WeatherNext: AI Model Achieves Breakthrough in Cyclone Prediction

WeatherNext is an AI model that predicts the track, intensity, and wind structure of tropical cyclones. A study published in Nature achieved prediction accuracy levels roughly one day ahead of existing models, effectively delivering the accuracy of a 2-day forecast at a 3-day lead time.

Over the past 50 years, tropical cyclones have caused more than 700,000 deaths and $1.4 trillion in economic losses worldwide. Prediction accuracy and warning timing are critical factors in reducing damage.

WeatherNext Cyclones uses global atmospheric conditions from Hurricane Milton in October 2024 as a starting point to iteratively predict global weather patterns and detailed cyclone tracks up to 15 days in advance. It generates 1,000 ensemble scenarios to provide regional probability maps ranging from tropical storms to hurricane-force winds.

Google DeepMind and Google Research developed the model in collaboration with the US National Hurricane Center (NHC), CIRA, the UK Met Office, and various national meteorological agencies. In the case of Hurricane Melissa in 2025, it predicted rapid intensification and landfall in Jamaica, helping to secure time for the NHC to issue advance warnings.

Google has open-sourced WeatherNext 2 and WeatherNext Cyclones, which were used during the hurricane season. This supports researchers and regional meteorological agencies in disaster preparedness, renewable energy expansion, and extreme weather prediction.