The ocean surface boundary layer (OSBL) plays a vital role in regulating the exchange of mass and energy between the atmosphere and the ocean interior, primarily through vertical turbulent mixing. These mixing processes are unresolved in ocean climate models, requiring parameterizations that often rely on ad hoc components, introducing uncertainties in climate projections. In this talk, I will present advancements in an energetics-based parameterization of vertical mixing for the OSBL in NOAA’s Geophysical Fluid Dynamics Laboratory (GFDL) ocean model (MOM6). Leveraging machine learning, I will demonstrate how neural networks trained to emulate eddy diffusivity profiles derived from a high-fidelity second moment closure scheme enhance the OSBL mixing parameterization. These neural networks replace ad hoc components while preserving the conservation laws fundamental to ocean model equations, ensuring suitability for climate simulations. The improved parameterization reduces biases in mixed-layer depth and modestly improves tropical upper ocean stratification in global ocean-only simulations. Additionally, interpretable approximate equations, developed as a cost-effective alternative to the neural networks, achieve similar performance improvements. This work demonstrates how machine learning can be used to improve sub-grid parameterizations while remaining consistent with the physical principles required for climate modeling.