Intelligent solutions to monitor and predict ocean health

Abstract: The ocean is central to the health of the planet, but open questions about what drives the circulation hinder our understanding and ability to monitor changes. This talk highlights work towards addressing three critical challenges combining novel computational tools, theory, and data. First, I show how sparse data from the surface ocean can be used to infer subsurface circulation in climate models, and how associated data-driven exploration can advance theoretical understanding of the global ocean circulation. The developed method pioneers dynamical regime identification in model equation-space and a mapping from surface to depth with an ensemble multilayer perceptron and layerwise relevance propagation. Second, I turn to the problem of model bias from difficulty in capturing the effect of small-scale dynamics on the larger ocean system. Using a manifold representation of model equation space, I propose a machine learning based scale-aware method for parameter tuning. Third, I address the question of using and developing trustworthy machine learning for scientific progress. I present a workflow for accelerating scientific discovery that combines quantifying prediction uncertainty with verification of neural network predictions using input feature attribution and oceanographic theory. While rooted in computational and oceanographic expertise, the methods and workflows presented are widely applicable.

Questions? Contact Dave Wright djwright@mit.edu