DeepMind's WeatherNext AI Predicts Hurricanes a Day Earlier

Google DeepMind's WeatherNext model predicted Hurricane Melissa would hit Jamaica as a Category 5 storm five days before landfall, with 80% confidence — while the storm was still only a Category 1.

A day of extra warning, on average

In a paper published Thursday in Nature, DeepMind and Google Research report that WeatherNext forecasts a cyclone's track, intensity, and wind structure with what they call state-of-the-art accuracy. On average, the model gives forecasters a full extra day of lead time: its three-day forecasts are as accurate as prior models' two-day forecasts, according to DeepMind's own blog post. DeepMind describes that jump as roughly a decade's worth of meteorological progress compressed into one model.

"Even a few hours can make a difference," Mike Brennan, director of the US National Hurricane Center, told Wired. Organizing evacuations and staging supplies are time-sensitive decisions, and Brennan said pushing forecast accuracy out by a day "is really valuable."

Why intensity, not just track, was the hard part

Earlier AI weather models were already decent at predicting where a storm would travel, but not how strong it would get, according to Kate Musgrave, a paper co-author and tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere. Track prediction relies on global-scale data like cold fronts and prevailing winds; intensity depends on fine-scale local ocean and atmospheric conditions that "we just don't get from these global models," Musgrave told Wired.

WeatherNext closes that gap with a single model trained end-to-end on, per DeepMind's blog, nearly 20 terabytes of global atmospheric data plus the historical IBTrACS database of nearly 5,000 past storms. Notably, the researchers say they don't fully understand why it works: DeepMind scientist Ferran Alet told Wired the model uses lower-resolution atmospheric data than traditional intensity forecasts require, and "when we told the community that our model was only using relatively coarse resolution, they were shocked."

What's shipping, and the caveat attached

DeepMind is open-sourcing the WeatherNext 2 and WeatherNext Cyclones models used during this hurricane season, aiming to let outside researchers build on the work. The model now generates 1,000 possible scenarios per storm, up from 50 last year, to help forecasters account for sudden shifts in a storm's path.

Brennan cautioned against treating any single model as definitive: "There's no guarantee that one model, because it did well last year or really did well for this particular storm, is necessarily going to be the best model for the next season or the next storm." He added that turning a track or intensity forecast into an actual warning still requires human forecasters to translate the numbers into what the impacts will be on the ground.