Artificial intelligence writes weather history: Google DeepMind revolutionizes with GenCast

With the introduction of GenCast, an AI-based weather forecasting model, Google DeepMind is setting new standards in weather forecasting. GenCast significantly surpasses traditional methods by providing forecasts for up to 15 days and more accurately anticipating extreme weather events. The tool takes an innovative approach by evaluating various scenarios to accurately assess trends from wind energy generation to tropical cyclone movements. This probabilistic technique marks a turning point in weather forecasting with modern machine learning models. According to Ilan Price, a research specialist at Google DeepMind, GenCast could enrich practical weather forecasting systems and help decision-makers better prepare for upcoming weather events. Particularly noteworthy is GenCast's use of 'ensemble' forecasts, which represent different possible outcomes—a technique previously primarily used in conventional forecasting systems. The model is trained with four decades of data from the European Centre for Medium-Range Weather Forecasts (ECMWF). Impressively, GenCast performed better than the ECMWF in 97.2 percent of cases with its 15-day forecasts. This performance builds on DeepMind's past successes, particularly the GraphCast model, which had already recorded successes when it was released last year. GraphCast stood out with better forecasts on 90 percent of the metrics for three to ten days. AI-driven forecasting models, which typically operate faster and more efficiently than traditional methods, demonstrate their potential here: GenCast produces forecasts in just eight minutes—compared to several hours with classical methods. However, researchers see further optimization possibilities, especially in forecasting severe storms and enhancing data resolution according to the latest ECMWF upgrades. The significance of GenCast is also underscored by the words of the ECMWF, which described the development as a significant milestone. The ECMWF has already integrated elements of GenCast into its own AI system. Even though GenCast delivers groundbreaking results, the discussion about the optimal mix of AI and traditional physics remains intriguing. Google is banking on a hybrid technique, as demonstrated in July with the introduction of the NeuralGCM model, which combines machine learning and physical modeling. Steven Ramsdale from the British Met Office emphasizes the potential of these exciting developments. Nevertheless, he sees the greatest value in a hybrid approach that unites humans, physics, and AI.