infrastructure · environment · family: the model was never trained on this
fire models stop where buildings start
No model can predict how wildfire interacts with built infrastructure
Problem statement
Wildfire behavior models (FARSITE, PHOENIX, FlamMap) predict fire spread through vegetation based on fuel type, terrain, weather, and moisture. Structural vulnerability models predict when buildings ignite based on construction materials, defensible space, and ember exposure. These two modeling communities work independently, yet the critical destruction pattern in wildland-urban interface (WUI) fires is the coupled interaction between fire and the built environment: structures ignite from ember showers, become fuel sources that generate their own firebrands, which ignite neighboring structures in cascading structure-to-structure fire spread that can outpace the wildland fire itself. NSF's FIRE program makes "Next Generation Coupled Fire Models" its first focus area (the coupling it names there is between weather and fire), saying that understanding wildland fires and smoke "requires new predictive approaches and models to capture the full spatial and temporal range of unresolved dynamics" and pointing to "the current large uncertainties associated with fire models." Its wildland-urban interface focus area lists, among examples of the research it will support, work to "Test and model behaviors of building materials, structures, and infrastructures under wildfire loads" and to create "algorithms for evaluating cascading failures and community-scale vulnerability." The program counts "timely warnings" and "better management of limited human/equipment resources" among the impacts of improved models.
Why this matters
NSF's 2022 Dear Colleague Letter says wildland fire is becoming "an increasingly prevalent and pressing phenomenon nationally and globally," and the FIRE program says it is "rapidly accelerating in complexity beyond our current understanding." The program describes the WUI as "a critical nexus of potential human-fire interaction, where lives, infrastructure, and communities are often most vulnerable to wildfire." Improving "representation of fire in modeling and forecasting" was one of five critical needs identified at the Wildfire and the Biosphere Innovation Lab that NSF supported in May 2021.
What’s been tried and why it hasn’t worked
Physics-based fire simulators can model ignition and fire behavior for an individual building with high fidelity, but at a computational cost that makes neighborhood- or community-scale simulation impractical for operational decision-making. Empirical structure-vulnerability ratings use scoring rubrics that predict relative vulnerability but not ignition timing or firebrand generation, which limits their use in spread models. Agent-based models that treat burning structures as ember sources are conceptually appealing but lack validated ember generation and transport data for different building types and construction materials. The fundamental barrier is that the relevant physics spans spatial scales from ember ignition to community-wide fire spread and requires coupling atmospheric dynamics, combustion, heat transfer, and structural failure, each from different modeling communities with different tools and validation approaches. The FIRE program's modeling focus area likewise asks for "Innovative models, methods and algorithms for fire phenomena at disparate spatio-temporal scales."
What would unlock progress
A reduced-order model for structure-to-structure fire spread that captures the essential feedback (structure ignition → firebrand generation → downwind structure ignition) without requiring full physics simulation of each building. Validated empirical data on firebrand generation rates, sizes, and transport distances from burning structures of different construction types — data that can only come from controlled full-scale experiments or careful post-fire forensic analysis. Integration of high-resolution building and vegetation data (LiDAR, tax assessor records, aerial imagery) with fire behavior models to enable community-specific vulnerability assessment.
Entry points for student teams
A student team could develop a simplified agent-based model where burning structures generate embers that ignite downwind structures based on distance, wind speed, and construction type, calibrating it against post-fire damage data from a well-documented WUI fire (Camp Fire, Marshall Fire) and comparing predictions to observed destruction patterns. Alternatively, a team could analyze high-resolution damage assessment data from a recent WUI fire to characterize the structure-to-structure propagation network, identifying which building attributes (roofing, siding, vegetation clearance) most strongly predict whether a building served as a firebrand source for its neighbors. Relevant disciplines include fire protection engineering, atmospheric science, data science, and urban planning.
Genes
Structural cousins: stuck the same way, in other fields
Sources
NSF FIRE Program (Fire Science Innovations through Research and Education) PD 25-345Y; NSF CMMI Engineering for Civil Infrastructure (ECI) Program; NSF NHERI; accessed 2026-02-15 go to source ↗
Working notes
Notes made while writing and checking this brief, shown as written. They use working shorthand.
The NSF FIRE program (PD 25-345Y) supports "convergent research, education and networking activities to improve understanding, prediction and resilience to wildland fire and its interactions with communities, infrastructure and the natural environment." FIRE-NET projects "will build new collaborative teams to synthesize aspects of wildland fire science and develop strategies for tackling key gaps." The NSF ECI program supports research on the behavior of physical civil infrastructure "under conditions caused by single or multiple extreme hazard events (extreme weather, windstorms, earthquakes, tsunamis, storm surges, landslides, and fire, including wildland-urban interface fire)." The 2024 Natural Hazards Research Summit theme was "Research-to-Impact: Building Partnerships to Strengthen Community Resilience." Related problem: wildfire-wui-fire-codes-unproven.md addresses the building code gap for WUI; this brief addresses the modeling gap that prevents understanding community-scale fire dynamics.
Reconciliation 2026-09-30: Fact repair after reading the whole write-up against every source it names; a second reader went over the repairs on the same day, re-fetched every source, and tightened four of them. All sources were fetched on 2026-09-30 and read in full: the Dear Colleague Letter on the Source line (https://www.nsf.gov/funding/opportunities/dcl-planning-proposals-catalyze-innovative-inclusive-wildland/nsf22-122, archived by NSF, dated September 19, 2022); the FIRE program page for PD 25-345Y (https://www.nsf.gov/funding/opportunities/fire-fire-science-innovations-through-research-education); the ECI program page, which NSF now serves as the archived PD 19-073Y (https://www.nsf.gov/funding/opportunities/eci-engineering-civil-infrastructure); the NHERI program page (https://www.nsf.gov/funding/opportunities/nheri-natural-hazards-engineering-research-infrastructure); and, for the summit theme, the NHERI network's own news page at DesignSafe (https://www.designsafe-ci.org/community/news/2024/may/nsf-nheri-milestone-the-2024-natural-hazards-research-summit/). These are program descriptions and a funding letter; none of them gives loss figures, housing counts, named fires, named simulators or named rating systems, so the specific figures in the write-up had no source on the page. (1) The problem statement said "No coupled model captures this feedback loop, meaning fire agencies cannot predict" which communities will see structure-to-structure propagation. None of the sources says either thing, so both are removed. In their place the paragraph reports what the FIRE page says: "Next Generation Coupled Fire Models" is its first focus area (the coupling named there is "weather-fire coupling," not fire and buildings); new models are needed "to capture the full spatial and temporal range of unresolved dynamics"; fire models carry "current large uncertainties"; the wildland-urban interface focus area gives, as examples of research it will support, testing and modeling "behaviors of building materials, structures, and infrastructures under wildfire loads" and "algorithms for evaluating cascading failures and community-scale vulnerability"; and "timely warnings" and "better management of limited human/equipment resources" are among the impacts it expects from improved models. (2) Removed because no source contains them: $34.7 billion in insured losses between 2017 and 2024; the list of four fires; 30.8 million homes in the WUI in 1990 and over 44 million in 2020; that post-fire investigations "consistently find" structure-to-structure spread to be the dominant destruction mechanism; that operational predictions "ignore this mechanism entirely"; and that the 2023 Lahaina fire destroyed a town in hours. The paragraph now uses only what the sources say: the Dear Colleague Letter's statement that wildland fire is becoming "an increasingly prevalent and pressing phenomenon nationally and globally," the FIRE page's description of the WUI, and the need to improve "representation of fire in modeling and forecasting," one of five critical needs identified at the Innovation Lab NSF supported in May 2021. (3) Removed from the "tried" section because no source names them: NIST's WUI modeling research, the FDS simulator and "a single building simulation takes hours," and IBHS ratings. The sentences now describe physics-based simulators and empirical vulnerability ratings in general terms; that general description is not in any source either (see the end of this note). (4) "Six orders of magnitude in spatial scale (millimeter-scale ember ignition to kilometer-scale community fire spread)" is not in any source; the sentence now says the physics spans scales from ember ignition to community-wide spread, and a separate sentence quotes the FIRE page's request for "Innovative models, methods and algorithms for fire phenomena at disparate spatio-temporal scales." (5) The Source Notes quoted FIRE-NET as targeting "building new collaborative networks to understand wildland fire science and develop strategies to tackle gaps." Those words are not on the page; the page says FIRE-Network projects "will build new collaborative teams to synthesize aspects of wildland fire science and develop strategies for tackling key gaps," and the quotation now reads that way. (6) The Source Notes quoted the ECI program as supporting research "under conditions caused by extreme hazard events including wildland-urban interface fire." The page's words are "under conditions caused by single or multiple extreme hazard events (extreme weather, windstorms, earthquakes, tsunamis, storm surges, landslides, and fire, including wildland-urban interface fire)," and the quotation now carries the full phrase. Confirmed in the sources and left as written: the FIRE program description quoted in the Source Notes (word for word on the page); PD 25-345Y as the FIRE program number; the 2024 Natural Hazards Research Summit theme "Research-to-Impact: Building Partnerships to Strengthen Community Resilience" (on the NHERI DesignSafe news page, published May 29, 2024; the summit was held May 14 and 15, 2024 at the University of Maryland); the Dear Colleague Letter's title, date and address; and the sibling brief wildfire-wui-fire-codes-unproven.md, which exists. Not in any source and left as general description because it names no figure: the model names FARSITE, PHOENIX and FlamMap; the account of how wildland and structural models divide the work and of the structure-to-structure feedback loop; the general statements about physics-based simulators, empirical vulnerability ratings and agent-based models in the "tried" section; and the "What Would Unlock Progress" section. The entry points name the Camp Fire and Marshall Fire as well-documented WUI fires; no source names them, and they are left as suggestions. The long title's "No Model Can Predict" and the display title are absolutes the sources do not make; they were left unchanged for Beth's decision. Genome tags untouched.