Weather forecasting: stunning progress thanks to artificial intelligence

They are called FourCastNet, Pangu, GraphCast, GenCast, FuXi or AIFS. They are the ‘ChatGPT’ of weather forecasting and perhaps soon, under other names, of climate models. In less than two years, these models, which anticipate changes in the atmosphere in ten days, have turned an existing field on its head.

The artificial intelligence (AI) revolution started in this sector in February 2022 with a preprint from the American graphics card company Nvidia, which introduced FourCastNet, a competitor to traditional models. Huawei then described Pangu in November 2022, before Google Deepmind released GraphCast on Christmas Eve, which will result in a publication in Sciencewhich marks the beginning of a new era. “We took this article in the face, recalls Marc Pontaud, director of higher education and research at Météo-France. We, like the other important centers in this area, had in mind that it would be possible, but we did not expect it to happen so quickly. »

The specialists are in a hurry. “I had to adapt quickly”testifies Mariana Clare, from the European Center for Medium-Range Weather Forecasts (ECMWF), who works on the uncertainties of different models. “Our recruitment profiles have evolved, with the search for new skills”adds his colleague Zied Ben Bouallegue, specialist in prediction verification.

More “dice draws”

In April 2023, several European countries (notably France, Norway, Sweden and Switzerland) decided to join forces to build a model on the scale of Western Europe. Six months later, ECMWF adapted GraphCast for its own model, AIFS. Finally, Météo-France predicts that a prototype at a scale of less than two kilometers will be developed before the end of 2024, and will be validated in 2025.

The benefits of these new technologies are impressive. While a simulation on a dedicated supercomputer takes two to three hours, on a computer with a single graphics card it takes only a minute to get a very close result. However, these savings in time and energy will not reduce the carbon footprint of this activity, as specialists will take advantage of these gains to make more predictions. Because these phenomena are inherently unstable, we can better estimate the uncertainty the more “dice” there are. Instead of the current fifty predictions every six hours, there could be a thousand each time.