Google open-sources AI model for earlier hurricane warnings

Google has made the code and model weights for WeatherNext publicly available on GitHub, opening up an AI system designed to improve tropical cyclone forecasting. The model, developed by researchers at Google DeepMind and Google Research, can generate a 15-day forecast of a storm's track and intensity in less than a minute using Google's TPU hardware. This speed could help forecasters quickly assess the probability of rare but devastating events.

WeatherNext was created with contributions from the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, the UK Met Office, and other weather agencies. The work is described in a study published in Nature and summarized in a blog post from Google.

Forecasting tropical cyclones is challenging because their path is often best predicted by global atmospheric models, while their intensity depends on localized thermodynamic processes near the storm's core. WeatherNext combines both approaches in a single model, trained on nearly 20 terabytes of global atmospheric data and historical storm records from the International Best Track Archive for Climate Stewardship.

According to the researchers, the ability to generate a forecast in under a minute allows forecasters to evaluate a distribution of possible outcomes, including tail risks. This is the second generation of WeatherNext, and Google's research teams have also been working on AI systems to predict flash floods.