Coastal Families Could Get About a Day More Hurricane Warning From an AI Forecast

Anyone who has lived on a low stretch of coast knows the arithmetic of a hurricane watch. Board up now or wait. Leave tonight or in the morning. Every one of those choices is easier with more time, and the forecast is what sets the clock.

A study published in Nature on August 6, 2026 reports that an AI forecasting system from Google DeepMind, called WeatherNext Cyclones, reaches the accuracy of today’s leading models for storm track, strength and size about a day or more earlier. Its co-authors include the US National Hurricane Center, Colorado State University and the UK Met Office. During the 2025 Atlantic season, experimental forecasts from it went to the Hurricane Center, and forecasters there began folding its guidance into their forecast process. It is experimental guidance for human forecasters, not an official Hurricane Center model. People still make the call.

Here is the extra day in one number. For a storm’s track five days out, the system gives just over 30 hours more warning at that level of accuracy than the ECMWF ensemble, one of the best tools forecasters use. On the ground, a day like that can be the difference between an orderly evacuation and a frantic one.

Three things to know

• It is more accurate where it counts. Tested on storms from 2023 to 2025, its average five-day track error was 230 km, against 370 km for the ECMWF ensemble. It beat NOAA’s high-resolution hurricane model, HAFS, on intensity and on the size of the storm’s wind field, and it did better at flagging rapid intensification, the sudden strengthening that makes some storms so dangerous. The authors describe the overall gain as “comparable to the progress seen in the last decade of operational development.”
• It learned from real storms. The model trained on about 40 years of global weather analysis from the European Centre for Medium-Range Weather Forecasts (ECMWF) together with an expert-curated archive of nearly 5,000 observed tropical cyclones. Learning both at once let it infer storm strength from coarse global weather data, where earlier AI weather models struggled and tended to underestimate peak winds.
• It is wrong in different places. Its errors line up with those of traditional physics-based models about as much as, and often less than, those models line up with each other. In the authors’ words, it “typically has different failure modes.” Added with an optimized weight to the consensus forecasts the Hurricane Center relies on, it improved track errors in that consensus by an average of 28 percent and intensity errors by an average of 6 percent.

Because it scales cheaply, it can also produce ensembles of up to 1,000 possible futures instead of the usual 50, which the authors say is better at capturing rare events. Anyone can look under the hood: the code and model weights are public on GitHub, and live forecasts are posted on Google’s Weather Lab.

Who feels the difference

Tropical cyclones, the paper notes, have caused more than 700,000 deaths and US$1.4 trillion in losses over the past 50 years. Behind those numbers are households deciding whether to leave and towns deciding when to tell them to. Better forecasts mean better warnings about who should go and when. The paper reports that its forecasts have already been adopted by forecasters at the Hurricane Center and other World Meteorological Organization tropical cyclone centers.

It will not yet say how much rain will fall or how high the water will climb; rainfall and storm surge, which cause much of the damage, are on the authors’ list of next targets. But for a family on that low stretch of coast, a track forecast that firms up a day sooner is a day more to pack the car, check on the neighbors and get out of the storm’s way.