Peetak Mitra, Los Alamos National Laboratory | Gavin Portwood, Los Alamos National Laboratory
In large-scale geophysical flows, the relatively small-scale processes of turbulence and mixing can have a leading-order impact on the prediction of (for instance) ocean circulation and global energy budgets. Such predictions are critical components of weather and climate simulation — geophysical problems where small-scale models help offset the otherwise prohibitively expensive computational cost of simulation. These turbulence closure models attempt to capture dynamics that have complex functional dependence on a potentially broad range of large-scale flow parameters. However, models and frameworks are often phenomenological and heuristic in nature, such that robust model calibration to simulation, observation, and experiment data is a challenge. We'll explain how a nonintrusive supervised GPU-driven machine learning framework, such as Neural ODE, can help improve the state of turbulence models in canonical geophysical flows. We'll also discuss the interpretability of these machine-learning models and provide a roadmap to create more general frameworks for modeling such physics.