132 / 2023-08-15 19:29:31

Application of deep-learning parameterization of turbulent mixing in the upwelling region east of Hainan Island

oceanic mixing parameterization,deep learning,SHAP,coastal upwelling

Uncertainty associated with oceanic mixing parameterizations poses a significant challenge in ocean and climate modeling, particularly in the upwelling region off the eastern coast of Hainan Island. The conventional physical-driven parameterization methods have not yielded satisfactory results due to limited understanding of the underlying processes. However, recent advancements in deep learning techniques and the availability of extensive turbulence data offer opportunities to explore data-driven approaches for parameterizing oceanic mixing.To address the complex dynamics of the coastal upwelling region, we propose a subregional deep learning turbulent mixing parameterization scheme. This scheme utilizes hydrological survey data collected during the summer of 2012 in the east shelf sea area of Hainan Island for subregional training and modeling. It incorporates the normalized depth (D) as an input to the deep learning model. The results reveal that the proposed model exhibits improved predictive performance, with the correlation (r) increasing from 0.49 to 0.62 and the root mean square error (rmse) decreasing from 0.56 to 0.50. Comparative analysis against traditional parameterization schemes, such as the G89 scheme and MG03 scheme, demonstrates that the data-driven approach achieves higher accuracy and generalization.Additionally, we employ the SHAP (Shapley Additive Explanations) model to provide insights into the proposed model. The analysis reveals the importance order of input features in our study's model as follows: D, U, ρ, N², S², Ri. Furthermore, it elucidates the effects of each input feature on the prediction results, offering valuable guidance for turbulence parameterization. By integrating machine learning into the investigation of oceanic turbulent mixing, this research showcases the applicability of constructing deep learning-based parameterization schemes utilizing limited observational data and well-established physical processes. These findings open avenues for enhancing turbulent mixing parameterization schemes in ocean models.