Wildfire seasons in Canada are becoming increasingly long, intense, and devastating. The last four years have far exceeded the average burned area of previous decades, with nearly 15 million hectares burned in 2023, twice the size of New Brunswick.
According to Natural Resources Canada (NRCan), the wildfire season could extend by more than a month by 2100 in certain forested regions of Canada. It is therefore urgent to strengthen capacities for preventive measures and land use planning. It is within this context that Canadian Forest Services of NRCan has partnered with Mila to develop a solution capable of accelerating national fire hazard mapping.
The Challenge of Mapping Fire Hazards on a National Scale
Wildfire hazard mapping determines the propensity of an area to burn, as well as the intensity and speed of spread based on weather, topography, and vegetation. Hazard is different from risk, which arises when a hazard threatens people, property, or economic activities.
Using these maps, authorities can plan long-term targeted preventive interventions. These interventions include creating firebreaks to limit the spread of fires or planting less combustible vegetation, such as deciduous trees instead of conifers. Hazard mapping is also essential for assessing the risks surrounding communities, homes, and strategic infrastructure.
NRCan maps hazards using the BurnP3+ simulator, which is powered by the physical equations of fire behavior. However, to cover Canada at extremely high resolution, BurnP3+ must simulate millions of individual fire ignitions under various weather conditions. It took NRCan teams over a year to produce the national hazard map in 2026, with the previous one dating back to 2017.
"Wildfire hazard depends on climate and vegetation, two factors that change over time. Maps must therefore be updated frequently, but with the time and computational weight involved, this is almost unrealistic," says Yan Boulanger, a forest ecology scientist at NRCan and the project supervisor.
Accelerating Mapping with AI
Faced with the growing frequency and intensity of fires, NRCan called upon Mila’s applied research team to evaluate if an AI model can help accelerate the hazard mapping process. To achieve this, the NRCan team provided BurnP3+ simulation maps obtained during this year's national mapping to serve as training data for an AI model.
Motivated by the potential impact, Mila scientists led a cross-disciplinary effort with forestry experts to develop an AI model that emulates BurnP3+ in seconds, running 806 to 3400 times faster depending on the hardware used. For example, a task that normally takes one year using BurnP3+ would only take between 2 and 11 hours at this speed using the AI model.
"The emulator generates imperfect approximations, but with high fidelity and at speeds incomparably faster than BurnP3+. The initial reaction to the model's speed was simply 'wow'," explains Tom Swystun, a modeling specialist at NRCan.
The model also learned to approximate burn probabilities based on the territory's spatial patterns: "The model is capable of predicting that if there is a lake in a given area, there will be less probability of burning due to its protective effect," states Yan Boulanger.
Tom Swystun asserts that the risks of misinterpreting the model are rather low, since it reproduces what BurnP3+ does, but much faster.
Importantly, this new tool replaces neither BurnP3+ nor the expertise of NRCan teams. The model acts as a complementary tool in the analysts' toolkit: “The AI model built by Mila can be used upfront by testing several parameters, which would require several months of computation with BurnP3+. Once a scenario seems very likely to us, we can calculate it precisely with BurnP3+", says Yan Boulanger.
The next objective is to distribute the tool to the BurnP3+ user community and continue conducting several tests to fully realize its capabilities and understand its operational limitations.
A Positive Partnership for Canadian Sovereignty
The partnership between the two institutions is not new: initiated in 2020 around a first project in the mining sector, the relationship progressively expanded from energy to Earth observation, leading up to the launch of this forestry project in the fall of 2025.
The economic and social impacts of this project are major. By reducing computation times and costs, the AI model allows provinces, municipalities, and industry partners to better target preventive land management investments, as well as to better protect populations and infrastructure.
Designed specifically to adapt to the country's forestry information systems, this tool strengthens Canadian sovereignty by offering the federal government an autonomous, long-term planning capacity in the face of increasing climate hazards.
Finally, the methodology, the developed code and AI models are open-source, allowing research to progress on this type of problem and enabling its application to other countries or other fields of environmental risk management.
This project embodies Mila’s conviction that the greatest impact in AI research is achieved when AI scientists work hand-in-hand with domain specialists across Canada to solve our most critical challenges.
For more details, Code, models, and paper can be accessed via the project page here.
