Published October 20, 2026 | Version v1

Climate downscaling with likelihood functional flow matching

  • 1. ROR icon Lappeenranta-Lahti University of Technology
  • 2. Dell Technologies

Description

# CD-FFM data record: preprocessed data, model weights and results

Data accompanying *Climate downscaling with likelihood functional flow matching* (Liu, Toropainen and Boy).
Code: [GitHub / Zenodo DOI to be added]. Everything here is used by that code; no further software is needed.

## Use

Every archive is laid out like the code's data directory `$CDFFM_ROOT`, so extracting it in place is enough:

```bash
export CDFFM_ROOT=/path/to/cdffm            # same as for the code
sha256sum -c SHA256SUMS
for f in cdffm_static.tar cdffm_test_2019.tar cdffm_checkpoints.tar; do tar xf $f -C $CDFFM_ROOT; done
# then, in the code repository:
cd scripts && python -u fast_eval.py --run w5_flowmap --ckpt best --members 16 --ode-steps 2 --year 2019 --tag test
```

| Archive | Contents | Needed for |
|---|---|---|
| `cdffm_static.tar` | `norm_stats.npz` (two-harmonic seasonal climatologies of NGCD and ERA5, 1981–2015; precipitation transform), `fast_data/static.npz`, `fast_data/static_rr.npz` (grid, ERA5 cell of every 1 km pixel, static descriptors), monthly ERA5–NGCD offset of the training years | everything |
| `cdffm_test_2016.tar`, `cdffm_test_2018.tar`, `cdffm_test_2019.tar` | `fast_data/year_<year>_meta.npz` (ERA5 inputs, dates) and `year_<year>_targets.npy` (standardized NGCD TG/TX/TN at the 1.33 million land pixels, float16); 2019 also `year_2019_rr.npy` (precipitation) | evaluating any model on the test years |
| `cdffm_checkpoints.tar` | `runs/<run>/`: weights-only checkpoint (exponential moving average) with the best validation score, training arguments and validation log, for all 18 models in the paper | evaluating any model without training |
| `cdffm_results.tar.gz` | `runs/<run>/eval_<year>_<tag>/metrics.json` and `psd.npz`, exact log scores, and the paper tables `paper_tables.md`, `paper_ci.md` | regenerating every table on a CPU |
| `cdffm_figure_data.tar` | map fields of the evaluated models (`runs/.../fields_<date>.npz`) and `figures/paper_data/*.npz` | regenerating every figure on a CPU |

Run names (e.g. `w2e_static_pl24_l8` = CD-FFM, final model; `w5_flowmap` = its distilled flow map) are explained
in the code repository (`configs/train/<run>.json`, field `_role`). The training years (2001–2015, about 60 GB
preprocessed) are not included; the code downloads and preprocesses them from the Copernicus Climate Data Store.

## Sources and licence

The arrays are derived from the Nordic Gridded Climate Dataset (NGCD, Copernicus Climate Change Service dataset
`insitu-gridded-observations-nordic`) and ERA5 (Hersbach et al., 2020). Generated using Copernicus Climate Change
Service information [2026]. Neither the European Commission nor ECMWF is responsible for any use that may be made
of the Copernicus information or data it contains. Derived data and model weights in this record: CC BY 4.0.

Files

README.md

Files (13.5 GB)

Name Size
md5:3e2e973ea0ced5800c1f1b411b5656c7
1.4 GB Download
md5:37665ccf036d912173919b4edf787b9e
1.9 GB Download
md5:6f5c58e45d1bbc4b123de2c20d1cf2de
1.5 MB Download
md5:8345660a772420ccdd3e6c7b6de6d639
363.8 MB Download
md5:a62a7ca4129db5c3c83abc8b40f9f4ba
3.0 GB Download
md5:a526eb052472d51ed2be35e168fc4938
3.0 GB Download
md5:d3c29642ef37432e49e9bbdf17c001ee
3.9 GB Download
md5:b6af4b07d6e8203cd4de98c6237c97ec
2.8 kB Preview Download
md5:b6a9b3e4c46a279057b8c713c77770a2
604 Bytes Download