Google DeepMind's WeatherNext Cyclones improves cyclone forecasts by 24 hours

Google DeepMind and several weather agencies have developed WeatherNext Cyclones, a model that predicts the path, intensity, and wind structure of tropical cyclones. It was trained on nearly 20TB of global atmospheric data and about 5,000 historical storms. Using functional generative networks, it can generate a 15-day forecast in under a minute on TPUs, and now produces up to 1,000 predictions at a time.

The company says the model extends prediction lead time by 24 hours, from 2 days to 3 days, which is like a decade of progress in meteorology. During the 2025 hurricane season, it correctly predicted Hurricane Melissa would strengthen and hit Jamaica, giving the U.S. National Hurricane Center time to warn local teams.

Google has open-sourced WeatherNext 2, WeatherNext Cyclones, and WeatherNext 2-mini, along with code and model weights for researchers worldwide.