Google DeepMind WeatherNext Gives Hurricane Forecasts an Extra Day
Google DeepMind’s WeatherNext gave hurricane forecasters an extra day, accurately flagging Hurricane Melissa’s Category 5 Jamaica strike five days ahead.
Summary
A Nature paper published Thursday and covered on August 8, 2026, says WeatherNext, built by Google DeepMind and Google Research, predicts cyclones with unprecedented accuracy, averaging one extra day of lead time. Its three-day forecasts match conventional models at two days, an advance researchers say historically took a decade. In October 2025, models split over whether a Caribbean storm would stay weak and reach Haiti or intensify toward Jamaica. WeatherNext assigned an 80 percent chance of a Category 5 Jamaica landfall five days ahead. Hurricane Melissa then caused catastrophic flooding and landslides, but the forecast enabled earlier warnings and marked the US National Hurricane Center’s first Category 5 prediction while a storm was Category 1.
Lead author Ferran Alet said the team overcame scarce cyclone examples by training on abundant weather and cyclone data. Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere, said hurricanes require global information for tracks but local atmospheric and ocean conditions for intensity, which earlier AI models handled poorly. Retrospective gains persisted in operations despite WeatherNext using unexpectedly coarse atmospheric inputs, and researchers do not know which intensity signal its black box detects. Its probabilistic scenarios address butterfly effects and rose from 50 per storm last year to 1,000 now, beyond the computing capacity of existing numerical models. Mike Brennan, director of the US National Hurricane Center, said the time supports evacuations, supplies and resource staging, but no model is guaranteed to lead next season or storm, and experts must translate forecasts into lethal impacts. Google DeepMind is open-sourcing the hurricane-season WeatherNext models for research, improvement and investigation of cyclone physics.
Positives
- WeatherNext’s three-day cyclone forecasts match conventional models’ two-day accuracy, delivering an advance that historically required a decade.
- Five days before Hurricane Melissa’s landfall, WeatherNext gave an 80 percent probability of a Category 5 strike on Jamaica.
- Hurricane Melissa became the first Category 5 storm predicted by the US National Hurricane Center while it was still Category 1.
- WeatherNext increased its probabilistic output from 50 scenarios per storm last year to 1,000 now.
- Google DeepMind is open-sourcing the hurricane-season WeatherNext models so researchers can improve them and study cyclone physics.
Risks & concerns
- Hurricane Melissa caused catastrophic flooding and landslides across Jamaica despite the earlier warning.
- Scarce cyclone records make extreme-event AI training difficult, forcing researchers to combine limited cyclone examples with broader weather data.
- Researchers cannot explain which signals WeatherNext extracts from coarse atmospheric inputs to predict storm intensity so accurately.
- Mike Brennan warned that WeatherNext’s past performance does not guarantee it will be the best model next season or for the next storm.
- Hurricanes require global track data and fine local intensity data, a multiscale challenge earlier AI models failed to solve adequately.