Better anticipated heat waves, storms planned to the millimeter, and all without exploding the energy bill: the weather models boosted with artificial intelligence open up new perspectives. At a time when the climate is unleashed, these tools boosted in the code may well become our best allies against disasters. And Google and Microsoft have understood this.
After a first breakthrough in 2023 of a Huawei learning model, the two firms in turn developed AI capable of obtaining, in a few minutes, better forecasts than those produced in a few hours by the powerful traditional calculators of the major international agencies. These performances, experimental and not yet available for the general public or even professionals, however illustrate the rapid progress of research.
Prevent storms
According to Google in December, its Gencast model, trained on historical data, has been capable of providing weather and extreme events over a period of 15 days with unrivaled precision. If Gencast had been operational in 2019, it would, for more than 1,300 climate disasters, exceeded in 97 % of cases the forecasts of the world reference, the European center of meteorological forecasts in the medium term (ECMWF, in English).
Another model called Aurora and developed by a Microsoft laboratory in Amsterdam, still on historical data, has become the first AI model to systematically better predict the trajectory at five days of cyclones than seven state forecast centers, according to the results published in the Nature Reference Reference Review.
For Doksuri in 2023, the most expensive Pacific typhoon to date (more than $ 28 billion in damage), Aurora was able to determine with four days of anticipation that the storm was going to strike the Philippines when the official forecasts of the time saw him head north of Taiwan.
A faster advance than expected
“In the next five to ten years, the Holy Grail will consist in building systems capable of working directly with observations”, satellites or other, “in order to generate high resolution forecasts wherever we wish”, while many countries are devoid of reliable alert system to this day, said Paris Perdikaris, the main author of Aurora, in a video published by nature.
It was predictable that the models of AI would one day compete with the classic physical models, but “we did not think it would happen so early,” said Laure Raynaud, IA researcher at Météo-France, in full development of the AI declination of her arpeggles and aroma models.
A much faster method
The so -called “physical” models, developed for decades, consist in injecting in powerful computers the myriads of observations or weather archives, then applying the laws of physics transformed into mathematical equations, to deduce forecasts. But that requires hours of calculation on overpowered and energy -consuming computers.
All our articles on artificial intelligence
An IA learning model garners these same data, but its neural network self-suits and deduces its forecasts in a “completely statistical” manner, without recalculating everything, explains Laure Raynaud. Thanks to the earnings of speed and quality, “we may be able to calculate our forecasts more often by day”, especially for thunderstorms, devastating and very difficult to predict, explains the researcher. Météo-France targets forecasts with AI on a scale of a few hundred meters.
The ECMWF European Center also develops its IA model, “about 1,000 times less expensive in calculation than the traditional physical model”, said to AFP Florence Rabier, general manager of the center which provides forecasts to 35 European countries.
