Download
GraphCast, developed by Google DeepMind, is a research-grade weather forecasting framework that employs graph neural networks (GNNs) to generate medium-range global weather predictions. The repository provides complete example code for running and training both GraphCast and GenCast, two models introduced in DeepMind’s research papers. GraphCast is designed to perform high-resolution atmospheric simulations using the ERA5 dataset from ECMWF, while GenCast extends the approach with diffusion-based ensemble forecasting for probabilistic weather prediction. Both models are built on JAX and integrate advanced neural architectures capable of learning from multi-scale geophysical data represented on icosahedral meshes. The package includes pretrained model weights, normalization statistics, and demonstration notebooks that allow users to replicate and fine-tune weather forecasting experiments in Colab or on Google Cloud TPUs and GPUs.
Features
- Implements GraphCast and GenCast architectures for data-driven weather forecasting
- Pretrained model weights and normalization data available via Google Cloud Bucket
- JAX-based differentiable simulation framework using graph neural networks
- Colab-ready demonstration notebooks for quick experimentation and learning
- Compatible with ERA5 and HRES datasets for historical and operational fine-tuning
- Supports execution on TPUs and GPUs for scalable model training and inference
Project Samples
