Towards Using Machine-Learning to Improve Sub-Mesoscale Predictability
Abstract
Approved for Public ReleaseWhile the oceanic mesoscale (20-300 km) is responsible for roughly 95%of the horizontal kinetic energy, the submesoscale (<20 km) and internaltides are responsible for roughly 95% of the vertical kinetic energy(mostly as turbulent kinetic energy). How much is due to thesubmesoscale vs. the internal waves is an open research question.Understanding the role of the submesoscale in the ocean requiressignificant resources in observations and computer simulation. Toresolve both the horizontal and vertical resolutions required for thesubmesoscale demands significant computing power. Moreover, the role ofthe submesoscale in ocean prediction is an open research question.The Philippine Sea is a region characterized with bothstrong mesoscale (horizontal eddy energy) and combinedsubmesoscale/internal wave (vertical eddy energy) that creates achallenging environment to model and predict. For numerical prediction,there are a number of questions, including but not limited to: i) aresub-grid parameterizations robust enough to characterize the unresolvedsubmesoscale to provide improved prediction?; ii) which observationsconstrain numerical models to best represent these processes?; iii) whatresolutions are appropriate for capturing the submesoscale energy?I propose to conduct /basic/ research to develop an application of a newmachine learning technique, physics-informed neural networks to areplace the current parameterized schemes of vertical mixing that arecurrently used in ocean models. This advanced neural network willinclude an update stage that will allow for the assimilation of new data(such as dissipation from a vertical microstructure profiler) andprovide online estimations of the vertical mixing based on the model scurrent state. In addition to this, I propose an /applied/ research taskto utilize a second machine learning framework, 4D-Var, to assimilateall available data (including from ARCTERX) into a 2.5 km ocean model(submesoscale-permitting)of the region to assess the skill ofpredictability with a coarser model. Then, utilizing the new mixingscheme, we can directly compare whether the physics-informed neuralnetwork improves our estimate of the effects of the submesoscale andimproves predictability.The growth of instabilities from the submesoscale alter the flow leadingto a cascade that makes prediction at all scales more difficult. Theseproblems pose a significant challenge to operational oceanography toresolve all scales that may influence US Navy operations. Should theneural network approach improve predictability, it will allow forimproved predictions of the ocean without requiring additionalcomputation from existing numerical models.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2021
- Source ID
- N000142112709
Entities
People
- Brian A Powell
Organizations
- Office of Naval Research
- United States Navy
- University of Hawaiʻi System