Small scale vertical mixing processes in the upper ocean region cannot be resolved by ocean models due to the lack of computing resources, and are thus parameterized. Existing models have a few ad hoc approximations which cause uncertainty in climate change simulations. In this breakthrough preprint , Aakash Sane and co-authors - Brandon Reichl, Alistair Adcroft, and Laure Zanna - replace such an approximation using data driven neural networks (NN), leading to improved physics in an ocean model simulation and a reduction of model bias in the Tropics. Moreover, the NN implementation in MOM6 is stable, and does not violate any conservation laws.