Dataset for paper "Generalizable neural-network parameterization of mesoscale eddies in idealized and global ocean models" by Pavel Perezhogin, Alistair Adcroft and Laure Zanna

This repository contains the following data: factor-*-train.zip Is the training data generated from global ocean model CM2.6 factor-*-test.zip and factor-*-validate.zip are held out datasets separated from training dataset in time factor-*-metadata.zip is additional information (horizontal grid, ocean depth, ...) Factor number above stands for the coarsening ratio between high-resolution and coarse grids Data required to plot Figure 1 is in offline-skill-Figure-1.zip Offline evaluation of the subgrid model used in the Paper for online simulations is shown in offline-skill-of-ANN-used-in-online-simulations.zip Offline skill of a bigger ANN trained on all data and with dimensional scaling is provided offline-skill-ANN-32-32-neurons.zip Online metrics required to reproduce Figures 2,3 and 4 are in online-simulation-metrics.zip NW2-*.zip is raw simulation data in the idealized configuration NeverWorld 2 Training dataset factor-*-train.zip contains the following variables: u and v [m/s] are horizontal velocities on Arakawa C grid SGSx and SGSy [m/s^2] are zonal and meridional subfilter forcing such that du/dt=SGSx, dv/dt=SGSy Txx, Txy, Tyy [m^2/s^2]are horizontal subfilter momentum fluxes such that du/dt = d Txx / dx + d Txy / dy, dv/dt = d Txy / dx + d Tyy / dy rel_vort_h, sh_xx and sh_xy_h [1/s] are relative vorticity (rel_vort_h = dv/dx - du/dy), horizontal tension (sh_xx=du/dx - dv/dy) and horizontal shearing strain (sh_xy_h=du/dy + dv/dx). Index h stands for the horizontal interpolation to the center of the computational cell where thickness (h) resides. div=du/dx + dv/dy [1/s] is horizontal divergence deformation_radius [m] is the first baroclinic Rossby deformation radius. deltaU [m/s] is the scale of vertical shear NH [m/s] is the vertical integral of the Buoyancy frequency The deformation radius, vertical shear and Buoyancy frequency can be used in future studies to address generalization to different turbulence regimes. Additional features, which are potentially useful for training (Coriolis, beta effect) can be found in factor-*-metadata.zip.