GraphMet: Learning Spectral Compositional Koopman Operators for Global-to-Regional Weather Forecasting
🎉 Accepted at NeurIPS 2026!
We are excited to announce that our paper, "GraphMet: Learning Spectral Compositional Koopman Operators for Global-to-Regional Weather Forecasting," has been accepted to the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2026).
GraphMet introduces a spectral Koopman-based framework for global-to-regional weather forecasting, replacing deep message-passing processors with a lightweight, structured linear operator to better preserve fine-scale atmospheric variability over long forecast horizons.
Code Release — Coming Soon
The GraphMet-R implementation will be publicly released in the coming days.
The release will include:
- Pretraining regional fine-tuning pipelines.
- Deterministic and probabilistic forecasting workflows.
- Evaluation scripts and benchmark configurations.
- Documentation and reproducibility instructions.
Key Highlights
- 32% reduction in the 15-day activity deficit compared with GraphCast.
- 11–14% improvement in regional RMSE at 5.5 km resolution.
- ~14× lower inference cost compared with diffusion-based probabilistic forecasting baselines.
- 14,336-parameter Koopman processor replacing a 16-layer message-passing architecture.
- Analytic probabilistic forecasting without expensive diffusion sampling.
Citation
The official BibTeX citation will be added following publication of the NeurIPS 2026 proceedings.
License
This code is licensed under CC BY-NC 4.0.