DeepMind's WeatherNext matches a decade of cyclone forecasting progress

Forecasting a hurricane has long meant running two different kinds of model: a coarse global model to guess where the storm will go, and a separate high-resolution model to estimate how strong it will get. Google DeepMind says its WeatherNext model, described in a paper published in Nature on August 6, does both at once, along with the storm's wind structure, from one system. The headline result is a full extra day of lead time: a three-day forecast is now about as accurate as what earlier methods could manage for two days. DeepMind frames that jump as roughly a decade of normal meteorological progress.

The numbers behind it are specific. Three-day track error is around 100 kilometers, and intensity error is about 11 knots. The model was trained on roughly 20 terabytes of atmospheric data and close to 5,000 historical storms from the IBTrACS database, and it generates a 1,000-member ensemble in under a minute on a single TPU. It runs at a 28-by-28 kilometer resolution, about 100 times coarser than traditional physics models, which is part of why it is so fast. Last year it called Hurricane Melissa's rapid intensification and Jamaica landfall well ahead of the storm.

Speed matters here as much as accuracy, because forecasters can run many scenarios cheaply and see the range of plausible outcomes rather than a single track.

Why it matters

If you work in emergency management or disaster response, an extra day of reliable warning is time to evacuate, move equipment, and stage supplies. The cheap ensembles also let forecasters show a spread of outcomes instead of one line on a map, which changes how confidently a warning can be issued.