Applying machine learning in devising a parsimonious ocean mixing parameterization scheme

Abstract

Ocean mixing parameterization schemes have a strong impact on the accuracy of numerical model results. Traditionally, physics-based parameterizations consider a host of physical oceanic parameters which affect and interact with each other. The development and deployment of machine learning provide an opportunity to propose a new, unified ocean mixing parameterization scheme. This study uses K-profile parameterization (KPP) output from the Regional Ocean Modeling System (ROMS) and uses temperature vertical diffusion coefficient (AKt) to train a backpropagation neural network (BPNN) in calculating AKt by KPP. Crucially, it was shown that it is feasible to estimate the ocean mixing coefficient through T alone. This study is a new application of machine learning in ocean dynamic parameter estimation, which makes it possible to propose an ocean mixing parameterization with less dependence on dynamic parameters.