Google DeepMind open-sources WeatherNext cyclone forecasting models

Google DeepMind says its WeatherNext AI system can give cyclone forecasters roughly an extra 24 hours of useful predictive accuracy. The company has published the work in Nature and released WeatherNext 2 and WeatherNext Cyclones as open-source models.

The system predicts a storm’s path, strength and wind structure in one model, addressing a longstanding split between global models that capture atmospheric currents and local systems designed for cyclone intensity. DeepMind says WeatherNext’s three-day forecasts perform comparably to the two-day forecasts of earlier approaches—an improvement it equates to about a decade of progress.

Built for probabilistic forecasting

WeatherNext Cyclones can generate 1,000 possible outcomes for each storm, producing localized probability maps for tropical-storm and hurricane-force winds. Its forecasting process can extend up to 15 days, and the model was trained on nearly 20 terabytes of atmospheric data alongside records of almost 5,000 historical storms from the IBTrACS database.

Tests on storms from 2023 and 2024 reportedly showed more than a day of lead-time improvement across track, intensity and wind-structure predictions. The system also supported the U.S. National Hurricane Center during the 2025 season, including forecasts related to Hurricane Melissa’s rapid intensification and landfall in Jamaica.

For AI builders, the release offers a practical example of combining large-scale environmental data, specialized historical observations and ensemble generation in a single forecasting workflow. Researchers and local weather agencies can now inspect, adapt and potentially deploy the models for disaster preparation, renewable-energy planning and other applications where better uncertainty estimates matter.

Source: Google DeepMind Blog