A deep learning investigation of zonal jets

Ibrahim, Nurul Bin (2024) A deep learning investigation of zonal jets. Memorial University of Newfoundland. (Unpublished)

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (7MB)

Abstract

Zonal jets play a crucial role in influencing the dynamics and transport properties of planetary atmospheres and oceans. Despite their importance, the precise mechanisms governing their formation, equilibration, and maintenance remain a topic of ongoing investigation. This thesis aims to leverage deep learning techniques to gain new insights about zonal jets and develop efficient predictive models. We focus on applying these techniques to a low-order model (LOM) of the stochastically excited two-dimensional Boussinesq system, which can capture the essential interactions between the zonal jet and the associated turbulence. Various deep learning architectures like feed-forward neural networks, statistics-informed neural networks, and physics-informed neural networks are developed and trained on large ensembles of LOM simulations to analyze the dynamics of zonal jets. We demonstrate that such deep learning models can capture the underlying statistical properties and predict the long-term behavior of these jets. The capabilities of the trained models are also explored by extracting physical insights based on the analysis. For example, we can use our neural network to understand the physical mechanism behind jet maintenance. These findings have implications for improving our understanding and ability to predict geophysical turbulence. We also compare our deep learning results with the results of the statistical state dynamics (SSD) theory of turbulence. This comparison is used to test the correspondence of SSD with the real behaviour of the stochastic system and identify areas where it could be improved.

Actions (login required)

View Item