Application of data-driven mixing parameterization scheme in a regional ocean model

Abstract

In this study, based on the data-driven parameterization proposed in previous studies, we implemented a data-driven vertical turbulence parameterization scheme (backpropagation neural network, BPNN) into a regional ocean model and compared the simulation results with those obtained using the traditional physics-driven scheme (K-profile parameterization, KPP). The Kuroshio-Oyashio Confluence Region (KOCR) with rich ocean dynamics was selected as the study region. Comparisons of modeled temperature, salinity, and velocity outputs show that the two parameterization schemes produce similar spatial and temporal distribution characteristics. For the sea temperature, the maximum difference appeared at the subsurface layer (3.6 °C) while the maximum mean difference appeared at the surface layer (0.1 °C). Overall, sea temperature in the surface layer simulated by the model adopting the BPNN parameterization scheme was warmer. For the salinity, the maximum difference (0.28 psu) and maximum mean difference (1 × 10−3 psu) appeared in the subsurface layer. In the surface layer, when the model adopting BPNN parameterization scheme overestimated sea temperature compared to the KPP scheme, the result was usually accompanied by overestimated salinity. In the bottom layer, when the model adopting the BPNN parameterization scheme simulated temperature overestimation, the result corresponded to underestimated salinity. For the velocity, the maximum difference (0.31 m/s) and maximum mean difference (-3 × 10−3 m/s) between the two parameterization schemes appeared in the surface layer, and the high values of velocity difference generally occurred in sea areas with energetic dynamic processes. The KPP scheme and the model adopting BPNN parameterization produced similar mixed layer depth (MLD) spatial distribution and ocean fronts spatial distribution, but the latter parameterization scheme simulated an overall shallower mixed layer. For the ocean fronts, the differences between the two parameterization schemes (O∼10−5) is smaller than the intensity of the fronts (O∼10−4). This study presents a new attempt to apply the data-driven parameterization scheme in ocean numerical models.