DeepMind published research in Nature on August 6 showing that its WeatherNext AI model achieves state‑of‑the‑art accuracy for tropical cyclone track, intensity and wind structure — producing three‑day forecasts that match the accuracy previous models reached at two days, effectively adding roughly 24 hours of reliable warning.
WeatherNext’s performance and warning benefit
WeatherNext runs a 15‑day, 1,000‑member ensemble in under a minute on a single Tensor Processing Unit, where comparable physics‑based systems would take hours on a supercomputer. Rather than a single deterministic output, the model returns a probability distribution across 1,000 coherent future scenarios, helping forecasters identify low‑probability, high‑impact events such as rapid intensification. The model family includes a Cyclones variant operating on 28‑kilometer input data and a smaller variant using 111‑kilometer inputs; DeepMind notes that why accuracy survives such coarse resolution remains an open research question.
Field test: Hurricane Melissa
During the 2025 Atlantic hurricane season, the National Hurricane Center ran WeatherNext alongside operational guidance and used it to forecast Hurricane Melissa’s rapid intensification and catastrophic landfall in Jamaica. Melissa struck southwestern Jamaica on October 28, 2025, as a Category 5 hurricane with sustained winds of 185 miles per hour. World Bank and Inter‑American Development Bank assessments put physical damage at $8.8 billion, equivalent to 41 percent of Jamaica’s 2024 GDP; the storm killed at least 45 people and displaced tens of thousands. According to the report, the earlier warning enabled by AI forecasting gave communities critical time to prepare.
Architecture, training and prior work
WeatherNext builds on DeepMind’s earlier GraphCast, a 2023 graph neural network model published in Science that outperformed the European Centre for Medium‑Range Weather Forecasts’ operational system on 90 percent of tested variables. Unlike GraphCast’s single deterministic forecasts, WeatherNext uses Functional Generative Networks, a diffusion‑based approach that samples from the joint distribution over 15‑day global weather trajectories so each ensemble member is a dynamically coherent alternative future. The system was co‑trained on roughly 20 terabytes of global atmospheric data alongside the IBTrACS historical database of nearly 5,000 tropical cyclones, collapsing the traditional two‑pipeline approach (coarse global models for track and high‑resolution regional models for intensity) into a single unified model.
Open source, products and deployment
DeepMind released the code and model weights on GitHub under the Apache 2.0 license, with accompanying materials under Creative Commons Attribution 4.0; both permits allow commercial use with attribution, a change from earlier non‑commercial weather releases. Three variants are available — WeatherNext Cyclones, WeatherNext 2, and a compact WeatherNext 2‑mini capable of running in a free Google Colab notebook — and live outputs are accessible through Weather Lab within the Google Earth AI umbrella. The model family is already being integrated into Google consumer products including Search, Gemini, Pixel Weather and the Maps Platform Weather API. DeepMind says future development may extend the approach to other extreme weather events.
The report frames WeatherNext as a potential multiplier for early warning systems that have already driven a threefold reduction in cyclone deaths over the past five decades, while noting that gains will not be evenly distributed and that open‑sourcing the weights removes one barrier to deployment.
