Google WeatherNext AI model has achieved a major improvement in cyclone forecasting, providing more than 24 hours of additional predictive lead time for cyclone tracks intensity and wind structure, according to research published in Nature.
The model’s three day forecasts are reportedly as accurate as earlier forecasting systems were for only two days. Researchers say the improvement is roughly equivalent to a decade of progress in cyclone forecasting.
WeatherNext Cyclones was tested against historical storms from 2023 and 2024 and trained using nearly 20 terabytes of atmospheric data and records of almost 5,000 historical storms. It can generate a 15 day forecast in less than a minute and produce 1,000 possible scenarios for each cyclone to capture uncertainty and rare high impact events.
AI could strengthen early warnings for dangerous storms
The model combines forecasts of large scale atmospheric conditions with detailed information about cyclone behaviour. It can predict a storm’s path, strength and wind structure using relatively coarse 28-by-28 kilometre input data which researchers say challenges the assumption that very high spatial resolution is always necessary for accurate intensity forecasting.
The technology was also used during the 2025 hurricane season, when it helped forecast Hurricane Melissa’s rapid intensification and landfall in Jamaica, giving authorities additional time to prepare.
Google has now open sourced WeatherNext Cyclones, WeatherNext 2 and the smaller WeatherNext 2 mini model including their code and model weights. The company says the technology could help researchers, meteorological agencies and other organisations develop better local forecasting and early warning systems.
Despite the advances, researchers stress that AI forecasts are intended to support not replace official meteorological agencies and human forecasters.