Researchers at Mila, the Quebec AI institute, and Natural Resources Canada (NRCan) say they have built an AI model that reproduces Canada’s national wildfire hazard maps in seconds rather than hours. In their tests it ran about 806 times faster than the BurnP3+ simulator on an ordinary CPU and about 3,432 times faster on a data-centre GPU.
The work is set out in a preprint posted to arXiv on 16 October and marked “under review”. It has not been peer reviewed, and every figure in it comes from the authors’ own tests. The model is a copy of an existing simulator, not a new way to forecast fires, and the team says it is built to support expert judgment rather than replace it.
In brief
- What happened: Mila and NRCan’s Canadian Forest Service published an AI “surrogate” that imitates the BurnP3+ fire simulator, with public code and a desktop app.
- Why it matters: Hazard maps that took a day and a half to compute can be redrawn in under a minute, so planners can test more scenarios.
- What’s unknown: Independent review, how well it handles the most extreme hazard, and whether NRCan will use it operationally.
The paper says one model stands in for 30,000 simulated fires
BurnP3+ is the simulator NRCan uses to estimate wildfire hazard. It works by running thousands of simulated fires across a landscape using Monte Carlo methods. The authors write that these workflows “can take days to weeks to process a single regional modelling unit”.
The surrogate is a deep learning model of a type called a U-Net. It takes in fuel, terrain, ignition likelihood and regional weather, and predicts three of BurnP3+’s outputs at once: how likely each spot is to burn, how intense a fire there would be and how fast it would spread. It was trained on BurnP3+ results for 44 regions, with five more for validation and five for testing. The paper says a standard region covers 130,000 square kilometres.
The speed figures come from one test region. According to the paper, BurnP3+ needed 38.9 hours, 32 CPU cores and a peak of 100.8 GB of memory to complete 30,000 iterations there. The surrogate ran the same region end to end in 40.8 seconds on an Nvidia L40S GPU, 93.9 seconds on a MacBook and 173.6 seconds on four CPU cores. The authors put the speed-ups at about 3,432 times, 1,491 times and 806 times respectively. On the CPU it used 13% of BurnP3+’s host memory.
The paper reports close but imperfect agreement
The maps are not identical to the simulator’s. The paper sorts hazard into 13 classes. The surrogate picked exactly the same class as BurnP3+ about 48% of the time, was within one class about 87% of the time and within two classes about 97% of the time. For the four highest classes together, the overlap with BurnP3+ was about 85%.
At the very top, it struggles. An appendix says the highest burn-probability values are “systematically underestimated”, and that the model “underrepresents the long upper tails” of fire intensity and spread rate. For the single most extreme hazard class, the overlap score was zero. The authors link this partly to how the hazard scale is set by a few exceptionally high reference values.
The team also tested “what if” changes, such as replacing burned forest with a different fuel type. The surrogate moved in the same direction as BurnP3+, but agreement was weaker for spread rate, which the paper says “remains a limitation”.
The code and trained models are on GitHub. The README offers a point-and-click desktop app for Windows and Apple-silicon Macs that it says produces maps “in minutes instead of hours of simulation”, exported for standard mapping software. The app is not yet signed with a paid developer certificate. The training data is restricted. The paper says unrestricted release risks raw hazard maps being “easily misinterpreted by commercial entities (e.g., insurance companies)”, and that the data’s custodians may consider requests for research.
BetaKit calls it a replacement; the paper does not
BetaKit reported on 23 October that the model is “meant to replace” BurnP3+. The paper is more cautious. It says the trained surrogate takes the place of the simulator’s many iterations in the workflow, but concludes that it “inherits BurnP3+’s limitations and is intended to complement, not replace, expert judgment in high-stakes decision-making”.
BetaKit quoted Tom Swystun, an NRCan forest systems modelling specialist and one of the paper’s co-authors, from a release: “The emulator generates imperfect approximations, but with high fidelity and at speeds incomparably faster” than the existing tool.
BetaKit also gave the context, citing Public Safety Canada’s October update: more than 5,400 wildfires burned about 4.5 million hectares in Canada this year, and 33,750 people were evacuated, including more than 8,800 from First Nations communities. BetaKit said the season was set to be the fourth-worst on record.
Fast maps turn hazard modelling into a planning tool
The value here is less the headline multiple than what it allows. A map that needs a day and a half of computing gets made rarely and used as a fixed reference. One that takes under a minute can be redrawn whenever a planner wants to ask a question: what happens if burned forest regrows as a different fuel type, or if one region’s weather is swapped for another’s. Those are the kinds of tests the authors ran, and they pitch exactly that use, “interactive scenario exploration”. A free desktop app also lowers the barrier for regional teams without a computing cluster.
A copy of the simulator inherits its blind spots
The catch is in the word surrogate. The model learns to imitate BurnP3+, so it can be no better than BurnP3+, and the paper’s own results show it is somewhat worse where it matters most: the rarest, most intense hazard. Being within one class most of the time is fine for broad planning, but a tool that flattens the extremes could understate risk in exactly the places a community would most want warning. That is why the “complement, not replace” framing in the paper matters more than BetaKit’s shorter version. The results also rest on one benchmark region, so the speed and accuracy picture may shift elsewhere. TSN has made the same point about other fire and weather tools, such as ICEYE and OroraTech’s wildfire satellites and Vaisala’s AI lightning forecasts: claims need published results from real use.
Peer review and NRCan’s own use are the next tests
Two things will show whether this is a research result or a working tool: a reviewed version of the paper, and NRCan publishing hazard maps that were made with it. Neither has happened yet. Other public-benefit tools appear in TSN’s Tech for Good weekly round-up and on the Tech for Good hub.
Sources
- Accelerating Wildfire Hazard Mapping Across Canada with an AI Surrogate Modelling Framework, Abdelwahed et al., arXiv 2610.30170, submitted 16 October 2026 (preprint, under review)
- AISurrogate_wildfireriskmapping code repository and README, Mila Tech Transfer, GitHub
- Mila, federal researchers develop AI tool to better predict wildfire behaviour, BetaKit, 23 October 2026
Related stories
- ICEYE and OroraTech Pair Radar and Thermal Satellites on Wildfires
- Vaisala Says NVIDIA AI Stretches Lightning Warnings to Three Hours
- Tech for Good This Week: AI Diagnosis, Cleaner Grids and Scam Busts
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
