Machine learning approach towards predicting turbulent fluid flow using convolutional neural networks - Final thesis

Aug 5, 2026

Upload by Patrick

Using convolutional neural networks, this thesis proposes a novel framework to predict turbulent fluid flow through an array of obstacles. To capture key characteristics of turbulence, the study applies a convolutional autoencoder neural network to forecast the first ten POD modes of the flow. Results indicate strong predictive performance for the first two POD modes, with reduced accuracy for the remaining eight modes. Predicted POD modes are nevertheless accurate enough to reconstruct turbulent flow, preserving large-scale details from the original simulation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages178

Reading Time7 h 29 min

AccessFree to download

Abstract

Acknowledgement

Declaration

Machine learning approach towards predicting turbulent fluid flow using convolutional neural networks Vinh Vu June 9, 2023 Version: Final Machine learning approach towards predicting turbulent fluid flow using convolutional neural networks Vinh Vu The School of Mathematics and Statistics, and the Department of Mechanical Engineering Supervisors Robertus Erdelyi, Yi Li, Franck Nicolleau and Andrew Nowakowski Vinh Vu Machine learning approach towards predicting turbulent fluid flow using convolutional neural networks June 9, 2023 Supervisors: Robertus Erdelyi, Yi Li, Franck Nicolleau and Andrew Nowakowski University of Sheffield The School of Mathematics and Statistics, and the Department of Mechanical Engineering iv Abstract Using convolutional neural networks, we present a novel method for predicting turbulent fluid flow through an array of obstacles in this thesis. In recent years, machine learning has exploded in popularity due to its ability to create accurate datadriven models and the abundance of available data. In an attempt to understand the characteristics of turbulent fluid flow, we utilise a novel convolutional autoencoder neural network to predict the first ten POD modes of turbulent fluid flow. We find that the model is able to predict the first two POD modes well although and with less accuracy for the remaining eight POD modes. In addition, we find that the ML-predicted POD modes are accurate enough to be used to reconstruct turbulent flow that adequately captures the large-scale details of the original simulation. v Acknowledgement There are many people for who I am grateful for their support throughout my PhD life, to the extent that another thesis could be written about their support. I would first like to thank my Masters and first Mechanical Engineering PhD supervisor, Dr Franck Nicolleau, for initially guiding my route to research and encouraging me to pursue my interest in fluid mechanics and turbulence. My interest in using computational methods would not have started without his encouragement. I would also like to thank Dr Andrew Nowakowski for his support in my work. I also would like to thank Professor Robertus Erdelyi for his support. Aside from support and the occasional funny stories and jokes, he has helped me in obtaining the funding for this PhD, which led me to start my PhD life. Finally, I would like to give my greatest thanks to Dr Yi Li, who has helped and guided me throughout my research. Yi has been very kind and patient, and I have been lucky to have him as a supervisor who could introduce me to advanced fluid dynamics research and interesting methods which have aided in my rapid development. I wish I could still learn more from him. I would like to acknowledge both EPSRC and the Faculty of Science Doctoral Academy for their financial support. For this work, I had access to many different HPCs. I would like to acknowledge the IT Services at The University of Sheffield for the provision of services for High-Performance Computing. This work used the Cirrus UK National Tier-2 HPC Service at EPCC [http://www.cirrus.ac.uk](http://www.cirrus.ac.uk funded)[ funded](http://www.cirrus.ac.uk funded)[ ](http://www.cirrus.ac.uk funded)[by the University of Edinburgh and EPSRC](by the University of Edinburgh and EPSRC) ([EP/P020267/1](EP/P020267/1)). This project made use oftime on Tier 2 HPC facility JADE, funded by EPSRC (EP/P020275/1). This project made use oftime on Tier 2 HPC facility JADE2, funded by EPSRC (EP/T022205/1). I would like to thank all the PhD students on H-Floor in Hicks building for their kind friendship and support. Particularly, Hope, Poppy and Callum who has been with me since I joined the Maths Department. Finally, I would like to express my gratitude to my grandparents, parents, my sisters and my girlfriend. Without their support and encouragement over the past few years, it would be impossible for me to complete my study. vii D

FAQs

What problem does the thesis address?

The thesis targets prediction of turbulent fluid flow through an array of obstacles using machine learning models.

How does the proposed method represent the turbulent flow?

It uses a convolutional autoencoder to predict the first ten POD modes of turbulent flow.

How accurate are the predictions and what can they be used for?

The model predicts the first two POD modes well and the remaining eight with less accuracy, but the predicted POD modes can still reconstruct turbulent flow that captures large-scale details.

Aug 5, 2026

Upload by Patrick