In a paper published in Nature in 2021, DeepMind and the UK Met Office introduced a Deep Generative Model of Radar (DGMR) for precipitation nowcasting - predicting rain over the next zero to 90 minutes, the short window where conventional numerical weather models struggle and where decisions about floods, transport, and energy are most time-sensitive.
DGMR takes about 20 minutes of past radar observations and generates probabilistic forecasts over a region up to roughly 1,536 by 1,280 kilometers. Because it is a generative model, it produces sharp, spatiotemporally consistent rainfall fields rather than the blurry, smeared-out predictions that earlier deep-learning approaches gave when they averaged over uncertainty. In a structured evaluation, more than 50 expert meteorologists at the Met Office ranked DGMR first for accuracy and usefulness in 88 percent of cases against two competitive methods - an unusually direct test of forecast value by the people who actually use forecasts.
The lead author was Suman Ravuri, and the work (arXiv:2104.00954) became an influential demonstration that generative modeling, not just deterministic regression, was the right tool for short-range, high-resolution weather prediction. It sits alongside DeepMind’s later GraphCast and GenCast as part of a broader shift from physics-only forecasting to machine learning.