CORDEX-ML-Bench

A benchmarking dataset for data-driven regional climate downscaling

CORDEX-ML-Bench is a standardized benchmark for evaluating machine-learning approaches to regional climate downscaling. It is aligned with the World Climate Research Programme's CORDEX framework and constitutes the first phase of a community initiative to advance data-driven downscaling toward operational readiness, complementing future dynamical downscaling efforts under CMIP7.

The benchmark targets downscaled daily maximum temperature (tasmax) and precipitation (pr) at ~10 km resolution (a 20x increase in resolution) across three pilot regions: the European Alps, New Zealand, and Southern Africa. In the accompanying paper, 40 independently developed ML configurations — spanning traditional ML, convolutional U-Nets, vision transformers, graph neural networks, and generative models (diffusion, flow matching, GANs) — were evaluated under a perfect-model experimental design.

• Paper: CORDEX-ML-Bench: A Benchmark for Data-Driven Regional Climate Downscaling — Experiment Design and Overview (Rampal et al., 2026)
• Original data / supplementary material: Zenodo record 20985924 (DOI: 10.5281/zenodo.20985924)
• Benchmark code, data loaders & training infrastructure: WCRP-CORDEX/ml-benchmark
• Evaluation suite: jgonzalezab/cordex-bench-eval

Dataset Description

Each of the three geographic domains pairs coarse-resolution atmospheric predictors with high-resolution regional-climate-model (RCM) predictands, driven by two different global climate models (GCMs) — one used for training/testing, and a second held out purely to test cross-GCM transferability.

Predictors (coarse resolution, ~200 km, 16×16 grid; 16 variables total): - u, v — zonal / meridional wind components - q — specific humidity - t — temperature - z — geopotential height - (each at 850, 700, and 500 hPa) - Static field: orography (topography, ~10 km)

Predictands (high resolution, ~10 km, 128×128 grid): - tasmax — daily maximum temperature - pr — daily precipitation

Geographic domains

Data splits

• Training data — predictors and predictands for two benchmark experiments:
• ESD_pseudo-reality — standard empirical-statistical downscaling setup (historical period only)
• Emulator_hist_future — physical emulation setup (historical + future periods)
• Test data — three time periods (historical, mid-century, end-century), each provided with:
• perfect predictors — upscaled from the RCM (matching the training distribution)
• imperfect predictors — taken directly from the driving GCM

Underlying files are regional NetCDF archives (~30 GB total, ~5–10 GB per domain) at daily temporal resolution, released on Zenodo as NZ_domain.zip, ALPS_domain.zip, and SA_domain.zip.

Uses

This dataset is intended for training and benchmarking machine-learning models for regional climate downscaling / super-resolution — including empirical-statistical downscaling (ESD) and RCM emulation — and for evaluating cross-domain and cross-GCM generalization ("transferability") of such models.

Dataset Creation

Curated by Neelesh Rampal, José González-Abad, Peter Gibson, Francois Engelbrecht, Jessica Steinkopf, and Caroline Hardy, on behalf of a broader community effort involving 36 authors and the CORDEX Task Team on Machine Learning for Downscaling, as part of the CORDEX/CMIP7 community initiative. See the paper for full details on experimental design, preprocessing, and evaluation metrics.

Licensing

Released under a Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

@dataset{rampal2025cordexmlbench,
  author       = {Rampal, Neelesh and González-Abad, Jose and Gibson, Peter and
                   Engelbrecht, Francois and Steinkopf, Jessica and Hardy, Caroline},
  title        = {{CORDEX-ML-Bench: A benchmarking dataset for data-driven
                   regional climate downscaling}},
  year         = {2025},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.17957264},
  url          = {https://doi.org/10.5281/zenodo.20985924}
}

@article{rampal2026cordexmlbench,
  title   = {CORDEX-ML-Bench: A Benchmark for Data-Driven Regional Climate
             Downscaling -- Experiment Design and Overview},
  author  = {Rampal, Neelesh and González-Abad, José and Addison, Henry and
             Baño-Medina, Jorge and Bettolli, Maria Laura and Blasone, Valentina
             and Booth, Ben and Coppola, Erika and Di Gioia, Serafina and
             Oldham-Dorrington, Joshua and Doury, Antoine and Engelbrecht,
             Francois and Fuentes-Franco, Ramón and Gibson, Peter B. and
             Glawion, Luca and Hardy, Caroline and Ivanov, Mikhail and Lee,
             Hugo Kyo and Legasa, Mikel N. and Olmo, Matias and Orr, Andrew and
             Polz, Julius and Rogers, Martin S. J. and Schillinger, Maybritt
             and Sharma, Shivani and Soares, Pedro M. M. and Sobolowski, Stefan
             and Steinkopf, Jessica and Tang, Wenchang and Tian, Jr-Ben and
             Tomé, Ricardo and Wang, Ko-Chih and Wang, Yi-Chi and Watson,
             Peter A. G. and Wetherell, Tom and Widmann, Martin and Gutiérrez,
             José M.},
  journal = {arXiv preprint arXiv:2606.29172},
  year    = {2026},
  note    = {Submitted to Journal of Advances in Modeling Earth Systems (JAMES)}
}

More Information

• Data-loading and training walkthroughs: data_download.ipynb and experiments.ipynb in WCRP-CORDEX/ml-benchmark
• Region-specific preprocessing:
• NZ domain: nram812/CORDEXBench-nzdomain-preprocessing
• ALPS domain: jgonzalezab/CORDEXBench-alpsdomain-preprocessing