FastNet: Machine Learning Weather Prediction
As a Research Associate at the Alan Turing Institute, I work on FastNet, a machine learning weather prediction model developed jointly by the Turing and the Met Office as part of the AI4NWP (Artificial Intelligence for Numerical Weather Prediction) programme. The project aims to transform UK weather forecasting by combining AI models with traditional physics-based forecasting. The model takes its name from the Fastnet sea area in the Shipping Forecast.
Project Highlights
- Global Forecasting with GNNs: FastNet is a graph neural network (GNN) that produces global weather forecasts and is benchmarked against established numerical weather prediction methods.
- Physical Consistency: We design loss functions, including a spherical harmonic loss, horizontal gradient terms and improved wind representation, so that forecasts are not only accurate but also physically realistic.
- Cross-Institution Collaboration: The Turing leads AI model design and cloud scaling, while the Met Office contributes meteorological expertise and forecast evaluation metrics.
- Next Steps: Building on its global results, the project is moving towards high-resolution regional forecasts for the UK.
My Contribution
- Encoder graph construction: I owned the study comparing radius-based and k-nearest-neighbour (KNN) edges between the input grid and FastNet’s mesh, at two resolutions (O96 and N320).
- Scale: Experiments used distributed multi-GPU training on NVIDIA A100s with Azure Machine Learning, and the results are included in the FastNet v1.0 paper.
Technical Details
- Machine Learning Methods: Graph neural networks, physics-informed loss design and spherical harmonic representations.
- Data: Trained on the ERA5 reanalysis dataset, which combines satellite, ship and surface observations.
Publications
- Technical overview and architecture of the FastNet Machine Learning weather prediction model, version 1.0 (arXiv, 2025; equal contribution)
- FastNet: Improving the physical consistency of machine-learning weather prediction models through loss function design (Artificial Intelligence for the Earth Systems, 2026)