Opportunities and pitfalls in automated calibration of ocean models

In this talk we argue that machine learning methodologies — which systematically minimize the error between model predictions and extensive data — have the potential to accelerate the development of theory-based parameterizations. We provide an example that uses optimization with Ensemble Kalman Inversion to develop a theory-based boundary layer turbulence closure that is sparsely parameterized compared to typical data-driven parameterizations and extrapolates surprisingly well. We then turn to the next step and ultimate goal, which is to use observations to guide parameterization development rather than LES, thereby addressing missing physics in the LES and edging closer to the ultimate goal of climate prediction. But there be dragons: because calibration explicitly embraces compensating errors, calibration in the global context may fail if the model is missing significant physics, such as submesocale restratification processes. We discuss the implications of this pitfall and strategies to overcome it.