CEDDAR

Controllable Ensemble Diffusion Downscaling for Atmospheric Rainfall

CEDDAR is a diffusion-based statistical downscaling framework for generating high-resolution precipitation ensembles over Denmark from low-resolution ERA5 precipitation.

It is built around the Elucidated Diffusion Model (EDM) and includes a modular evaluation suite tailored to precipitation (probabilistic, spatial, distributional, and temporal diagnostics).

---

Overview

CEDDAR learns a conditional generative mapping:

ERA5 (low resolution) → DANRA-scale (high resolution)

Key ideas:

• Diffusion-based generative modelling of precipitation fields
• Ensemble generation for uncertainty quantification
• Scale control through sampling parameters (e.g., $\sigma^{*}$)
• Evaluation suite focused on precipitation skill (CRPS, MAE, spatial structure)

---

Example outputs

CEDDAR produces:

• Spatially structured high-resolution precipitation fields
• Ensemble variability across members
• Probabilistic evaluation maps (e.g., mean CRPS)
• Date-based diagnostics with CRPS and MAE

The reduced reproducibility run (repro/02_reduced_run) generates figures similar to the examples above.

Two pretrained models (B0 and B1_GSDF_RGBCE) are available along with a small test dataset (Data_DiffMod_small) for quick testing and reproducibility at https://zenodo.org/records/18643307.

---

Quickstart (Reproducibility)

Install locally:

pip install -r requirements.txt
pip install -e .

Download Data_DiffMod_small from Zenodo () and place it next to the repository:

project_root/
    CEDDAR/
    Data_DiffMod_small/

01_small_test (CPU, smoke test)

bash repro/01_small_test/run_small_test.sh

02_reduced_run (CPU/GPU/LUMI HPC)

Local (CPU/GPU):

bash repro/02_reduced_run/run_reduced_local.sh  # Local CPU/GPU

LUMI HPC (GPU):

sbatch repro/02_reduced_run/run_reduced_lumi.sh  # LUMI HPC (GPU)

See the respective README files in repro/ for expected runtimes and sanity ranges.

---

Usage

CEDDAR is configuration-driven, with modular runners for local and HPC execution, with example usage:

python -m sbgm.cli.main_app --config path/to/config.yaml --mode training

For working examples, use the reproducibility workflowa in repro/ or the full ablation and model bash scripts in bash_scripts/ or bash_ablations/.

---

Data

Full datasets are not included due to size and licensing. A small example dataset (Data_DiffMod_small) is available on Zenodo for testing and reproducibility purposes. The full dataset used in the paper can be reproduced by following the data processing steps outlined in the paper and code, starting from ERA5 precipitation data or by contacting the authors for access to the processed dataset.

---

Citation

@misc{CEDDAR,
    author          = {Quistgaard, Thea and collaborators},
    title           = {CEDDAR: Controllable Ensemble Diffusion Downscaling for Atmospheric Rainfall},
    year            = {2026},
    howpublished    = {GitHub repository}, 
}

---

License

MIT License. See LICENSE file for details.