land-use planning.

That planning can include creating firebreaks or planting less combustible vegetation, such as deciduous trees instead of conifers. But the underlying conditions do not stand still. Yan Boulanger, an NRCan forest ecology scientist supervising the project, says changing climate and vegetation demand frequent updates that the existing computing burden makes difficult.

Learning from the simulator’s own maps

The forestry project began in fall 2025. NRCan’s Canadian Forest Service worked with Mila’s applied research team to investigate whether AI could accelerate the mapping process. NRCan supplied simulation maps from the 2026 national mapping effort as training data. Mila’s scientists and forestry specialists used those outputs to develop an emulator—a model that approximates what BurnP3+ produces.

The model also learned spatial patterns associated with burn probability. Boulanger gives the example of a lake: the model can predict a lower probability of burning because of its protective effect. NRCan modeling specialist Tom Swystun describes the outputs as imperfect approximations with high fidelity, rather than exact reproductions of the simulator’s calculations.

Explore quickly, then calculate precisely

The intended workflow puts AI before the detailed simulation, not in its place. Boulanger says analysts could test multiple parameters with the emulator, exploring scenarios that would otherwise require months of BurnP3+ computation. Once they identify a scenario they consider very likely, they can calculate it precisely with BurnP3+. The tool is designed to complement both the simulator and NRCan’s experts.

Mila says the methodology, code and AI models are open-source, with access through the project page linked in its announcement. That makes the work available for further research and potential applications outside Canada or in other areas of environmental risk management. Those extensions are possibilities enabled by the release, not demonstrated outcomes described in the announcement.

The next objective is distribution to the BurnP3+ user community, alongside more testing to understand the model’s capabilities and operational limits. Open-source access therefore does not mark the end of validation. The stated next step is to establish how this faster approximation can be used within the community’s existing analytical work.