WeatherNext is a machine learning model for global medium-range atmospheric and cyclone forecasting developed by Google DeepMind and Google Research.
The tool addresses the challenge of generating accurate weather forecasts across multiple timescales and phenomena. WeatherNext 2 operates at 0.25-degree resolution and produces deterministic forecasts along with probabilistic cyclone predictions. The repository provides pretrained model weights and inference code, allowing users to generate forecasts by initializing the model from operational weather data. The tool also includes implementations of prior-generation models GraphCast, which uses graph neural networks for deterministic forecasting, and GenCast, which applies diffusion-based methods for ensemble forecasting.
Developers should adopt this tool if they need to run weather forecasts locally or integrate them into applications, though the README emphasizes that accessing daily forecast data feeds through Google Cloud, WeatherLab, or OpenMeteo may be more practical for many use cases. The repository suits research applications, operational weather services, and systems requiring cyclone track predictions. The tool distinguishes itself by offering multiple model variants trained for different purposes—WeatherNext 2 is fine-tuned on operational ECMWF data for direct initialization from real-time conditions, while the cyclone-specialized variant focuses on tropical cyclone forecasting with dedicated training.
The project maintains active development with multiple model variants available as pretrained weights. Documentation spans technical reports, model guides, and specialized papers on cyclone forecasting. The repository serves as the primary home for the WeatherNext family, consolidating current and legacy models in a single codebase with clear separation between different model architectures and training approaches.
