Global Estimation of Subsurface Eddy Kinetic Energy of Mesoscale Eddies Using Machine Learning

Abstract Oceanic eddy kinetic energy (EKE) is a key quantity for measuring the intensity of mesoscale eddies and developing mesoscale eddy parameterizations in ocean circulation models. Three decades of satellite observations enable a global assessment of sea surface EKE. However, due to the sparseness of in situ observational data, subsurface EKE with a spatial filter has not been systematically studied. Subsurface EKE can be inferred theoretically and numerically from sea surface observations, but is limited by the problem of decreasing correlation with sea surface variables as depth increases. In this study, we propose a dual‐branch neural network approach to reconstruct the monthly mean subsurface EKE using sea surface variables and subsurface climatological variables (e.g., horizontal filtered velocity gradients), inspired by the Taylor‐series expansion of subsurface EKE. Four neural network models are trained on a high‐resolution global ocean reanalysis data set: surface input fully connected neural network model (FCNN), surface input residual neural network model (ResNet), dual‐branch fully connected neural network model (DB‐FCNN), and dual‐branch residual neural network model (DB‐ResNet). The proposed DB‐FCNN and DB‐ResNet models integrate the surface input variables and the vertical profiles of subsurface variables. The DB‐ResNet model outperforms the FCNN, ResNet, DB‐FCNN, and traditional physics‐based models in both regional and global reconstruction of subsurface EKE in the upper 2,000 m. In addition, the DB‐ResNet model performs well for both regional and global observational data based on transfer learning. These results reveal the potential of the DB‐ResNet model for efficient and accurate reconstruction of subsurface oceanic variables.