文献类型：期刊文献

英文题名：Parameterization of turbulent mixing by deep learning in the continental shelf sea east of Hainan Island

作者：Hu, Minghao[1];Xie, Lingling[1,2,3];Li, Mingming[1,2,3];Zheng, Quanan[4]

机构：[1]Guangdong Ocean Univ, Coll Ocean & Meteorol, Lab Coastal Ocean Variat & Disaster Predict, Zhanjiang 524088, Peoples R China;[2]Key Lab Climate Resources & Environm Continental S, Zhanjiang 524088, Peoples R China;[3]Minist Nat Resources, Key Lab Space Ocean Remote Sensing & Applicat, Zhanjiang 524088, Peoples R China;[4]Univ Maryland, Dept Atmospher & Ocean Sci, College Pk, MD 20742 USA

年份：2025

外文期刊名：JOURNAL OF OCEANOLOGY AND LIMNOLOGY

收录：SCI-EXPANDED(收录号：WOS:001386858700001)、、Scopus(收录号：2-s2.0-85213808351)、WOS

基金：* Supported by the National Natural Science Foundation of China (No. 42276019) and the Guangdong Provincial Observation and Research Station for Tropical Ocean Environment in Western Coastal Waters (No. GSTOEW)

语种：英文

外文关键词：ocean turbulent mixing; parameterization; continental shelf sea; deep learning; SHapley Additive Explanations (SHAP)

外文摘要：The uncertainty of ocean turbulent mixing parameterization comprises a significant challenge in ocean and climate models. A depth-dependent deep learning ocean turbulent mixing parameterization scheme was proposed with the hydrological and microstructure observations conducted in summer 2012 in the shelf sea east of Hainan Island, in South China Sea (SCS). The deep neural network model is used and incorporates the Richardson number Ri, the normalized depth D, the horizontal velocity speed U, the shear S-2, the stratification N-2, and the density rho as input parameters. Comparing to the scheme without parameter D and region division, the depth-dependent scheme improves the prediction of the turbulent kinetic energy dissipation rate epsilon. The correlation coefficient (r) between predicted and observed lg epsilon increases from 0.49 to 0.62, and the root mean square error decreases from 0.56 to 0.48. Comparing to the traditional physics-driven parameterization schemes, such as the G89 and MG03, the data-driven approach achieves higher accuracy and generalization. The SHapley Additive Explanations (SHAP) framework analysis reveals the importance descending order of the input parameters as: rho, D, U, N-2, S-2, and Ri in the whole depth, while D is most important in the upper and bottom boundary layers (D <= 0.3 & D >= 0.65) and least important in middle layer (0.3