DeepMind's WeatherNext AI cyclone forecaster gains an extra day of accuracy, Nature paper says

Google DeepMind and collaborators including the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office have published a paper in Nature reporting that its WeatherNext AI model gives forecasters an extra day of predictive accuracy for tropical cyclones. On average, the model's three-day forecasts of a cyclone's track, intensity and wind structure are as good as what prior models managed at two days — an improvement the team compares to roughly a decade of meteorological progress in track and intensity forecasting.

The team writes that predicting cyclones has traditionally required two different modelling techniques. A cyclone's track — where it goes — is steered by large-scale atmospheric currents, best captured by coarser global models. Its intensity — how strong it becomes — is driven by highly localised, fine-scale thermodynamic processes around its core, best captured by specialised, high-resolution local models. WeatherNext is a single model that improves both global weather forecasting and cyclone track, intensity and wind structure with what the team says is state-of-the-art accuracy.

The model was co-trained on two data modalities — global weather dynamics and expert-curated historical cyclone observations — using nearly 20 terabytes of global atmospheric data and the IBTrACS database spanning nearly 5,000 historical storms. It was evaluated on cyclones from 2023 to 2024 against other top weather models. During the 2025 hurricane season it helped the National Hurricane Center predict Hurricane Melissa's rapid intensification and landfall in Jamaica, according to DeepMind.

WeatherNext uses Functional Generative Networks to produce ensembles of forecasts. Last year the system produced 50 predictions at a time, matching global physics models; this year the ensemble was scaled to 1,000 members per cyclone. A single 15-day forecast now runs in less than a minute on a TPU, the team says, so forecasters can quickly evaluate a probability distribution of tail-risk scenarios such as rapid intensification.

Unlike traditional systems, WeatherNext operates at a resolution of 28 by 28 kilometres — about 100 times coarser than traditional models — and a smaller variant called WeatherNext 2-mini works at 111 by 111 kilometres. Why the model still produces accurate cyclone predictions at this resolution surprised the researchers and is an open question. DeepMind is open-sourcing WeatherNext 2 and WeatherNext Cyclones alongside the paper.