Identifying Spatial Patterns of Biases in Ocean Turbulence Parameterizations using Unsupervised Machine Learning

Turbulent mixing caused by breaking internal waves is the primary driver of the vertical heat transport and the global overturning circulation of the ocean, and is critical for understanding the role played ocean dynamics in the climate system. To circumvent the complexities in obtaining in-situ measurements of mixing, simplified parameterized models to estimate the rate of mixing are widely used by utilizing relatively easily collected oceanic properties such as temperature and velocity as inputs. However, inaccuracies in predictions by these simplified models arise when certain assumptions in the model are violated. In this study, by incorporating data collected from a global suite of ship based observations, we use a data-driven approach to identify the spatial distribution of two distinct regions in the ocean where large biases in the predictions by the simplified models are possible. Extending this approach, future studies could potentially identify the underlying causes of such disparities to further improve representation of ocean dynamics in coupled climate models.