A Fire Can Create Its Own Thunderstorm: Researchers Now Able to Predict It One Day Ahead

A cloud capable of generating its own lightning formed above Gironde on July 24, 2026, a first in France. In the face of this phenomenon that classical meteorology cannot forecast, researchers are developing algorithms able to detect it up to 24 hours before it appears.

In Gironde, a wildfire has forged its own self-sustaining storm

Originating from Saumos, in Médoc, on July 22, 2026, the fire ravaging the Arcachon basin has crossed a rare threshold: it now creates its own conditions for propagation. “A phenomenon never observed in France,” summarized Marc Vermeulen, head of the Gironde departmental fire and rescue service.

Concretely, the heat released by the flames lifts a convective column so powerful that it draws in surrounding cool air to feed the combustion, like a giant self-feeding chimney. The fire then escapes the sole logic of the external wind.

Why simulation software cannot foresee this type of blaze

Convective fire has the peculiarity of generating its own weather, notably its own wind. Its trajectories become erratic: it no longer tends to follow the direction of the dominant flow reported by Météo-France, which complicates the work of ground teams.

Beyond a certain altitude, the water vapor carried by the plume condenses into a thunderstorm cloud. This cloud produces its own lightning and triggers ember attacks far upstream from the flame front, opening up new, unpredictable hotspots.

The simulation software used by emergency services rests on a simple assumption: a fire driven by external wind and terrain. This feedback of the fire on itself escapes them almost entirely, invisible as long as you only look at synoptic wind maps.

Pyrocast, the program that detects the precursors of a fire thunderstorm

The Pyrocast pipeline, developed notably by researchers at the University of Cambridge, relies on a database gathering satellite imagery and environmental data for more than 148 pyroCb events that occurred in North America, Australia, and Russia between 2018 and 2022.

Random forests, convolutional neural networks, and autoencoder-pretrained models were compared to predict, six hours in advance, the formation of a pyroCb. The best among them reached a ROC-AUC of 0.90, a score signaling highly reliable discrimination.

A 2026 version expands the inventory to 214 events in the United States and Canada between 2013 and 2020, tested on 91 cases from 2021, with a targeted 24-hour forecast window. In the United States, NOAA foresees smoke within a comparable horizon, and an Indian team has adapted Pyrocast to the Meteosat satellite.

A useful alert, but still far from a crystal ball for firefighters

A score of 0.90 does not turn an algorithm into a crystal ball. These models are trained on a few hundred global events, a tiny sample compared with the diversity of terrains and vegetation. A Landes pine forest fire does not have the same thermal signature as an Australian wildfire or a Canadian boreal fire.

A twelve- to twenty-four-hour alert would allow repositioning aerial assets before the tipping point, to prepare preventive evacuations and to inform air traffic. It remains to be seen whether French rescue services, still ill-equipped to integrate these models, will be able to turn this statistical lead into operational decision-making.