DeepMind publishes results for WeatherNext hurricane forecast AI

In short: Google DeepMind researchers say their WeatherNext AI model can predict hurricanes earlier and more accurately, and they plan to open-source the version used during hurricane season.

What happened

Researchers from Google DeepMind and Google Research published a paper in Nature showing results for WeatherNext, an AI model (software trained on lots of past data) that forecasts tropical cyclones, including hurricanes.

They say WeatherNext gives forecasters about one extra day of useful warning, on average, compared with existing forecasting systems. Put simply, its three day forecast is about as accurate as older systems’ two day forecast.

The article points to Hurricane Melissa in October 2025 as a real world example. Five days before landfall, WeatherNext predicted with 80 percent confidence that the storm would hit Jamaica as a Category 5 hurricane. Forecasters at the US National Hurricane Center used the model as one input to help issue earlier warnings.

A key claim is that WeatherNext can predict both a storm’s path and its intensity, meaning how strong it will get. Intensity is often harder to forecast. The researchers say the model does this even though it uses lower-resolution weather data (a less detailed picture, like a blurrier map).

DeepMind also says it will open-source the WeatherNext models used during hurricane season, meaning other researchers can download them, test them, and try to improve them.

Why it matters

An extra day can help officials decide on evacuations, move supplies, and prepare emergency teams. But experts caution this is still just one tool, and humans still have to translate a forecast into what it means for real places and real people.

Source: Wired