Between technological promises and operational realities, Laure Raynaud, researcher at the National Center for Research in Meteorology, draws up a nuanced inventory of this digital transformation. It also highlights the current limits of the AI models used by Météo France. For several years, the establishment has been trying to master AI tools to improve the effectiveness of their forecasts. However, we are still far from the crystal ball.
Geographic precision: the weak link in AI
Artificial intelligence models, despite their spectacular advances in many areas, still struggle to compete with traditional methods in meteorology. “ If there is a forecast that is not good, with humans, we can today try to identify the source, whereas with AI models, we do not yet know how they work. to achieve a result » explains Raynaud to Tech&Co .
The granularity of the forecasts constitutes the main obstacle: when Météo France operates to the nearest kilometer, the most sophisticated AI systems display a margin of error of several tens of kilometers. A crippling gap for forecasting extreme eventslike floods. The traceability of decisions also poses a problem: unlike current physical models, AI decision-making processes remain largely opaque.
A gradual and controlled transformation
The public establishment nevertheless remains committed to a measured innovation approach. For five years, Météo France has been developing internal expertise in AI, while preserving its technological independence. This caution is illustrated in particular in its partnership policy: while collaborations with hardware suppliers like NVIDIA are possible, proposals from giants like Google DeepMind (developer of Alphafold) come up against the desire to protect strategic meteorological data.
A position that demonstrates a long-term vision, where AI would complement, rather than replace, human expertise. As Raynaud points out: “ We will always have the human footprint on the forecasts. Today, we have physicists, but tomorrow, it will be data scientists, who will analyze the data “. For this transition to be effective, the researcher insists on the need to continue training employees so that they can master these new generation tools and take full advantage of them.
Météo France is therefore well aware of the potential of artificial intelligence; on the other hand, its practical application to weather forecasting is a more recent development. This transformation will therefore take place smoothly, and the institution does not yet seem ready to completely change its practices overnight, but prefers to adopt a more gradual approach. A completely relevant and responsible strategy to minimize the risk of malfunctions while ensuring better quality of forecasts.
- Météo France is exploring AI to improve its forecasts, but its models still lack the geographic precision to predict extreme events like floods.
- The institution favors a progressive approach and maintains its technological independence, limiting partnerships to protect its data.
- Ultimately, AI will certainly complement, without replacing, human expertise.
