Published November 2025 | Version v2

Models and samples for ML emulators of Met Office UK CPM

  • 1. ROR icon University of Bristol
  • 2. ROR icon Met Office
  • 3. ROR icon Nvidia (United Kingdom)

Description

Contains model weights and samples for emulators of the Met Office's UK CPM

  • diffusion.tar.gz: a diffusion-model-based emulator (CPMGEM, workdirs/score-sde/subvpsde/ukcp_local_pr_12em_cncsnpp_continuous/bham-4x_12em_pSTV) and a variant using less data (workdirs/score-sde/subvpsde/ukcp_local_pr_1em_cncsnpp_continuous/bham-4x_1em_pSTV_lowdata)
  • u-net.tar.gz: a deterministic U-Net emulator (workdirs/score-sde/deterministic/ukcp_local_pr_12em_plain_unet/bham_pSTV-ema-gradcl-256-batch)

Also included are samples of coarsned CPM rainfall downscaled using BCSD, used for comparison (bcsd.tar.gz: workdirs/bcsd-v1/ukcp_local_pr_12em-ws3).

For each model, samples can be found in samples/<CHECKPOINT_ID>/<DATASET>/<XFM_ID>/<SPLIT>/<ENSEMBLE_MEMBER>. e.g. in workdirs/score-sde/subvpsde/ukcp_local_pr_12em_cncsnpp_continuous/bham-4x_12em_pSTV there are folders like samples/epoch_20/bham64_gcm-4x_12em_psl-sphum4th-temp4th-vort4th_pr-historic/bham64_gcm-4x_12em_psl-sphum4th-temp4th-vort4th_pr-pixelmmsstan/test/13

In some cases there are samples path is like samples/<CHECKPOINT_ID>/<DATASET>/<XFM_ID>/<SPLIT>/<ENSEMBLE_MEMBER>/<CONFIG_HASH> where CONFIG_HASH is a hash of the emulator config used to generate the samples. This became useful as tried alternative approaches to generating samples with the same fitted models but mostly these are slightly experimental and not essential to the published work.

For each model, checkpoints directory holds snapshots of model weights from training.

Samples and models were created using code from https://github.com/henryaddison/mlde (in particular, the v0.2.2 tag).

Files

Files (16.9 GB)

Name Size
md5:27b012874462150aa37f4f7507fe2f81
1.0 GB Download
md5:7dfcb54b828c5aefe428bb4f0e681fb2
14.8 GB Download
md5:1e955c7a576175249b7a1fcc09ca1ce7
1.1 GB Download

Additional details

Funding

UK Research and Innovation
UKRI Centre for Doctoral Training in Interactive Artificial Intelligence EP/S022937/1