Reconstruct. Predict.
Understand wildfire.
Physics-grounded wildfire intelligence on a living digital twin.
01 — Terrain
Everything starts with the ground
A 5 m national elevation model, resampled and slope-resolved. Not a backdrop: the terrain is what turns a breeze into a run.
02 — Fuel
Fuel is measured, not assumed
Sentinel-2, PNOA LiDAR and the national forest map, classified into Scott & Burgan models for the state of the vegetation in August 2019 — not for today.
03 — Weather
Wind, resolved over the terrain
Reanalysis forcing downscaled by a mass-consistent solver to a 50 m grid, every 30 minutes — including the drainage breeze that turned the fire at night.
04 — Ignition
The hour nobody wrote down
The report records the detection. The ignition is unknown — so it is inferred, with an interval, and labelled as an inference wherever it appears.
05 — Physics
Hundreds of runs, not one
A Latin hypercube over ignition time, fuel moisture, wind bias and spread rate. The spread of the spaghetti is the honest answer.
06 — Observation
Every satellite pass corrects the state
Six overpasses fall inside the window. A detection proves the fire had arrived; its absence proves nothing — and the scoring is asymmetric because of it.
07 — Prediction
The ensemble collapses into bands
What survives the observations is a distribution, shown as ensemble percentiles under stated hypotheses. Never as a certified forecast.
One fire. Thousands of observations. One evolving state.
NOW
The current fire state: arrival time, active front, uncertainty.
+15 / +30
Short-horizon spread under the assimilated wind field.
+1 h
Ensemble bands, widening as the observations run out.
Capabilities
RECONSTRUCT
Rebuild a past fire from the record and the physics, with every layer carrying its provenance class.
OBSERVE
Fuse satellite detections, perimeters and station data into a single evolving state.
PREDICT
Run ensembles, not single deterministic lines, and report the intervals.
RESPOND
Turn the state into arrival estimates at the assets that matter.
Decision support for qualified wildfire professionals. FireTwin does not replace judgement, and every figure it shows carries its interval.
Not another fire map
GIS shows where the fire is.
FireTwin models what the fire is becoming.
FireState
The canonical state — arrival time, front, uncertainty. Static today, assimilating tomorrow.
Model Bus
One normalised contract, many engines. Swap the physics without rewriting the pipeline.
Evidence Graph
Every object knows where it came from, what it derives from and what contradicts it.
Physics meets AI
Physics-grounded AI
Physics
Rothermel surface spread over measured fuel, terrain and a downscaled wind field. Auditable, citable, wrong in ways you can name.
Learning
Statistics where the record supports it — never before three fully observed fires, and never with a fire split across train and test.
We say physics-grounded AI, not AI-powered. The difference is which one is load-bearing.
Case 001
La Granja de San Ildefonso · Segovia · Spain · 2019
Accident-investigation style: only canonical figures, each with its source and its class. Where a number is disputed, all the versions are kept.
| Item | Value | Class |
|---|---|---|
| Burned area (official register) | 385.77 ha | OFFICIAL |
| Copernicus EMSR375, 6 Aug | 370.2 ha | OFFICIAL |
| Own dNBR perimeter | 352.0 ha | DERIVED |
| Detection | 4 Aug, 14:46 | OFFICIAL |
| Ignition time | under inference | RECONSTRUCTED |
| Satellite detections | 96 | OBSERVED |
| Point of origin | as declared in the report | OFFICIAL |
Five different burned-area figures coexist in the record and none is “the right one”; all are kept and labelled. FireTwin reconstructs when, where and how. Never who.
CASE 002
Selection criterion: observational richness — Copernicus EMS activations with multiple delineations.
CASE 003
Machine learning stays gated until three fully observed fires exist.
Who it is for
FIRE AGENCIES
Post-incident reconstruction, training scenarios, and a state that can be interrogated after the fact.
LAND & FORESTRY
Fuel state over time, treatment counterfactuals, exposure of stands and infrastructure.
UTILITIES & INFRASTRUCTURE
Arrival-time estimates at lines, substations and access roads, with intervals attached.
INSURANCE & CLIMATE RISK
Physically grounded event reconstruction for loss attribution and scenario work.