GenCast: Diffusion-based Ensemble Weather Forecasting New

GenCast (Google DeepMind, 2023; Nature 2024) is a generative machine learning model for medium-range global weather forecasting. It frames forecasting as a conditional diffusion process on a 0.25-degree latitude-longitude grid and produces a full 15-day probabilistic ensemble (50+ members) in roughly 8 minutes on a single Cloud TPU, compared with hours on supercomputers for physics-based ensembles such as ECMWF's ENS.

The model is trained on ERA5 reanalysis and initialized from conventional analysis. In the GenCast paper (arXiv:2312.15796), it was shown to outperform the best operational physics ensemble on the great majority of standard verification targets, including extreme-event risk.

Licensing note: the code (released alongside GraphCast in the same repository, google-deepmind/graphcast) is under the Apache License 2.0. The trained model weights are made available under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license - commercial use is forbidden.

Why it matters for physics AI: GenCast shows that generative ML can deliver skillful probabilistic forecasts and establishes a template (diffusion on the sphere) that complements or exceeds traditional NWP, marking an early step toward operational ML ensembles.