bris-forecaster

Bris Forecaster

This repository contains the Bris forecaster checkpoints and matching training and inference configurations.

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

Contents

• bris-crpsfft_inference.ckpt: inference checkpoint
• bris-crpsfft_training.ckpt: training checkpoint artifact
• configs/config_inference.yaml: inference configuration
• configs/config_training.yaml: training configuration
• pyproject.toml: pinned Python project metadata for inference
• uv.lock: locked dependency set for the inference environment

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

The configurations in configs/config_inference.yaml and configs/config_training.yaml are the references for concrete settings.

Usage

Training is performed with the Anemoi codebase.

Training and fine-tuning of this model require the forked Anemoi Core branch evenmn/anemoi-core:feat/crps-fft-loss rather than current upstream Anemoi Core:

https://github.com/evenmn/anemoi-core/tree/feat/crps-fft-loss

That branch carries the custom spectral CRPS loss used by the training config (anemoi.training.losses.CRPSFFTLoss; see training/src/anemoi/training/losses/afcrps_fft.py in the fork). It is based on an older Anemoi version, so training or fine-tuning against newer upstream Anemoi releases is not expected to work without porting the loss implementation and any related training code.

Environment

This repository includes a pyproject.toml and uv.lock for the pinned environment.

Setup with uv:

1. Install uv.
2. Ensure you have SSH access to github.com/metno/bris-inference, since the bris dependency is fetched from Git.
3. Create the environment with uv sync.

You can then run commands inside the environment with uv run.

This pinned environment is primarily for inference and artifact inspection. If you want to train or fine-tune, set up Anemoi from the fork/branch above and run training from that checkout while using this repository's configs/config_training.yaml as the reference configuration.

From the forked Anemoi checkout, run training with this repository's config on the Hydra search path, for example:

uv run anemoi-training train --config-path=/path/to/bris-forecaster/configs --config-name=config_training.yaml

Inference is intended to be performed with Anemoi Inference.

The inference checkpoint is included here, and the Anemoi Inference config will be added once that interface is finalized.

In practice, this repo is meant to be used as:

• the published checkpoint artifact for inference
• companion configs showing how the model was trained and run

If you need operational details, use the config directly rather than this README.

The environment is pinned in this repository through pyproject.toml and uv.lock.

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:

Even Marius Nordhagen, Håvard Homleid Haugen, Aram Farhad Shafiq Salihi, Magnus Sikora Ingstad, Thomas Nils Nipen, Ivar Ambjørn Seierstad, Inger-Lise Frogner, "High-Resolution Probabilistic Data-Driven Weather Modeling with a Stretched-Grid," arXiv:2511.23043, 2025.

Reference: https://arxiv.org/abs/2511.23043