Statistical-dynamical analyses and modelling of multi-scale ocean variability

• Main Author: Agarwal, Niraj
• Format: Text
• Language: unknown
• Published: Imperial College London 2021
• Subjects: Pacific; North Atlantic
• Online Access: https://dx.doi.org/10.25560/95810; http://spiral.imperial.ac.uk/handle/10044/1/95810

This thesis aims to provide a comprehensive analysis of multi-scale oceanic variabilities using various statistical and dynamical tools and explore the data-driven methods for correct statistical emulation of the oceans. We considered the classical, wind-driven, double-gyre ocean circulation model in quasi-geostrophic approximation and obtained its eddy-resolving solutions in terms of potential vorticity anomaly and geostrophic streamfunctions. The reference solutions possess two asymmetric gyres of opposite circulations and a strong meandering eastward jet separating them with rich eddy activities around it, such as the Gulf Stream in the North Atlantic and Kuroshio in the North Pacific. This thesis is divided into two parts. The first part discusses a novel scale-separation method based on the local spatial correlations, called correlation-based decomposition (CBD), and provides a comprehensive analysis of mesoscale eddy forcing. In particular, we analyse the instantaneous and time-lagged interactions between the diagnosed eddy forcing and the evolving large-scale PVA using the novel `product integral' characteristics. The product integral time series uncover robust causality between two drastically different yet interacting flow quantities, termed `eddy backscatter'. We also show data-driven augmentation of non-eddy-resolving ocean models by feeding them the eddy fields to restore the missing eddy-driven features, such as the merging western boundary currents, their eastward extension and low-frequency variabilities of gyres. In the second part, we present a systematic inter-comparison of Linear Regression (LR), stochastic and deep-learning methods to build low-cost reduced-order statistical emulators of the oceans. We obtain the forecasts on seasonal and centennial timescales and assess them for their skill, cost and complexity. We found that the multi-level linear stochastic model performs the best, followed by the ``hybrid stochastically-augmented deep learning models''. The superiority of these methods underscores the importance of incorporating core dynamics, memory effects and model errors for robust emulation of multi-scale dynamical systems, such as the oceans.