Doctor of Oceanography and Coastal Sciences (POCS)


Department of Oceanography & Coastal Sciences

Document Type



Turbulence and its associated mixing effects in the ocean surface boundary layer are crucial to ocean environment and the earth system. However, the small-scale nature of turbulence presents a formidable challenge in directly resolving those phenomena in realistic ocean models. Consequently, realistic ocean models rely on parameterization schemes in ocean models, developed from high-resolution turbulence-resolving models to represent the mixing effects of small-scale turbulence. Inadequate understanding of turbulent mixing incorporated into those parameterization schemes lead to substantial deviations of model outputs from observations.

This dissertation first investigates turbulent mixing behavior in an idealized frontal zone using large-eddy simulations. The frontal zone is represented by a constant temperature gradient in x-direction and is subject to varying wind-wave conditions. The alignment between wind and the frontal zone critically affects submesoscale instabilities, thermal fluxes, horizontal and vertical turbulent mixing processes. Thus, a promising direction to improve turbulent mixing schemes within a frontal zone is to consider the alignment between winds and frontal zones.

Addressing the existing challenge to improve mixing parameterizations through conventional physics-based methods, this dissertation explores the use of deep neural network schemes to represent the vertical turbulent mixing in the ocean surface boundary layer. Two approaches are attempted. In the first approach, solutions from turbulence-resolving large eddy simulation models driven by realistic forcing conditions observed in an ocean station are used to train the neural networks. The results demonstrate that a well-tuned neural network scheme predicts turbulent mixing trend more accurately than some traditional physics-based parameterizations. However, granting a neural network scheme fully flexibility in predicting turbulent mixing properties results in the accumulation of spurious errors over time, thus undermines its reliability for long-term predictions. The other approach is to develop a neural-network-embedded parameterization while retaining the core framework of traditional parameterization schemes. The physics within traditional parameterzations is retained while the parameters that require tuning and are uncertain in traditional parameterizations are obtained using neural networks. The neural-network-embedded parameterization scheme is stable during long-term integration. It outperforms traditional schemes and mitigates some common limitations like high sea surface temperature and shallow mixed layers.



Committee Chair

Jun-Hong Liang

Available for download on Monday, April 07, 2031

Included in

Oceanography Commons