WeatherNext: DeepMind's AI Gives You an Extra Day to Prepare for Cyclones

Google DeepMind has announced a breakthrough in forecasting tropical cyclones — the hurricanes and typhoons that over the past 50 years have claimed more than 700,000 lives and caused $1.4 trillion in damage. In a paper published in Nature, the researchers showed that their WeatherNext model has achieved record accuracy in predicting a cyclone's track, its intensity and its wind structure.

The key achievement is measured in time: on average, the model gives forecasters an extra full day. A three-day WeatherNext forecast turns out to be as accurate as earlier models were only two days out. By the authors' estimate, that is equivalent to roughly a decade of progress in meteorology.

One model instead of two

Cyclone forecasting used to require a trade-off. A storm's track depends on global atmospheric currents, which were better captured by coarse global models. Its intensity, meanwhile, is governed by subtle thermodynamic processes near the cyclone's core, which called for specialized high-resolution models. WeatherNext combines both tasks in a single AI model.

It was trained simultaneously on two types of data: nearly 20 terabytes of global atmospheric observations and the historical IBTrACS database, which covers about 5,000 storms. To gauge uncertainty, the model uses what are known as Functional Generative Networks and builds ensembles of forecasts. Where last season the system produced 50 scenarios at a time, it now produces a thousand, making it possible to capture rare but destructive events such as a storm's rapid intensification.

One curious detail that surprised the scientists themselves: for accurate intensity forecasts, the model needs only a resolution of 28×28 km — a hundred times coarser than traditional methods. A compact version, WeatherNext 2-mini, runs at 111×111 km and can be launched on a single TPU in the free Colab. Why this is possible remains an open question for now.

Real hurricanes

The work has already proved useful in practice. During the 2025 hurricane season, the model helped the U.S. National Hurricane Center forecast the rapid intensification and landfall of Hurricane Melissa in Jamaica, giving people on the ground time to prepare. The model generates a full 15-day forecast in less than a minute on a TPU.

The work was carried out together with forecasters from the National Hurricane Center, the CIRA institute and the UK's Met Office. DeepMind is now open-sourcing the code and weights of the WeatherNext 2 and WeatherNext Cyclones models, in the hope of accelerating progress across the entire meteorological community. The authors stress that official warnings should still be obtained from national weather services, and that the AI is meant to work in tandem with the experience of human experts.

Source: Google DeepMind