DeepMind open sources WeatherNext AI to boost cyclone forecasting accuracy

Google DeepMind has released its WeatherNext artificial intelligence models into the public domain, marking a significant shift in how tropical cyclones are tracked and predicted. The open-sourced release includes WeatherNext Cyclones and WeatherNext 2, following a breakthrough in forecasting cyclone tracks, intensity, and wind structure. The technology offers an average of one additional day of predictive accuracy compared to previous systems, a leap that DeepMind equates to a decade of meteorological progress.

The models were put to the test during the 2025 hurricane season, where they assisted the National Hurricane Center in forecasting the rapid intensification and landfall of Hurricane Melissa in Jamaica. This capability enabled the centre to issue advance warnings, providing critical preparation time for emergency teams on the ground. The success of the system has prompted the release of code and model weights for academic research, operational forecasting, and the development of specialised local models by global weather agencies.

WeatherNext Cyclones bridges the traditional gap between global atmospheric modelling and localised intensity forecasting. While track prediction relies on massive global currents, intensity is driven by fine-scale thermodynamic processes. The model achieves state-of-the-art accuracy using input data with a resolution of 28x28km, which is 100 times coarser than traditional models. It was trained on nearly 20 terabytes of global atmospheric data and the IBTrACS database, spanning nearly 5,000 historical storms.

To capture inherent weather uncertainty, the model utilises Functional Generative Networks to produce ensembles of predictions. During the 2025 season, the system scaled its ensemble size to 1,000 members, allowing forecasters to evaluate the probability distribution of rare but consequential scenarios, such as rapid intensification events. A single 15-day forecast can now be generated in less than a minute on a TPU, significantly accelerating the evaluation of potential tail-risks.

DeepMind is also releasing WeatherNext 2-mini, a compact version operating at a coarser 111x111km resolution. This variant is designed to run on a single TPU via a free public Colab notebook, making the technology accessible to individual researchers and smaller institutions. The models are part of Google Earth AI, with visualisations available through the refreshed Weather Lab interface, which now includes global weather forecasts alongside cyclone tracks.

The research was co-developed by Google DeepMind and Google Research teams in collaboration with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office. By making these tools openly available, the organisations aim to empower the research community to build more resilient communities against extreme weather events.