How Google DeepMind's Open-Source WeatherNext AI Gives Forecasters An Extra Day To Prepare For Cyclones

Hyderabad: Google DeepMind's WeatherNext, an artificial intelligence (AI) model, has achieved state-of-the-art accuracy in forecasting the track, intensity and wind structure of tropical cyclones. According to a research paper, published in the Nature journal, WeatherNext Cyclones (WN-C), an open sourced specialised AI model gives forecasters an extra day's worth of predictive accuracy. It offers three-day (72 hours) forecasts that now match the reliability of older models that could only achieve two days (48 hours) predictions.

Tropical cyclones, also referred to as hurricanes or typhoons, have caused more than 700,000 deaths and $1.4 trillion (roughly £1.1 trillion) in economic losses worldwide over the past 50 years.

Google DeepMind said the improvement delivered by WeatherNext is roughly equivalent to a decade's worth of meteorological progress. The research was developed jointly by teams at Google DeepMind and Google Research, working alongside forecasters at the US National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office and other weather agencies globally.

Real-world impact during Hurricane Melissa

The model has already proven its worth in practice. During the 2025 hurricane season, WeatherNext assisted the National Hurricane Centre (NHC) in making a landmark forecast for Hurricane Melissa, by correctly predicting the storm's rapid intensification and its landfall in Jamaica. This allowed authorities to issue an early warning, giving ground teams crucial extra time to prepare. Building on this, the model is now being used to generate 1,000 possible scenarios per cyclone, helping forecasters weigh up a wider range of outcomes.

How the WeatherNext Cyclones model works?

Traditionally, forecasting a cyclone's track and its intensity required two separate approaches, since track is shaped by broad global atmospheric currents while intensity depends on fine-scale processes near the storm's core. WeatherNext bridges this gap using a single model trained on nearly 20 terabytes of global atmospheric data, combined with the IBTrACS historical database covering close to 5,000 past storms.

The system uses what Google calls Functional Generative Networks (FGNs) to rapidly generate ensembles of forecasts that reflect weather's inherent uncertainty. A full 15-day forecast can now be produced in under a minute on a single TPU chip.

It is worth noting that WeatherNext Cyclones operates at a resolution of just 28x28 km, which is around 100 times coarser than traditional high-resolution models, but still it achieved state-of-the-art accuracy, a result researchers say has surprised scientists and remains only partly understood.

Open access for researchers

Alongside the Nature paper, Google DeepMind is releasing the code and model weights for WeatherNext Cyclones, WeatherNext 2, and a lightweight version called WeatherNext 2-mini, which can run in a free public Colab notebook on a single TPU. The company said the aim is to let academic researchers, forecasting agencies and nonprofits build on the technology to improve predictions for a wide range of weather events.

Users can also explore live forecasts through Weather Lab, a recently redesigned platform that now includes global weather forecasts for temperature, precipitation and wind speed, alongside cyclone tracking, all part of Google's broader Earth AI initiative.

Google DeepMind said it hopes the release will foster a more collaborative forecasting ecosystem, combining AI with the expertise of human forecasters to help communities better prepare for extreme weather as climate patterns shift.