DeepMind's WeatherNext predicts hurricanes a day earlier in real-world test
AI nailed Category 5 Melissa's landfall 5 days ahead—a day's extra warning.
In October 2025, Hurricane Melissa formed over the Caribbean, with models split on its path. Google DeepMind's WeatherNext AI forecast a Category 5 landfall in Jamaica with 80% confidence five days out—and was right. The catastrophic storm triggered flooding and landslides, but early warnings gave communities precious time to prepare. A new paper in Nature shows WeatherNext consistently outperforms traditional models, delivering a full day of extra lead time.
The model's breakthrough is intensity prediction, something previous AI struggled with. By training on global weather data—not just rare cyclone events—WeatherNext captures small-scale atmospheric and ocean signals from coarse-resolution inputs that physics-based models miss. During live tests at the US National Hurricane Center, it predicted Melissa's rapid intensification from Category 1 to Category 5, a first. Researchers admit it's a black box, but that opaque learning is already saving lives and reshaping hurricane forecasting.
- WeatherNext gave 5-day, 80% confidence forecast for Hurricane Melissa's Category 5 landfall in Jamaica
- Predictions are a full day more accurate than existing models: 3-day forecasts match current 2-day accuracy
- Trained on abundant weather data rather than rare cyclone events, enabling accurate intensity forecasting from low-resolution inputs
Why It Matters
Even a few extra hours of hurricane warning can save lives and cut evacuation and resource mobilization costs.