Application of deep learning in estimating the convective mixing induced by brine rejection

The open-water deep convection induced by brine rejection constitutes a primary process for deep water formation in high latitudes, playing a critical role in the global thermohaline circulation. Proper parameterization of subgrid-scale convective salt plumes arising from brine rejection is crucial for improving climate model simulations of ocean convection. Traditional physically driven parameterization schemes require numerous physical parameters. In this study, we aim to explore an objective, few-data-driven approach by using deep learning for subgrid scale salt plume parameterization. The Backpropagation Neural Network (BPNN) is trained using the output results of the Large Eddy Simulation (LES) model, which is well-suited for simulating the highly turbulent salt plumes. Results show that using the BPNN driven solely by salinity and latitude effectively captures essential information about the mixing coefficient induced by salt plumes. The application of deep learning provides a new perspective for proposing the subgrid scale parameterization of salt plumes.