Exploring the use of machine learning to parameterize vertical mixing in the ocean surface boundary layer

Abstract

In ocean and climate models, the simulation of upper-ocean temperature and salinity depends on mixing parameterizations for ocean surface boundary layer turbulence. Existing mixing parameterizations are based on physical principles with empirical parameters. However, they are still imperfect, leading to biases in the simulation of physical states in the upper ocean. In this study, we explore the use of the data-based machine learning technique, specifically, a deep neural network model, for the effects of vertical mixing in the ocean surface boundary layer. The model is trained using process-oriented simulations of the upper-ocean turbulence driven by realistic forcing conditions at the Ocean Station Papa that is a mid-latitude ocean climate station. The deep neural network model outperforms traditional physics-based parameterizations that relate the mixing effects to surface forcing using deterministic formulas. The deep neural network model is also used to explore two currently debated issues in the development of physics-based mixing parameterizations, including the representation of wave forcing and the history of forcing conditions.