Real-World, Perfect-Model, and Cross-Model Ocean Reconstructions with Generative Data Assimilation, 1950–2020

Monthly eight-member ensembles from score-based data assimilation (SDA) with a diffusion model trained on the MIROC6 large ensemble, produced for Medvedev, Cooper, and Manucharyan, "Reconstructing the Global Ocean from the Historical Observing Record with Generative Data Assimilation" (submitted to JAMES). The code is archived at https://doi.org/10.5281/zenodo.23683113. The archive contains three NetCDF files in physical units and the inference checkpoint. * `sda_perfect_model_miroc6_r47i1p1f1.nc`. Perfect-model test. The target is MIROC6 historical member r47i1p1f1, held out of training. January 1950 through September 2014 (777 months).* `sda_cross_model_cesm2_r11i1p1f1.nc`. Cross-model test. The same MIROC6-trained model, applied without retraining to CESM2 historical member r11i1p1f1. January 1950 through September 2014 (777 months).* `sda_realworld.nc`. Assimilation of real observations, namely HadSST4 sea-surface temperature, CRUTEM5 surface air temperature, ICOADS sea-level pressure, and EN4 temperature and salinity profiles. January 1950 through December 2020 (852 months).* `network-snapshot-001184000.pkl`. Diffusion-model weights used for all runs (training checkpoint at 1184 kimg), to be loaded with the code in the repository. Each file has 8 ensemble members. The horizontal grid is a regular 1° latitude–longitude grid, interpolated from the native HEALPix level-6 grid by nearest neighbour. Temperature and salinity use 10 depth bins. The upper nine are at 25, 75, 150, 300, 500, 700, 1000, 1500 and 2000 m. The stored depth 2500 marks the mean of levels below 2000 m. Variables are thetao, so, tos, sos, psl, zos, tas and siconc. Sea-surface salinity, sea-surface height and sea ice were not assimilated directly. The assimilation used 128 steps. Model truth is not included. For questions, contact Andrei Medvedev ([e-mail]).