“Extreme precision”: Google claims, Meteo France would be much worse than its tool for knowing what the weather will be

News JVTech “Extreme precision”: Google claims, Meteo France would be much worse than its tool for knowing what the weather will be

A probabilistic model to understand uncertainty

Google has launched a new offensive in the field of weather forecasting

with GenCast, an artificial intelligence model presented as revolutionary. According to the company, this probabilistic tool surpasses in precision and speed the current reference system, the Ensemble Prediction System (ENS) of the European Center for Medium-Range Weather Forecasts (ECMWF), and by extension, Weather forecasts France, which relies in part on this system. A statement which, if verified, could revolutionize our way of anticipating the vagaries of the sky.

GenCast, published in the scientific journal Nature

stands out for its probabilistic approach. Unlike deterministic models that provide a single forecast, GenCast generates a set of more than 50 possible weather scenarios, each associated with a probability. This approach makes it possible to understand the uncertainty inherent in weather forecasts, particularly in the long term. The user thus has a more complete vision of upcoming weather conditions and can make more informed decisions. The example of Typhoon Hagibis in 2019 illustrates GenCast’s ability to refine its predictions as the deadline approaches.

Generative AI for the weather

GenCast’s performance is based on the use of a diffusion model, a generative artificial intelligence technique. Adapted to the spherical geometry of the Earth, GenCast was trained on four decades of historical weather data from the ECMWF ERA5 archive. This massive dataset allowed the model to learn global weather patterns at a resolution of 0.25°. Google claims that GenCast is more accurate than ENS on 97.2% of the 1,320 combinations of variables and forecast times tested, and 99.8% beyond 36 hours.

Beyond the increased precision, Google highlights the speed of execution of GenCast. A single 15-day forecast can be generated in just 8 minutes on a Google Cloud TPU v5. In addition, all predictions can be calculated in parallel, which represents a considerable time saving compared to traditional models. This energy efficiency is a major asset.

Better anticipation of extreme events

The potential impact of GenCast is particularly significant in forecasting extreme weather events. Faced with heatwaves, cold waves, violent winds and cyclones, more precise and faster forecasts make it possible to anticipate risks and implement more effective prevention measures. Google points out that GenCast outperformed ENS in predicting these phenomena.

Improved weather forecasting also has significant implications for renewable energy. By optimizing the forecast of wind energy production, GenCast can contribute to better management of electricity networks and promote the integration of renewable energy sources.

Although Google presents GenCast as a major breakthrough, it is important to qualify these claims. The study published in Nature focuses on a comparison with the ENS and uses historical data. The performance of GenCast in real conditions remains to be demonstratedand a direct comparison with Météo France forecasts would be necessary.

Despite these reservations, GenCast represents an important step in the evolution of weather forecasting. Google is paving the way for more accurate, faster and more accessible forecasts. The opening of the source code and the upcoming release of predictions will allow the scientific community to explore the potential of GenCast. The future of weather forecasting looks decidedly technological, and collaboration between digital giants and national weather agencies will be essential.