Open-source wildfire spread model

A wildfire spread model that runs on your graphics card

FlameLength simulates how a wildfire spreads across real terrain, hour by hour, from its fuel, the slope and the weather. On a recent graphics card, 48 hours of a large fire takes a few seconds, in your browser or on an NVIDIA GPU.

Needs WebGPU: a recent Chrome or Edge, or Safari 26 or later.

The demo running: a 3D map of the Santa Monica Mountains with the Palisades fire on it five hours after its start, its fire line bright around a dark burn scar that reaches the houses below.
The Palisades fire’s first 48 hours as the demo runs them, three times as fast. Run the Palisades, Eaton and Woolsey fires from where they really started, change the wind and how dry the fuel is, or tap the map and start a fire of your own. Then hold a run against the mapped fire, or run it 8 times with the wind a little different each time.
48 hof fire per run, in about 2 seconds on an RTX 3090
30 mgrid cells, over real fuel, slope and hourly weather
23past fires scored against their mapped perimeters
3Los Angeles fires to replay in the demo, or start your own

What it shows

How far a fire could spread if nobody fights it

  • It isn't a forecast. Don't use it to decide whether to evacuate.
  • There's no firefighting in it. Fires that crews caught early run far too large.
  • Towns don't burn in it. A fire stops at the edge of a town unless embers carry it across.
  • It knows Southern California. Every fire it has been tested on is there, run on weather known after the fact.

How well it does

Scored against real fires, failures included

I score a run by how well the area it burned matches the fire's mapped perimeters, using the Sørensen overlap, where 1 is a perfect match. ELMFIRE, an established fire spread model, is the comparison.

Mean overlap with the mapped perimetersFlameLengthELMFIRE

19 Southern California fires each with its own windstorm held out of the fitting

FlameLength 0.448
ELMFIRE 0.393

The blind exam, 4 fires kept out of everything until the model was fixed, then opened once

FlameLength 0.137
ELMFIRE 0.097
Show as a table
TestFlameLengthELMFIRE
19 Southern California fires, each with its own windstorm held out of the fitting0.4480.393
The blind exam: 4 fires kept out of everything until the model was fixed, then opened once0.1370.097

It failed the blind exam. To pass, it had to beat ELMFIRE on three of the four fires, and it won two. All four fires were caught within a day, and every model, mine included, ran them far too large. The 19-fire score is also short of the 0.46 I set as the bar before the first score.

How it works

Published fire science, public data, two engines

Published physics

Rothermel's surface spread, crown fire after Van Wagner and Cruz, embers after Sardoy and Schroeder, and a level-set fire front. A handful of its numbers are fitted to 83 California fires.

Public data

LANDFIRE fuels and canopy, NOAA's hourly HRRR weather, gridMET fuel moisture, and fire perimeters mapped by NIFC and NASA. Every source is credited, with its terms.

Two engines

CUDA on NVIDIA cards, where it is fitted and scored against past fires, and a WebGPU port of the same core that runs in the demo, with the fitted settings or the published ones.

Open source

Everything behind the scores, in the open

The engine, the browser port, the scorer, the fitted numbers and every test record are on GitHub.

Code and fitted numbers: MITRecords and memos: CC BY 4.0