Google DeepMind and Google Research report in Nature that their WeatherNext AI model attains state-of-the-art accuracy for tropical cyclone track, intensity, and wind-structure forecasts. The teams say WeatherNext provides, on average, roughly one extra day of actionable lead time — their three-day forecasts match the accuracy previously attainable for two days — and they are releasing code and model weights publicly so researchers and forecasters can build on the work.
Who collaborated and where the results appeared
The work was co-developed by researchers and engineers at Google DeepMind and Google Research in collaboration with expert forecasters at the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office and additional weather agencies. The results are described in a paper published in Nature.
What WeatherNext achieves
WeatherNext is a single AI model that jointly predicts a tropical cyclone’s track, intensity and wind structure with what the authors describe as state-of-the-art accuracy. According to the paper, this improvement corresponds to roughly an extra day of lead time: three-day forecasts now achieve a level of accuracy comparable to what prior models could provide for two days. The authors characterize this scale of improvement as comparable to about a decade of meteorological progress.
Data, training and model design
The model was co-trained on two data modalities: global atmospheric dynamics and expert-curated historical cyclone observations. Training used nearly 20 terabytes of global atmospheric data together with the IBTrACS database, which spans nearly 5,000 historical storms. This combined training helps the model learn complex atmospheric patterns and the behavior of extreme weather.
WeatherNext uses Functional Generative Networks (FGNs) to produce ensembles efficiently and represent the inherent uncertainty of forecasts. The system can generate a single 15-day forecast in less than a minute on a TPU, enabling rapid evaluation of probability distributions for tail risks. Last year the system ran ensembles of 50 members; this year the team scaled ensembles to 1,000 members to better capture rare but consequential scenarios such as rapid intensification events.
Resolution and an unexpected finding
Traditionally, accurate intensity forecasts have been associated with very high spatial resolution. The paper notes that WeatherNext Cyclones attains strong performance using inputs at 28×28 km resolution — roughly 100× coarser than many traditional high-resolution local models — and that WeatherNext 2-mini also performs well at 111×111 km resolution. The authors acknowledge that it remains an open research question how the model produces such accurate results at this coarser resolution and invite the community to investigate further.
Real-world use: Hurricane Melissa (2025)
According to the paper, the model was used operationally during the 2025 hurricane season and assisted the National Hurricane Center with a notable forecast for Hurricane Melissa. WeatherNext reportedly predicted the storm’s rapid intensification and landfall in Jamaica, enabling the NHC to issue advance warnings and provide more time for on-the-ground preparation.
Open-source release and where to try it
Alongside the Nature paper, the teams are open-sourcing the code and model weights. Public releases include WeatherNext Cyclones (the model used during the hurricane season), WeatherNext 2 (an update operationalized in October) and WeatherNext 2-mini, a compact variant that can run on a single TPU and is available via a free Colab notebook. The authors aim for these releases to support academic research, operational forecasting and the development of localized or specialized models.
The group has also refreshed Weather Lab with a new interface and global forecasts; Weather Lab now visualizes WeatherNext predictions for temperature, precipitation, wind speed and cyclone tracks in a single view. Both Weather Lab and the WeatherNext models are part of Google Earth AI.
Why this matters
Tropical cyclones are among the most destructive weather phenomena: over the past 50 years they are associated with more than 700,000 deaths and around $1.4 trillion in economic losses worldwide. Because warnings and preparation often operate on tight timescales, an average increase of about one day in reliable forecast lead time can materially affect evacuation decisions, resource allocation and emergency response.
Invitation to the community
The authors invite researchers, meteorological agencies and operational partners to use and build on the open-source models, examine forecasts on Weather Lab, and collaborate to improve forecasting systems. They emphasize combining advanced machine learning with the expertise of human forecasters to create an ecosystem that can better protect lives and infrastructure as the climate changes.
Acknowledgements
The research was co-developed by Google DeepMind and Google Research. The teams thank collaborators NOAA/NWS/NCEP National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA) and the UK Met Office. The paper’s co-authors include Ferran Alet, Tom Andersson, Ilan Price, Stratis Markou, Andrew El-Kadi, Dominic Masters, Amy Li, Samier Merchant, Natalie Williams, Gregory Thornton, Ken MacKay, Olivia Graham, Akib Uddin, Ben Gaiarin, Devaja Shah, Elinor Kruse, Wallace Hogsett, David Zelinsky, John Cangialosi, Jonathan Martinez, James Franklin, Mark DeMaria, Kate Musgrave, Caroline L. Bain, Helen Titley, Jacklynn Stott, Remi Lam, Aaron Bell, Paul Komarek, Matthew Willson, Alvaro Sanchez-Gonzalez and Peter Battaglia.
Note: for official forecasts and warnings, consult your local meteorological agency or national weather service.
