Weather: how artificial intelligence will revolutionize forecasts

Will the way of predicting the weather be turned upside down in the near future with ever more precision at stake? The question arises in the face of the rise in power of artificial intelligence (AI). So much so that the subject will be raised on Thursday, November 23 in Toulouse, during a round table in which several experts in the field will participate, the title of which is intended to be explicit, to say the least: “When AI tackles forecasts weather: evolution or revolution?

This week, Google struck a big blow. The Californian company announced that its artificial intelligence department, DeepMind, was able to predict nine days in advance the precise trajectory of Hurricane Lee, which hit the east coasts of Canada and the United States in September . That is to say three days ahead of the European Center for Medium-Range Weather Forecasts (ECMWF), one of the world giants in the sector. Results published in the journal ScienceTuesday November 14.

“The quality of predictions always depends on the quality of the data”

Member of the research team at DeepMind and author of the article, Frenchman Remi Lam is very satisfied with the results obtained without falling into triumphalism. “The quality of predictions always depends on the quality of the data on which we train artificial intelligence,” he explains in an interview with Release. Graphcast (the tool developed by Google for almost two years by around fifteen people, Editor’s note) is trained on the basis of models developed by the ECMWF. “Understand, it will always take material to feed the beast.

Read also :
Storm Ciaran: “We have increasingly efficient digital models”… how Météo-France anticipated the depression

“Today, artificial intelligence cannot replace weather forecasting as it is done today,” says Laure Raynaud, head of AI and forecasting activities at Météo-France. “But of course, it opens up new perspectives.”

Predicting the weather remains a particularly complex exercise which is based on expertise in physical phenomena. “Calculating forecasting means understanding how the atmosphere works, physical laws,” explains Laure Raynaud. “This way of making forecasts is roughly the same since the 1950s, with changes over time. measure.”

Saving time in calculations

And where artificial intelligence could greatly facilitate the task of meteorologists is in the processing of data, the volumes of which are colossal. “With a ‘physical’ model, it will take an hour to calculate a 48-hour forecast while the AI ​​will need a minute,” explains the forecaster.

Huge saving of time therefore, but also of technological resources, since after training, GraphCast is for example capable of running “on a computer graphics card that fits in the hand”, underlines Rémi Lam. Much simpler than using super calculators, like those installed in Toulouse.

“What Google, and others, are offering is a new way of forecasting which is not based on physical knowledge but on learning statistical relationships through an algorithm,” adds Laure Raynaud. “Behind, there is will always need human expertise.

Read also :
Extreme heat: “Extremes still unimaginable in the 20th century”… Météo France is alarmed by the record temperatures the country is experiencing

The perfect prediction, a mirage

Jobs will evolve, but the fear of seeing a drop in staff numbers in Météo-France services nevertheless remains significant. “We are used to technological developments but this must be reflected in the quality of the public service mission that we carry out”, indicates François Giroux, CGT delegate at Météo-France on (symbolic) strike, Monday November 13, for alert about the drop in the number of employees within the entity. “We must not go too fast, too early, in automation,” she says.

As for the fantasy of a weather forecast capable of saying with accuracy, on the scale of the meter and the minute, where and when it will rain: it is not for now. “We can always improve but the uncertainty factors are such that zero uncertainty will never really exist,” concludes Laure Raynaud, who believes that before being able to master AI – “a bit of a black box for the moment” – you had to understand and master the tool.