DeepMind Says Its AI Can Predict Hurricanes Earlier Than Everyone Else

Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google DeepMind's WeatherNext model predicted Hurricane Melissa would hit Jamaica as a Category 5 hurricane five days before landfall with 80% confidence — the first time the National Hurricane Center predicted a Cat 5 when the storm was only at Category 1.
- The Nature paper published Thursday shows WeatherNext predictions three days out are as accurate as traditional models' two-day-out predictions, giving forecasters roughly one extra day of lead time — an improvement researchers say historically would take a decade of work.
- WeatherNext generates 1,000 potential scenarios per storm (up from 50 last year) to capture butterfly-effect deviations, and was trained on both general weather data and cyclone data to overcome the scarcity of extreme-event training data.
- The AI model uses lower-resolution atmospheric data than traditional models require yet still produces highly accurate intensity predictions, surprising both the DeepMind team and the broader atmospheric science community — researchers called it a "black box" suggesting something is happening in cyclone dynamics not previously understood.
- National Hurricane Center director Mike Brennan called the extra lead time "really valuable" for evacuations and supply staging, but cautioned that no single model is guaranteed to perform best for the next storm and emphasized human experts are still critical for translating forecasts into impact warnings.
- Google DeepMind is open-sourcing the WeatherNext models so researchers can use and improve on them, with lead author Ferran Alet hoping the open release could help uncover fresh insights into how cyclones work.
Why it matters: NHC director Mike Brennan called the extra day of lead time "really valuable" for evacuation and supply-staging decisions. Hurricane Melissa was the first time the NHC predicted a Category 5 storm when it was only Category 1 — and Google DeepMind is now open-sourcing WeatherNext so the broader atmospheric research community can interrogate and improve it.
Ask SkimNews


