Google Publishes WeatherNext Source Code
Late on the evening of August 7 at 19:30, Google announced the release of the code and parameters of its new AI model WeatherNext. This system is designed to predict the trajectory and strength of tropical cyclones up to two weeks in advance. For researchers, this is an important signal: open algorithms of this level can be studied and adapted to their own tasks without needing to wait for commercial releases.
The project was a collaborative effort between the Google DeepMind and Google Research teams, the National Hurricane Center of the USA, the meteorological service of the UK, and other organizations. WeatherNext focuses on two key characteristics of a cyclone — its movement path and intensity. Almost 20 terabytes of data were used for training, including information on the Earth's atmosphere and historical records of tropical storms.
A distinctive feature of the new model is its speed: a 15-day forecast is generated in less than a minute on specialized TPU processors. The current version is the second generation of WeatherNext, the first of which Google introduced in 2025. Now, the code and parameters of the model are hosted on GitHub and accessible to everyone. At the same time, Google teams are also working on applying AI to predict sudden flooding.
The development of WeatherNext is an important step in improving the accuracy of tropical cyclone forecasts, which directly impacts the quality of planning and preparation for natural disasters.
The opening of the code gives the scientific community the opportunity to study and enhance forecasting algorithms based on new data and technologies. This underscores the growing importance of artificial intelligence in meteorology and risk management — it increasingly affects the protection of populations and infrastructure in hurricane-prone areas.