Google DeepMind open-sources WeatherNext, an AI that forecasts cyclones a day earlier

Google DeepMind and Google Research have published a Nature paper on WeatherNext, an AI model that predicts a tropical cyclone’s track, intensity, and wind structure with state-of-the-art accuracy. The key result is roughly a full extra day of lead time: its three-day forecasts match the quality of what earlier models managed at two days, a jump the team likens to about a decade of conventional meteorological progress. A single model now handles both the large-scale atmospheric steering that governs a storm’s path and the fine-scale physics that drives intensity — problems previously split between separate coarse and high-resolution systems.

The approach combines co-training on nearly 20 terabytes of global atmospheric data plus the IBTrACS record of about 5,000 historical storms, with Functional Generative Networks that produce forecast ensembles to capture uncertainty. Notably, WeatherNext Cyclones operates at 28km resolution — roughly 100x coarser than traditional models — and a mini variant works at 111km, yet still delivers strong intensity forecasts; why coarse inputs work this well remains an open question. The system generates a 15-day forecast in under a minute on a TPU and has scaled from 50 to 1,000 ensemble members to catch rare events like rapid intensification. It contributed to a real forecast of Hurricane Melissa’s intensification and Jamaica landfall during the 2025 season.

DeepMind is releasing code and weights for WeatherNext 2, WeatherNext Cyclones, and a compact 2-mini version that runs in a free Colab notebook, alongside a refreshed Weather Lab visualization tool. The open-sourcing is aimed at meteorological agencies, academics, and nonprofits building localized or operational forecasting systems, extending a growing trend of AI weather models displacing purely physics-based numerical prediction.