SFNO-BVMC
Model Introduction
SFNO-BVMC reproduces the ensemble design of Huge ensembles - Part 1, combining spherical Fourier neural operators, multiple training checkpoints, and centered bred vectors. The approach efficiently creates large global ensembles carrying initial-condition and model uncertainty.
Paper: Huge ensembles - Part 1: Design of ensemble weather forecasts using spherical Fourier neural operators
https://doi.org/10.5194/gmd-18-5575-2025
Model Description
The model was proposed by researchers from Lawrence Berkeley National Laboratory, the University of California Berkeley, NVIDIA, Indiana University, and collaborating institutions. It was trained with global 0.25-degree ERA5 reanalysis data, using 1979–2015 for training, 2018 for validation, and 2020 for testing. The model is suitable for global medium-range ensemble weather forecasting and for studying initial-condition uncertainty, model uncertainty, and extreme-event probabilities. Its key feature is the combination of spherical Fourier neural operators, multiple training checkpoints, and paired centered bred vectors to construct large ensembles efficiently.
Usage Instructions
python scripts/fake_data.py
python scripts/train.py
python scripts/inference.py
python scripts/result.py
Synthetic data preserve the full channel and logical [721,1440] global-coordinate protocol through sampled original-coordinate tiles. Runtime outputs are written under result/ and are removed from the release package.
Citation and License
This repository is an independent engineering reproduction of the public SFNO-BVMC specifications, with code licensed under the Apache License 2.0.
The original paper is licensed under CC BY 4.0; the paper, official weights, and data remain subject to the licenses and terms of their respective projects.
- Downloads last month
- 20