{"resourceId":"mila-nrcan-burnp3-ai-emulator-806-3400x-faster-hazard-mapping-20261022","versions":[{"version":"external-ff98749a045c772fd5fefec6ea4159d001ddc19b613ee538eea108d11d8a8c15","resource":{"id":"mila-nrcan-burnp3-ai-emulator-806-3400x-faster-hazard-mapping-20261022","title":"Mila and Natural Resources Canada: AI emulator of the BurnP3+ wildfire simulator runs 806 to 3,400 times faster, as a screening tool beside it","organization":"Mila (Quebec AI Institute) with Canadian Forest Service, Natural Resources Canada","sector":"Wildfire and emergency management","geography":"Canada","publishedAt":"2026-10-22","publicationDate":"2026-10-22","eventDate":null,"sourceName":"Mila","sourceLabel":"Mila news: AI in the Service of Wildfire Management in Canada","sourceUrl":"https://mila.quebec/en/news/ai-in-the-service-of-wildfire-management-in-canada","evidenceClass":"independent-research","outcomeClass":"emerging","topics":["infrastructure","operating-model","governance-procurement"],"finding":"Mila and NRCan's Canadian Forest Service built an AI model trained on this year's BurnP3+ national simulation maps that emulates the simulator 806 to 3,400 times faster depending on hardware; Mila says a one-year BurnP3+ task would take 2 to 11 hours. NRCan says producing the 2026 national hazard map took over a year, and the previous map dated from 2017. NRCan staff describe the output as imperfect approximations with high fidelity, and say the tool replaces neither BurnP3+ nor analysts: it screens scenarios, and likely ones are recomputed in BurnP3+. Code and models are open source; wider distribution and tests of operational limits are next.","sledRelevance":"U.S. state forestry agencies, county emergency managers and wildfire-risk planners face the same slow, compute-heavy hazard mapping; an open-source emulator suggests faster mitigation and land-use scenario work, though it is tied to Canada's fuel system and simulator.","evidence":"Institutional announcement describing a federal-research collaboration; speed figures are stated by Mila; accuracy detail is in the companion arXiv preprint, not this page.","architectureImplications":"A deep learning surrogate runs in seconds to hours on a GPU, laptop or CPU, beside an authoritative physics-based simulator used for confirmation.","governanceImplications":"Positions AI output as screening, with the physics model and expert analysts retaining the final hazard determination, a useful precedent for agency rules.","securityPrivacyImplications":"Uses fuel, terrain, ignition and weather data rather than personal data; open-source code reduces lock-in but requires the agency to own model validation.","caveats":"Planning-scale hazard mapping, not real-time fire spread prediction; trained only on Canadian BurnP3+ outputs; NRCan says operational limitations are still being tested.","streamIds":["emergency-services"],"roles":{"sales":"For sales, the framing point is speed of planning, not prediction: NRCan needed over a year for its 2026 hazard map, and an AI emulator cuts comparable work to hours. State forestry and county emergency managers who update wildfire hazard maps rarely can relate. A bounded conversation is about scenario screening for mitigation and land-use planning. This source cannot support claims of better fire outcomes, real-time spread forecasting or U.S. accuracy, and the speed figures come from the developers.","engineering":"For pre-sales engineering, the design is a deep learning surrogate trained on outputs of a physics-based simulator, run on modest hardware, with the simulator kept for confirmation. Integration means feeding the same fuel, terrain, ignition and weather layers the simulator uses. U.S. agencies use different fuel models and simulators, so any reuse requires retraining on local simulator output. Proof of value must compare emulator and simulator results on held-out local regions, especially at high hazard levels.","delivery":"For delivery, the work is a validation-first pilot: retrain or test the open-source model against the agency's own simulator runs, then use it to screen mitigation scenarios before confirming them with the physics model. Ownership sits with the state forestry or wildfire-risk modeling lead. Analysts need training to read emulator output as approximate. Acceptance should rest on agreement with the simulator on held-out areas and time saved per scenario. Risks include over-trusting approximations and mismatch with U.S. fuel models."},"roleDetails":{"sales":{"customerProblem":"National or statewide wildfire hazard maps take months to over a year to recompute, so they go stale as climate and vegetation change.","stakeholders":["State forester or wildfire-risk lead","County emergency manager","Land-use planning department","GIS and modeling staff"],"discoveryQuestions":["How long does your last hazard map update take, and when was it last done?","Which simulator and fuel model do you rely on?","How many mitigation scenarios would you test if compute were not the limit?"],"valueHypothesis":"An AI emulator could let planners screen many more mitigation scenarios before confirming a few in the simulator (interpretation).","potentialEngagement":"A bounded validation of an emulator against the agency's own simulator outputs for one region.","unsupportedClaims":["Fewer fires or less burned area","Real-time spread forecasting","Accuracy on U.S. fuel models or simulators"]},"engineering":{"fit":"Fits agencies that already run Monte Carlo wildfire simulators for hazard mapping.","architectureIntegration":"Surrogate model consumes the simulator's fuel, terrain, ignition and weather inputs and outputs hazard layers; simulator retained for confirmation.","prerequisites":["Existing simulator outputs for training","Compatible geospatial input layers","GPU or adequate CPU capacity"],"constraints":["Trained on Canadian BurnP3+ outputs","Different U.S. fuel models","Planning use only, not incident forecasting"],"security":"Non-personal environmental data; agency must control model versions and validation records.","proofOfValue":["Agreement with simulator on held-out local regions","Error at the highest hazard classes","Runtime and cost per scenario"]},"delivery":{"work":"Retrain or test the open-source emulator on local simulator runs and use it to screen mitigation scenarios.","dependencies":["Simulator output archive","Modeling staff time","Agreement on validation thresholds"],"ownership":"State forestry or wildfire-risk modeling lead.","skillsAdoption":"Analysts learn to treat emulator results as approximate screens that need simulator confirmation.","governanceCheckpoints":["Validation sign-off before planning use","Rule that published maps come from the authoritative simulator","Periodic re-validation after fuel updates"],"acceptanceCriteria":["Proposed: agreed correlation with simulator on held-out regions","Proposed: measurable cut in time per scenario screened"],"risks":["Over-trusting approximations","Missed extreme hazard peaks","Mismatch with local fuel models"]}},"retrievedAt":"2026-10-24T21:58:55-05:00","enrichedAt":"2026-10-24T22:02:19-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Shifts analyst time from waiting on compute toward testing more scenarios; analysts still need simulator expertise to confirm results.","procurementImplications":"Open-source release lets agencies evaluate without licensing; any procurement would be for integration, validation and support services.","operatingModelImplications":"Supports a two-tier model: AI screens many scenarios quickly, and the authoritative simulator confirms the few that matter.","sourceVerification":{"responseId":"curl-mila-2026-10-24","openedUrl":"https://mila.quebec/en/news/ai-in-the-service-of-wildfire-management-in-canada","referenceExcerpt":"Importantly, this new tool replaces neither BurnP3+ nor the expertise of NRCan teams.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}