On December 4, 2024, DeepMind introduced GenCast, a machine-learning system for medium-range weather forecasting that produces probabilistic ensemble forecasts. The work was published in Nature as “Probabilistic weather forecasting with machine learning.” GenCast is a successor to GraphCast, which produced a single deterministic forecast; GenCast instead generates a spread of possible futures.
GenCast is a diffusion model - the same family that powers AI image and video generation - adapted to the sphere of the Earth. A single forecast is an ensemble of 50 or more trajectories, which together estimate the probability of different weather outcomes and so capture the chance of extreme events. The model was trained on four decades of the ECMWF ERA5 reanalysis archive, using data up to 2018 and tested on 2019.
Against the European Centre for Medium-Range Weather Forecasts’ own ensemble system, GenCast produced better forecasts on 97.2 percent of the 1,320 targets evaluated, and on 99.8 percent of targets at lead times beyond 36 hours, out to 15 days ahead. It generates a full 15-day forecast in about eight minutes on a single Google Cloud TPU v5 chip, where the traditional system needs hours on a supercomputer.
GenCast extended the case, begun by GraphCast, that learned models can match or beat decades of hand-built physics simulation for operational forecasting, including for the extreme events that matter most.