A data-driven investigation into the behaviour and parameterisation of mesoscale eddies

• Main Author: Bolton, T
• Other Authors: Zanna, L
• Format: Thesis
• Language: English
• Published: 2021
• Subjects: Turbulence; Ocean; Physics; Oceanography; Machine learning; North Atlantic
• Online Access: https://ora.ox.ac.uk/objects/uuid:6ec97e71-99bc-43d0-a73f-584723753b8c

Mesoscale eddies, turbulent oceanic features of length-scales 10-100km, are an important component of the climate system. They impact the large-scale circulation of both the ocean and atmosphere, as well as the transport of tracers such as heat and carbon. Due to the computational costs of running ocean models at high-resolution, climate models generally do not fully resolve mesoscale eddies, and will not be able to for the foreseeable future. We, therefore, need to parameterise the effects of mesoscale eddies on the large-scale flow. Eddy parameterisations are typically derived from physical theories and mechanisms; this approach can sometimes over-simplify the underlying dynamical process. The primary aim of this thesis is to uncover new ways of capturing the complex non-linear behaviour of mesoscale eddies. Using a hierarchy of numerical models, we take a data-driven approach to studying and parameterising mesoscale eddies. Using an observation-driven experiment, we investigate how the spatio-temporal variability of real mesoscale eddies affect tracer fluxes in the North Atlantic. In an idealised model, we demonstrate that modern machine learning algorithms, namely convolutional neural networks, can accurately represent eddy momentum fluxes. We also demonstrate that machine learning can construct expressions for eddy momentum and temperature fluxes while retaining interpretability. We find that, when implemented, data-driven parameterisations can emulate the scale-interaction between mesoscale eddies and the resolved flow, increasing the kinetic energy and improving the higher-order statistics of the velocity field. Our results show that data-driven algorithms, while respecting physical principles, can be used to parameterise the effects of mesoscale eddies, with implementation in more realistic models now a possibility. The predictive skill of data-driven algorithms is not the only advantage: our use of interpretable machine learning opens up the door for more knowledge discovery studies about eddies and ...