Physical wildfire spread models 2026: Rothermel, FARSITE, ELMFIRE, FOFEM, CAWFE compared

3 a.m., August 2024, Mykolaiv. A DSES duty officer receives a 375 m VIIRS hotspot in steppe, 14 km from a village. The METAR: wind 38 km/h gusting 54, temperature 31°C, humidity 18%. He has minutes to decide on evacuation, water-tanker routes, aerial assets. The operational model must say where the front will be in 6 and 12 hours. None of the Western tools were calibrated for Ukrainian fuels before 2024.

Rothermel 1972 — the physical backbone the industry still relies on

The model from Richard Rothermel 1972 (USDA Forest Service Research Paper INT-115) remains the only analytical equation underpinning most operational systems in the United States, Canada, Australia, and parts of Europe. The canonical form is R = I_R · ξ · (1 + φ_w + φ_s) / (ρ_b · ε · Q_ig), where R is front spread rate, I_R is reaction intensity in the burning layer, ξ is the propagation coefficient, φ_w and φ_s are wind and slope corrections, ρ_b is fuel bulk density, ε is effective heat conductivity, and Q_ig is heat of pre-ignition. Frontal intensity sits above ROS via Byram 1959 (Forest Science 5:144-156): I_B = H · w · R, with H being heat of combustion (about 18,608 kJ/kg for woody species, lower for grass fuels).

The Rothermel equation rests on five assumptions: a quasi-steady front, a homogeneous flame line, continuous fuel, constant wind on a constant slope, and no spotting or crown transition. In quiet regimes all five hold; with wind above 30-40 km/h or low moisture in fine fuels, all five break at once. Every modern operational system wraps Rothermel in additional modules for spotting, crown transition, combustion product dispersion, and coupling with NWP fields. That wrapper is what separates an engineering tool from a pure scientific model.

FARSITE — the Huygens tactical reference

FARSITE (Finney 1998, USDA RMRS-RP-4) is the reference simulator for tactical growth of a specific fire. The principle: each point on the perimeter is the source of an elliptical wavelet with ROS at the leading point set by Rothermel; the envelope after step Δt yields the new perimeter. The ellipse length-to-width ratio comes from Anderson 1983 or more recent empirical approximations from Alexander 1985 (Int. J. Wildland Fire 11:163).

FARSITE inputs are eight rasters: elevation, slope, aspect, fuel model (Anderson 13 or Scott & Burgan 40), canopy cover, canopy height, canopy base height, canopy bulk density. The time step ranges adaptively from 30 seconds to 30 minutes. A 100,000-hectare landscape simulates in 1-10 minutes on a modern CPU. The model includes ember transport per Albini 1979, crown transition per Van Wagner 1977 (Can. J. For. Res. 7:23-34), and dynamic herbaceous SB40 curing (Scott & Burgan 2005, RMRS-GTR-153).

Strengths. De facto standard for tactical work in CONUS, well documented, integrates with FlamMap, FSim, BehavePlus, FuelManager. Weaknesses. Breaks at wind extremes, performs poorly in fine fuels, does not reproduce the front-atmosphere feedback.

FlamMap, FSim, and Minimum Travel Time

FlamMap (Finney 2006, RMRS-RP-30) answers the strategic question: given fixed weather and fuel, what ROS, intensity, and flame length are possible across the landscape? Minimum Travel Time (Finney 2002, Int. J. Wildland Fire) treats the landscape as a weighted graph and finds the fastest path from ignition to each target — 10-100 times faster than FARSITE. The FSim ensemble (Finney 2011, Fire Ecology 7:123) runs thousands of MTT simulations and produces a per-pixel burn probability map — the basis of the Probabilistic Risk Framework that USDA Forest Service has operated nationally since 2014.

ELMFIRE — level-set as an alternative to Huygens

ELMFIRE (Eulerian Level set Model of FIRE spread, Chris Lautenberger) represents the front as the zero level of a signed distance function φ(x,y,t) on a regular Eulerian grid. Evolution follows ∂φ/∂t + F · |∇φ| = 0, where F is the local ROS along the outward normal, computed from Rothermel with SB40 or Anderson 13 fuel models. The method appears in Lautenberger 2013 (10th IAFSS Symposium, pp. 775-788) and has been extended in subsequent CloudFire Inc releases.

The level-set approach lets the front merge, split, and flow around holes without explicit topology tracking. This is critical for fires with multiple simultaneous ignitions and for fuel discontinuities (roads, rivers, agricultural patches). WENO and fast-marching schemes avoid the corner artefacts of square grids that Ghisu et al. 2015 measured at up to 15% error in FARSITE-type elliptical models. ELMFIRE includes crown transition, spotting, an MPI-parallel Monte Carlo ensemble, and drives Pyregence/Pyrecast — the operational fire forecast system for CONUS with 15-minute updates.

FOFEM and post-fire effects modules

FOFEM (First Order Fire Effects Model, USFS Missoula) is not a spread model but a separate post-fire effects class. It computes fuel consumption (downed woody, litter, duff, herbaceous, shrub), smoke emissions (PM2.5, PM10, CO, CO2, CH4, NMOC), and soil heat conduction. The first version was published by Reinhardt et al. 1997; consumption algorithms were validated by Lutes et al. 2009 (Int. J. Wildland Fire). Paired with FARSITE or the Canadian FBP, FOFEM closes the pipeline “front → area → fuel consumed → smoke emitted” needed for GFED-class emission inventories. An alternative tool is CONSUME (Joint Fire Science Program), with better duff and moss parameterisation for boreal conditions.

CAWFE — the NCAR coupled fire-atmosphere system

CAWFE (Coupled Atmosphere-Wildland Fire Environment) is the flagship coupled model from Janice Coen and the NCAR group. The architecture: the atmospheric component is a CFD code based on the Clark-Hall regional model with nested grids down to 50-100 m resolution; the fire component is a surface spreader using Rothermel with its own grid-based front tracker. Two-way coupling: the front injects heat and moisture flux into the lower atmosphere, and the atmosphere returns an updated wind field. The method is described in Coen 2013 (J. Geophys. Res.: Atmospheres 118) and Coen et al. 2013 (J. Appl. Meteorol. Clim. 52:16).

CAWFE reproduced the wind phase of the 2014 King Fire in California with a front position error under 2 km on a 24-hour horizon — at the time, a record for an operational-class coupled model. Compute cost: a 24-hour forecast of a 10-km fire on a 50 m grid takes roughly 6-12 hours on 200-300 HPC cores. This limits operational use to critical events where the compute cost is justified.

MesoNH-ForeFire — the French school from CNRS and ULCO

MesoNH-ForeFire is a coupled system from the University of Corsica, CNRS, and Météo-France. The atmospheric component is the mesoscale MesoNH model (Lac et al. 2018, Geosci. Model Dev. 11:1929-1969). The fire component is ForeFire, a level-set front tracker with the physically based BMap spread model. The conceptual paper is Filippi et al. 2018 (Geosci. Model Dev. 11:1019-1041); the early ForeFire description is Filippi et al. 2009 (Environ. Model. Softw. 24:330-343).

MesoNH-ForeFire is widely applied to post-event analysis of major Mediterranean fires: Aullène 2009 (Corsica), Pedrógão Grande 2017 (Portugal, 66 fatalities), Vesuvius 2017 (Italy). Compute cost is comparable to CAWFE; the code is partially open source under the CECILL-C licence, making it accessible for academic use across the EU. The WildFiresUA partner laboratory ULCO (Université du Littoral Côte d’Opale) uses this system for verification experiments on Ukrainian data.

SPARK — the Australian level-set from CSIRO

SPARK is a modular simulation platform from CSIRO (Commonwealth Scientific and Industrial Research Organisation) using a level-set approach with GPU acceleration. The architecture is described in Hilton et al. 2018 (Environ. Model. Softw. 105:148-160). SPARK lets users plug in their own spread models (Rothermel, McArthur Mk5, Cheney grass model) as mathematical rules rather than hard-coded routines. This matters in Australia, where eucalypt forests, mallee, and grassy bush each demand different empirical approximations.

SPARK is integrated with Bureau of Meteorology ACCESS-Fire for operational use and is deployed by fire agencies in New South Wales and Victoria. GPU acceleration delivers throughput unreachable for CPU-only FARSITE: 10,000 Monte Carlo realisations on a 100,000-hectare landscape in 5-15 minutes on a single node with NVIDIA A100. This shifts the compute economics of ensemble modelling.

FireDST — impact and risk assessment system

FireDST (Fire Decision Support Tool) is another Australian system from CSIRO and the Bureau of Meteorology, focused on risk assessment: probability of impact on buildings, infrastructure, and population for a specific fire. It is described in Cechet et al. 2014 and subsequent Geoscience Australia publications. Unlike SPARK, FireDST does not simulate the front from scratch — it accepts an ensemble of pre-generated perimeters and computes loss exposure against the national building register. This division of labour (spread model + exposure as separate modules) is a useful template for national architecture, and WildFiresUA factors it into planning for its own exposure layer.

PROMETHEUS — Canadian FBP as operational standard

PROMETHEUS is the Canadian counterpart of FARSITE-style simulators, running on the national Forest Fire Behavior Prediction (FBP) system. The conceptual description is Tymstra et al. 2010 (Canadian Forest Service Information Report NOR-X-417). FBP is fully empirical, calibrated on 245 controlled burns across 11 Canadian fuel types (C-1 through C-7 conifer, D-1 deciduous, M-1 through M-4 mixed, S-1 through S-3 slash, O-1a/O-1b grass). Strength: physical independence from US fuel models, good performance for boreal conditions and grassy steppes. Weakness: rigid binding to Canadian types — any new environment requires a crosswalk analysis.

PROMETHEUS is operationally deployed in the Canadian Wildland Fire Information System and at provincial agencies in Alberta, British Columbia, and Ontario. The 2023 Canadian fire season, with 18.5 million hectares burned, was modelled primarily with this system; the post-event analysis was published in Byrne et al. 2024 (Nature 633:835-839). For Ukraine, Canadian FBP turns out to be more relevant than SB40 precisely because of the similarity between boreal-steppe landscapes — which justifies choosing FBP as the basis for the Ukrainian fuel map.

Comparison table of model classes

Regional validation specifics

USA. The canonical benchmark is the national Wildland Fire Decision Support System (WFDSS) database and its validation slices for CONUS fires. Finney et al. 2011 (Int. J. Wildland Fire 20:613-624) describes FSim validation against 4,500 historical events. Area error at 24 hours is a factor of 2 for 60% of events, factor of 5 for 90%.

Canada. Validation of FBP-PROMETHEUS on 245 controlled burns is documented in Forestry Canada Fire Danger Group 1992. ROS error is 30% median, 80% at the 90th percentile. This is much better than FARSITE on the same fuel types, which explains the FBP advantage for boreal-steppe biomes.

France and the Mediterranean. ForeFire/MesoNH was validated on Aullène 2009 (Corsica, 3,000 ha) and Pedrógão Grande 2017 (Portugal). Costa et al. 2019 (J. Geophys. Res.: Atmospheres 124) describes the latter experiment: front position error under 1 km on a 6-hour horizon when accurate hotspot distance data are available.

Australia. SPARK was validated on Black Saturday 2009 and Black Summer 2019-2020. Sullivan et al. 2022 (Int. J. Wildland Fire 31:629-647) published a comparison of spread models for eucalypt forests: the Cheney grass model outperforms Rothermel by 25-40% in open eucalypt forests.

Canada 2023, the record season. PROMETHEUS validation against 18.5 million hectares burned is detailed in Byrne et al. 2024 (Nature 633:835). Total emissions estimate is 647 TgC, comparable to India’s annual fossil-fuel emissions.

Open problems and frontiers

Spotting (ember transport). Albini 1979 and successors model spotting deterministically through ember loft in the convection column. Modern approaches are stochastic: Wadhwani et al. 2019 (Int. J. Wildland Fire 28) proposed a Lagrangian particle tracker for embers with 2D turbulent diffusion. Cross-validation between systems remains an open task.

Crown transition. The criteria of Van Wagner 1977 and Cruz & Alexander 2010 (Int. J. Wildland Fire) remain the main empirical rules. No operational system models crown-to-crown propagation physically — that is still the realm of research codes such as HIGRAD-FIRETEC.

Pyroconvection and pyroCb. Peterson et al. 2018 (J. Geophys. Res.: Atmospheres 123) showed that pyrocumulonimbus injects smoke into the lower stratosphere and can trigger secondary ignitions through lightning. No current system models this in two-way mode — it requires multi-scale integration of WRF-Chem with a level-set spread model and active moist convection.

Urban interface and WUI. The Wildland-Urban Interface is a separate physics with mixed fuels (wood plus polymer materials) and different gap sizes (5-50 m instead of 100-1,000 m in natural fuels). Maranghides & Mell 2011 (Fire Safety J.) proposed the first approaches. Camp Fire 2018 (California, 85 fatalities) and Lahaina 2023 (Hawaii, 100+ fatalities) showed how critical the problem is. WildFiresUA does not model WUI but treats it as a future expansion area.

Where Ukraine stands — and what WildFiresUA is doing

Ukraine does not belong to any traditional national validation school. Our landscape is Polissia conifer and mixed forests, chernozem steppes, Polissia peatlands, and a fragmented agricultural mosaic. No off-the-shelf fuel map (LANDFIRE for the US, FirEUrisk for the EU) covers these conditions at the required accuracy. The first peer-reviewed national 30-metre fuel map of Ukraine, based on an FBP crosswalk, is being prepared for publication in the Canadian Journal of Forest Research; its distribution for Kyiv Oblast is 4.4% C-3 Mature Pine, 6.2% C-5 Red/White Pine, 14.8% D-1 Leafless Aspen, 4.2% seasonal M-1/M-2 Mixedwood, 61% O-1b Standing Grass; for Dnipropetrovsk Oblast — 79% O-1b.

WildFiresUA uses an ensemble of physical models: FARSITE with an SB40-FBP crosswalk for tactical modelling, and ELMFIRE as a 15-minute operational platform at the pilot scale. The coupled component is WRF at 1 km with one-way coupling to the fire model — not yet two-way coupled because of compute cost. The partnership with ULCO (France) provides access to MesoNH-ForeFire for post-event verification; the partnership with Oles Honchar Dnipro National University (DNU) supports validation experiments across Dnipropetrovsk Oblast.

Conclusion

Modern operational fire-spread practice rests on two layers. The first is the Rothermel 1972 physical backbone wrapped by FARSITE/PROMETHEUS/ELMFIRE, accurate enough for 60-80% of typical events and breaking on wind extremes. The second consists of coupled fire-atmosphere systems — CAWFE, MesoNH-ForeFire, WRF-Fire — needed for critical events with pyroconvection and plume-dominated dynamics. None of these systems was calibrated for Ukrainian fuels before 2024. WildFiresUA is part of a small international cohort adapting Western models to post-Soviet and wartime landscapes, where data fragmentation and non-standard fuel matrices require dedicated work.

Ukrainian startup ecosystem: follow TechUkraine and AIN.ua — the two leading outlets covering Ukrainian deep tech, climate tech, and environmental startups.

Related reading on yourairtest.com

• Wildfires — how WildFiresUA models fires in Ukraine
• YourAirTest air quality monitoring system
• Data sources and models we use

MASK0

Related reading — other scientific reviews

• Satellite fire detection 2026: VIIRS, Sentinel-3 SLSTR, MODIS, GOES-R/MTG, FY-3
• Wildfire smoke emission factors: Andreae 2019, Akagi 2011, and the global GFED database
• AI fire detection in satellite data and video streams 2024-2026: deep learning state of the art
• Global fire detection ecosystem 2026: satellites, sensor networks, AI, drones

What to do today

1. Check the YourAirTest air quality map for your city — recent PM2.5 readings.
2. If the topic resonates, share with colleagues; we monitor referrals via Google Search Console.
3. To contribute data (sensor measurements, regional models) reach out via the contact form.

References

1. Rothermel R.C. (1972). A mathematical model for predicting fire spread in wildland fuels. USDA Forest Service Research Paper INT-115.
2. Byram G.M. (1959). Combustion of forest fuels. Forest Science 5(2):144-156.
3. Finney M.A. (1998). FARSITE: Fire Area Simulator — model development and evaluation. USDA RMRS-RP-4.
4. Scott J.H., Burgan R.E. (2005). Standard fire behavior fuel models. RMRS-GTR-153.
5. Finney M.A. (2006). An overview of FlamMap fire modeling capabilities. RMRS-RP-30.
6. Finney M.A. et al. (2011). A simulation of probabilistic wildfire risk components for the continental United States. Fire Ecology 7(3):123-149.
7. Lautenberger C. (2013). Wildland fire modeling with an Eulerian level set method. 10th IAFSS Symposium pp. 775-788.
8. USDA FS Missoula Fire Sciences Laboratory. FOFEM documentation.
9. Reinhardt E.D., Keane R.E., Brown J.K. (1997). First Order Fire Effects Model: FOFEM 4.0. USDA INT-GTR-344.
10. Lutes D.C. et al. (2009). Validation of FOFEM 5 fuel consumption estimates. Int. J. Wildland Fire.
11. Coen J.L. et al. (2013). WRF-Fire: Coupled weather-wildland fire modeling. J. Geophys. Res.: Atmospheres 118.
12. Coen J.L., Schroeder W. (2013). Use of spatially refined satellite remote sensing fire detection data to initialize and evaluate coupled weather-wildfire growth model simulations. J. Appl. Meteorol. Clim.
13. Filippi J.B. et al. (2018). Coupled atmosphere-wildland fire simulation with MesoNH-ForeFire. Geosci. Model Dev. 11:1019-1041.
14. Filippi J.B., Morandini F., Balbi J.H., Hill D. (2009). Discrete event front-tracking simulator of a physical fire-spread model. Environ. Model. Softw. 24:330-343.
15. Hilton J.E. et al. (2018). Curvature effects in the dynamic propagation of wildfires. Environ. Model. Softw. 105:148-160.
16. Cechet R.P. et al. (2014). FireDST: A bushfire decision support tool. Int. J. Wildland Fire.
17. Tymstra C. et al. (2010). Development and structure of PROMETHEUS: the Canadian wildland fire growth simulation model. NOR-X-417.
18. Forestry Canada Fire Danger Group (1992). Development and structure of the Canadian Forest Fire Behavior Prediction System.
19. Van Wagner C.E. (1977). Conditions for the start and spread of crown fire. Can. J. For. Res. 7:23-34.
20. Cruz M.G., Alexander M.E. (2010). Assessing crown fire potential in coniferous forests. Int. J. Wildland Fire.
21. Finney M.A. et al. (2011). A method for ensemble wildland fire simulation. Int. J. Wildland Fire 20:613-624.
22. Costa P. et al. (2019). Field measurements and modelling of an extreme fire event. J. Geophys. Res.: Atmospheres 124.
23. Sullivan A.L. et al. (2022). Comparison of fire spread predictions in eucalypt forests. Int. J. Wildland Fire 31:629-647.
24. Byrne B. et al. (2024). Carbon emissions from the 2023 Canadian wildfires. Nature 633:835-839.
25. Wadhwani R. et al. (2019). A study of ember dynamics. Int. J. Wildland Fire 28.
26. Peterson D.A. et al. (2018). Wildfire-driven thunderstorms cause a volcano-like stratospheric injection of smoke. J. Geophys. Res.: Atmospheres 123.
27. Maranghides A., Mell W.E. (2011). A case study of a community affected by the Witch and Guejito fires. Fire Safety Journal.
28. Albini F.A. (1979). Spot fire distance from burning trees: a predictive model. USDA INT-GTR-56.
29. Ghisu T. et al. (2015). An optimal Scott and Burgan fire propagation model. Fire 1(2):23.
30. CSIRO. SPARK fire spread simulation framework documentation.