Deep Learning of Finescale Parameterizations of Internal Wave Dissipation

Accurate estimates of the intensity and spatial distribution of oceanic turbulence are critical to our understanding of the global overturning circulation. However, direct observation of these small-scale turbulent processes through microstructure measurement is too difficult and costly especially in order to provide sufficient sampling of the global ocean. Finescale parameterization provides an alternative approach to infer the dissipation rate of turbulent kinetic energy (ε) at the Kolmogorov scale based on the measurements of shear and/or strain at a much coarser resolution (~ 10 m) and relies on the theoretical transfer of nonlinear internal wave energy to smaller scales. Although the majority of ε in the open ocean appears to be driven by nonlinear wave-wave interactions, parameterization errors of 200% to 300% are typical in the upper ocean. Moreover, the assumptions involved in this parameterization are not universal throughout the ocean. In this study, we propose a data-driven finescale parameterization that employs deep convolutional neural networks and leverages a large compilation of microstructure profiles (available at the UCSD database) to predict ε. The neural network is trained using common hydrographic variables at a range of coarse resolutions consistent with the paradigm of finescale parameterization. The accuracy of the predictions is assessed relative to the theoretical estimates of finescale parameterization.