google-deepmind/graphcast
Graphcast
GraphCast is a machine learning model that uses graph neural networks to produce global weather forecasts up to ten days ahead at high spatial resolution. The system represents the Earth's surface as an icosahedral mesh, enabling message passing between mesh nodes to capture atmospheric dynamics, and combines this with a learned multiscale processor that operates across coarse-to-fine mesh resolutions.
The model is trained on historical ERA5 reanalysis data through a supervised learning objective, and its autoregressive rollout loop feeds predictions back as input to generate multi-step forecast trajectories while maintaining end-to-end differentiability. GraphCast includes a grid-to-mesh encoder-decoder that converts regular latitude-longitude grid data into the triangular mesh format for processing and decodes outputs back to grid coordinates.
The repository provides three pretrained model snapshots that can generate forecasts without training from scratch, along with a training pipeline for producing custom forecasting models. Conversion utilities are included to transform gridded climate data into the icosahedral mesh representation required by the graph network.
Features
- Weather Forecasting Graph Networks - Uses graph neural networks to forecast global weather conditions up to 10 days ahead at high spatial resolution.
- Differentiable Forecast Trajectories - Maintains end-to-end differentiability through the autoregressive rollout for gradient-based training.
- Weather Forecast Rollouts - Feeds model predictions back as inputs to generate a differentiable multi-step forecast trajectory.
- Autoregressive Forecast Generators - Feeds model predictions back as inputs to produce a differentiable multi-step forecast trajectory.
- Graph Neural Network Forecast Trainers - Trains graph neural networks from scratch on historical ERA5 reanalysis data to produce custom forecasting models.
- ERA5 Reanalysis Training Pipelines - Trains the graph neural network on historical ERA5 reanalysis data using supervised learning.
- Graph Neural Networks - Trains graph neural networks on gridded climate reanalysis data to produce custom forecasting models.
- Icosahedral Mesh Graph Networks - Represents the Earth's surface as a triangular icosahedral mesh for graph neural network message passing.
- Multiscale Mesh Processors - Uses a hierarchy of coarse-to-fine mesh resolutions to capture both global and local atmospheric dynamics.
- ERA5 Reanalysis Training Pipelines - Trains graph neural network parameters from historical ERA5 reanalysis data to produce custom forecasting models.
- Global Model Forecasts - Predicts global weather conditions days ahead using learned graph neural network models on mesh data.
- Graph Neural Network Forecasts - Predicts global weather conditions up to 10 days ahead at high spatial resolution using a learned graph neural network.
- Grid-to-Mesh Encoder-Decoders - Converts regular latitude-longitude grid data into icosahedral meshes for graph network processing.
- Pretrained Model Snapshots - Loads and executes pretrained model snapshots to generate weather forecasts without training.
- Pretrained Weather Forecast Models - Provides three pretrained model snapshots that generate global weather forecasts without training from scratch.
- Grid-to-Mesh Converters - Transforms regular latitude-longitude grid data into triangular icosahedral meshes for graph network processing.
- Grid-to-Mesh Spatial Modeling - Converts regular latitude-longitude grids into triangular icosahedral meshes for graph network processing.
