Weather: this French AI does better than Google with much fewer resources.

Weather forecasting using artificial intelligence has become, in three years, a battleground among giants. Google DeepMind launched GraphCast in 2023, then GenCast in 2025. The European Centre for Medium-Range Weather Forecasts (CEPMMT, based in Bonn) has been running its IFS ENS model for decades. To compete with these giants, one typically needs massive GPU farms and research budgets in the millions. A team at INRIA in Paris has just demonstrated the opposite.

How to do better than Google with a fraction of its resources

The trick lies in a simple, well-executed idea. Instead of training a generative model directly on raw atmospheric data (which costs a fortune in computation), the team took a two-step approach. The first step involves training a classical deterministic model, ArchesWeather, which predicts the average atmospheric state 24 hours later. The second step uses a technique called "flow matching" (a modern variant of diffusion models, those same models used to generate images) to model what the deterministic model does not capture: the uncertainty component, alternative scenarios, and trajectories the atmosphere could have taken.

On the WeatherBench benchmark, a test that measures prediction accuracy against real data from a year (specifically the year 2020), ArchesWeatherGen outperforms IFS ENS by an average of 5.3% on the CRPS score, the reference metric for probabilistic forecasts, across all key variables up to 10-day forecasts. When compared to Google's GenCast trained on an 111 km grid, the French model performs slightly worse in the short term (1 to 3 days) but gains an advantage from day 4 onwards. And facing the high-definition version of GenCast, which splits the atmosphere into 28 km grid cells and costs twenty times more to train, ArchesWeatherGen converges to comparable performance for 9 and 10-day forecasts.

An accessible model that opens the door to university labs

The total training budget for ArchesWeatherGen represents 45 days of GPU V100 computation, which is about 23 days on more recent A100 cards. To give an idea, Google's GenCast requires over 1,000 V100 days. The training dataset splits the globe into 167 km grid cells and weighs only a terabyte, compared to 36 times more for the fine 28 km grid used by the competition. The entire code, pre-trained models, and data pipeline are published on GitHub under an open license.

The team behind these works is led by Guillaume Couairon (who has been at Google DeepMind Paris) and Claire Monteleoni, within the framework of the "Choose France" AI chair. The computation was carried out on French infrastructures from GENCI and IDRIS, the CNRS computing center. To put it clearly: it's French public research, funded by public funds, that produces a competitive model compared to those of the best-equipped laboratories in the world.

ArchesWeatherGen will not replace Meteo-France's operational models tomorrow, which run at much finer resolutions. But it proves that a university lab can now produce world-class probabilistic forecasts with an accessible budget.

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : Science Advances/INRIA