WeatherNext: AI model achieves breakthrough in forecasting cyclones
The WeatherNext AI model from Google DeepMind achieves a breakthrough in cyclone forecasting, with three-day predictions as accurate as two-day forecasts from earlier models—equivalent to a decade of meteorological progress. The model has been open-sourced and was already deployed during the 2025 hurricane season, where it helped the National Hurricane Center predict Hurricane Melissa's rapid intensification.
Weather is one of the most demanding disciplines for current computational models and their hardware. Moreover, weather on the planet changes so rapidly that meteorological models quickly become outdated. AI could help us see things we currently cannot see, and do so even faster than we can today.
How does AI improve upon traditional cyclone forecasting methods?
Conventional approaches required separate models: coarser global models for tracking a cyclone's path and specialized high-resolution local models for predicting intensity. WeatherNext is a single unified AI model that predicts track, intensity, and wind structure together with state-of-the-art accuracy.
Why is improving cyclone prediction accuracy critical?
Tropical cyclones are among Earth's most destructive phenomena, causing over 700,000 deaths and $1.4 trillion in economic losses over the past 50 years. Each additional day of forecast lead time enables communities to prepare, evacuate, and protect lives and infrastructure.
How can WeatherNext work accurately with much coarser input resolution than traditional models?
WeatherNext requires only 28×28 km resolution—100 times coarser than traditional models—yet achieves better accuracy. This counterintuitive finding remains an open research question, and understanding how the model succeeds with reduced spatial detail is an area scientists are actively investigating.
What techniques does WeatherNext use to capture rare but dangerous cyclone behaviors?
The model was trained on 20 terabytes of global weather data and the historical IBTrACS database covering nearly 5,000 storms, learning complex atmospheric patterns. This year it generates 1,000 ensemble members per cyclone to capture rare scenarios like rapid intensification, which occurred during Hurricane Melissa in 2025.
- An Alien Mind: OpenAI's Chief Scientist on Intelligence We Don't Fully Understand — openai.com 69 % match
- November 2025 solar storm caused GPS errors up to 10 meters across continental US — sciencealert.com 69 % match
- LFM2.5-2.6B: small and capable local AI model — liquid.ai 69 % match
- WeatherNext
- Google DeepMind3
- Google Research
- National Hurricane Center
- Hurricane Melissa
- Nature
- UK Met Office
- GitHub15
- Weather Lab