Google's DeepMind and Google Research teams, working with the US National Hurricane Center, the UK Met Office and other agencies, have developed an AI weather model called WeatherNext designed to improve cyclone forecasts. Google is releasing both the code and model weights as open source on GitHub, allowing other scientists to build on the work, with the underlying research published in the journal Nature. The aim is to give forecasters more advance warning of a storm's likely path and strength, potentially helping communities prepare earlier for devastating hurricanes and typhoons.

Predicting cyclones has traditionally required two separate approaches: broader global models to track a storm's path and specialised local models to assess intensity based on core thermodynamics. WeatherNext combines both into a single system, trained on nearly 20 terabytes of global atmospheric data alongside historical records from the International Best Track Archive for Climate Stewardship, to produce a 15-day forecast of track and intensity in under a minute using a TPU. It follows last year's launch of WeatherNext's second generation, and builds on Google's wider work applying AI to weather risks, including flash flood prediction.