“Probabilistic weather forecasting with machine learning,” by Ilan Price and colleagues at Google DeepMind, was published in Nature in December 2024. It introduced GenCast, which differs from earlier machine-learning weather models in that it produces an ensemble of possible forecasts rather than a single best guess, capturing the uncertainty inherent in weather prediction.
GenCast is a diffusion model: it generates each forecast trajectory by sampling one time step at a time through an iterative denoising process. The paper reports that it produces a set of stochastic 15-day global forecasts at 12-hour steps and 0.25-degree resolution, covering more than 80 variables, in about eight minutes. GenCast had greater skill than ENS, the ensemble forecast of ECMWF, on 97.2 percent of the 1,320 targets evaluated, and it better predicted extreme weather, tropical cyclone tracks and wind-power production.
Ensemble forecasting matters because the most useful weather information is often the probability of a damaging event, not a single deterministic line. By beating the leading ensemble system while running far faster, GenCast extended the machine-learning advantage from average-case accuracy to the tail risks that drive decisions about evacuations, energy markets and disaster preparation.