Predicting storms and heat waves advances thanks to artificial intelligence

 Improving the accuracy of heatwave and storm forecasts, and reducing the significant energy consumption required, is a goal for various meteorological agencies. They are now relying on rapid advances in artificial intelligence models that enable them to prepare for disasters exacerbated by climate change.

After achieving initial progress in 2023 with a machine learning model developed by Huawei, Google and Microsoft have developed AI tools capable of producing better forecasts in minutes than those produced by conventional computers used by major international agencies, which take several hours.

This experimental performance, not yet available to the public or even professionals, is an indicator of the rapid progress in research.

Google announced last December that its GenCast model, trained on historical data, demonstrated unparalleled accuracy in predicting weather and extreme weather events over a 15-day period. If GenCast had been operational in 2019, it would have outperformed the global reference, the European Centre for Medium-Range Weather Forecasts (ECMWF), in 97% of cases for more than 1,300 climate disasters.

Another model, Aurora, created by a Microsoft lab in Amsterdam using historical data, became the first AI model to predict hurricane paths five days ahead of time, better than seven government forecasters, according to results published this week in the scientific journal Nature.

For Typhoon Duksuri in 2023, the costliest tropical storm in the Pacific to date (more than $28 billion in damage), Aurora was able to predict the storm four days before it reached the Philippines, while official forecasts at the time indicated it was heading north of Taiwan.

"In the next five to ten years, the ultimate goal will be to build systems capable of working directly with observations," whether satellite or otherwise, "to make highly accurate forecasts wherever we need them," says Paris Perdicaris, the lead inventor of Aurora, in a video published by the journal Nature. Many countries currently lack reliable warning systems.

AI models were expected to one day rival classical models, but "no one thought it would happen so quickly," Laure Raynaud, an AI researcher at the French meteorological agency Météo-France, told AFP, as they developed two AI-based versions of the Arpège and Arom models.

These so-called "physical" models, which have been developed over decades, work by feeding vast amounts of observational data or weather archives into powerful computers and then applying the laws of physics, transformed into mathematical equations, to produce forecasts.

Its drawbacks are that it requires hours of computation on energy-intensive computers.

An AI-based learning model collects the same data, but its neural network feeds itself and generates forecasts in a "purely statistical" way, without recalculating everything, according to Laure Raynaud.

"Thanks to gains in speed and quality, we may be able to calculate our forecasts more frequently, every day," the researcher says, especially for storms that are considered devastating and difficult to predict. Meteo France is working to provide AI-powered forecasts at a scale of a few hundred meters.

The European Centre for Medium-Range Weather Forecasts is developing its own AI model, which is "about 1,000 times less computationally expensive than a traditional model," Florence Rabbet, director general of the centre, which provides forecasts for 35 European countries, told AFP.

This AI model currently produces forecasts on a scale of about 30 square kilometers, which is certainly less detailed than Aurora's (about 10 square kilometers), but its first version is already operational and has been used since February by local meteorologists responsible for preparing alerts for the population.

These forecasts won't disappear anytime soon, according to Laure Raynaud, who says, "We will always need meteorologists to evaluate the data."

Florence Rabbet says, "When it comes to protecting people and property, I don't think we can do without human expertise.