Training data, models and predictions for the paper: Evaluating the prediction of subgrid eddy kinetic energy from a machine learning model

Data record for the paper Evaluating the prediction of subgrid eddy kinetic energy from a machine learning model. It holds what is needed to use the trained models and to remake the paper's figures without running the ocean model again. Training data: daily fields from a 1/16° NEMO simulation of the idealised DINO configuration, filtered and coarsened to 1/4°, model years 0061 to 0072 (360-day calendar), for four regions (SO_JET, SO_EAST, SO_GYRE, SO_TROP): three input features (relative vorticity, kinetic energy of the filtered coarse-grained flow, deformation radius over grid spacing), the target (subgrid eddy kinetic energy), and the mesh mask of each region. 164 trained U-Net models (Keras 3), each file carrying its training record: 15 trained on every combination of the four regions, 90 with one or two input features left out, and 59 from a sensitivity study of the training settings. Predictions of each model on the 367-day test period of each region (656 files). Figure inputs: seven further files used by the paper's figures, including a prediction on a fifth region not used in training. DINO configuration: the NEMO v4.2.1 namelist and XIOS output definition of the 1/16° run. The raw model output (about 5.9 TB) is not included. manifest.csv (one row per model: training regions, features, files) and md5sums.txt. The models are loaded with the code ml-parameterisation v1.0.0 (10.5281/zenodo.23815035). The paper's figures, with the code that makes each one from this record, are in a separate record (10.5281/zenodo.23814762). README.md describes the variables, units, regions, data split and things to know before use.