Bris-HourGlass

Bris-HourGlass

This repository contains the Bris-HourGlass (hourly temporal downscaler) checkpoints with matching training configs.

The intended use is training with Anemoi and forecast inference from the published model artifacts.

Contents

• configs/Bris-HourGlass_o96.yaml: global o96 pre-training config
• configs/Bris-HourGlass_n320.yaml: global n320 fine-tuning config
• configs/Bris-HourGlass_stretched.yaml: global+regional n320+2.5km stretched grid fine-tuning config
• Bris-HourGlass_n320_inference.ckpt: global n320 inference checkpoint
• Bris-HourGlass_n320_training.ckpt: global n320 training checkpoint (for further fine-tuning)
• Bris-HourGlass_stretched_inference.ckpt: stretched grid inference checkpoint
• Bris-HourGlass_stretched_training.ckpt: stretched grid training checkpoint

Scope

This is an artifact repository. It provides model weights and configs, but not input datasets.

The source code used for training is open and available through Anemoi Core: https://github.com/ecmwf/anemoi-core

Usage

Training is performed with the Anemoi codebase.

Training and fine-tuning of this model was done on the Anemoi Core branch ecmwf/anemoi-core/tree/feature/ens_interp.

https://github.com/ecmwf/anemoi-core/tree/feature/ens_interp

Porting the checkpoints to a newer version is not supported, but the functionality in that branch is now all on the main Anemoi Core, so for training new models, using main is recommended.

Notes

• bris-crpsfft_inference.ckpt is the checkpoint intended for inference.
• bris-crpsfft_training.ckpt is kept as a training artifact.
• The configs are included to make the artifacts easier to interpret and reuse.

Acknowledgements

Checkpoints were trained on the EuroHPC supercomputer LEONARDO, hosted by CINECA (Italy). Computing and storage resources were provided by EuroHPC through the Regular Access call EHPC-REG-2025R02-263.

Citation

If you use these artifacts, cite: https://arxiv.org/abs/2607.11457

---