Leveraging machine learning, realistic simulations, and in-situ observations to infer submesoscale transport from SWOT

Principle Investigator: K. Shafer Smith (New York University)

Co-Investigator(s): Dhruv Balwada, Abigail Bodner, Leah Johnson, Spencer Jones

Collaborator(s): Tatsu Monkman, Ryan Du, Andrew Fagerheim, Jingwen Lyu, Yue Wang

---

Submesoscale flows play a significant role in the ocean’s energy and tracer budgets, producing buoyancy fluxes that restratify the mixed layer, enhance mesoscale flows through an inverse energy cascade, and generate sharp fronts with intense flows that puncture the upper ocean’s stratification, opening pathways for the transport of heat, momentum and tracers between the ocean’s surface and interior. SWOT sea surface height (SSH) observations are revealing structures with scales of a few kilometers, suggesting that global observations of fronts and submesoscale flows are attainable. Because geostrophic balance is inaccurate at these scales, inferring upper-ocean submesoscale transport from SWOT remains a serious challenge. Moreover, submesoscale SSH reflects a mixture of fast ageostrophic wave signals that must be filtered, and flow signatures of ageostrophic transport-active fronts and vortices that contribute strongly to submesoscale transport.

The aim of our project is to build on our previous NASA-funded results to develop robust algorithms to estimate submesoscale velocity statistics and transport directly from SWOT observations of SSH. In phase I, we quantified vertical transport in submesoscale-permitting simulations, elucidating the partition between mesoscale vs. submesoscale, and balanced vs. unbalanced transport of tracers. In phase II, we developed methods to infer submesoscale velocity statistics directly from SSH using high-resolution simulations by (1) using Lagrangian filtering to isolate the non-wave velocity field so that the wave and non-wave components of the velocity can be learned separately; (2) developing a statistical method to identify frontal flows, enabling quantitative estimates of non-wave convergence; and (3) using the above to create a neural network that shows compelling skill in learning submesoscale velocity statistics directly from simulated SSH.

Now, we aim to: (1) develop and refine machine learning (ML) models to predict surface and subsurface flows and transport with the help of SSH and other surface and near surface variables. These ML models will be trained using a variety of realistic submesoscale-permitting simulations. Their accuracy and robustness will be enhanced by introducing dynamically informative quantities in the loss functions, and by experimenting with physics-aware architectures. (2) The ML models developed above will be used with, validated against, and refined by applying them to 1-day and 21-day repeat-cycle SWOT observations, in regions where ample ancillary data exist (e.g. High-Frequency Radar, ARGO, Global Drifter Program drifter array, dedicated campaign data with ADCP measurements) to provide a “truth” signal. We intend also to collaborate closely with, and use high resolution data assimilation products being created by other science teams in the US and France, should they be funded.

On the applications side, we aim to use the above: (1) dynamical diagnose balanced subsurface lateral and vertical currents; (2) quantify the upper ocean kinetic energy cycle; (3) estimate vertical buoyancy fluxes and mixed layer restratification globally; and (4) develop transport operators through which one can infer passive tracer transport for any tracer.