Machine learning-augmented turbulence models for the simulation of two-phase shear flows

Bertolotti, Luc (2022) Machine learning-augmented turbulence models for the simulation of two-phase shear flows. PhD thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (6MB)


This dissertation focuses on the investigation of co-current two-phase stratified gas-liquid shearing flows with a sharp interface for industrial applications and more specifically for the study of flows present in aero-engine bearing chambers. Predicting the behaviour of shear flows in the different parts of an aero-engine, such as bearing chambers, is crucial. As a matter of fact, in the context of the thermal management of the bearing chamber, methodologies to predict the oil film thickness distribution must be investigated in order to optimise the system lubrication and prevent from oil coking or degradation. The thickness distribution of wavy films is largely analysed in the industry, but turbulent two-phase flows remain very challenging to predict depending on the case. Carrying out experiments is often too expensive and computational fluid dynamics (CFD) still struggles with two-phase flow prediction especially when it involves the modelling of a sharp interface. Many CFD methods are employed to predict the oil film thickness distribution and interface velocity. However, the vastly used standard Reynolds-averaged Navier-Stokes equations (RANS) turbulence models are derived from semi-empirical methods of turbulence damping, which are inaccurate for wavy films with high gradients of velocity across the interface, thus impacting flow modelling in bearing chambers.

With the objective of improving the current RANS models from scale resolving methods, high-fidelity simulations were carried out in this dissertation using quasi-direct numerical simulation (qDNS) with OpenFOAM. The results of various high-fidelity simulation cases were employed to inform the interfacial turbulence in the widely used standard Wilcox’s RANS k−ω turbulence model. Two flow configurations based on experimental works exploring stratified flows in horizontal channels were investigated, namely the thick-film and thin-film configurations. Simplified 3D periodic versions of the channels were designed. Channels were filled with two distinct gaseous and liquid phases. Shearing flows and interface waviness were triggered by the strong interfacial velocity gradients, as the gaseous phase velocity was set at much larger values than the liquid velocity. Numerical results were compared with experiments in terms of mean velocity and turbulent energy and Reynolds stress.

First, a preliminary study was conducted in order to choose the most optimal methods for the investigated two-phase flows. Part of this preliminary work involved the domain simplification by an analysis of the fluctuation velocity auto-correlations in the periodic directions of the computational domain. Secondly, a methodology of a proof of concept was established and performed in order to demonstrate the RANS k−ω turbulence model capability to be driven by high-fidelity data for the prediction of accurate behaviours of two-phase shearing flows, in the thick-film configuration. Finally, a high-fidelity data-driven neural network was implemented in machine learning models with PyTorch and coupled with the Wilcox’s model through python scripts in order to inform and improve the interfacial turbulence of RANS simulations. This was accomplished by following the methodology of the proof of concept on new cases, in the thin film configuration. While the type of corrections predicted by the implemented machine learning models presented in this dissertation applies to the specific Wilcox’s turbulence model, it was proven that machine learning can be used effectively to assist and enhance CFD abilities to predict two-phase stratified flows with averaged models without increasing computational costs significantly as opposite to high-fidelity simulations or hybrid models. In the context of industrial studies, one can imagine the future development of new CFD methods involving couplings between averaged turbulence models and machine learning models for quick analyses and accurate results.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Jefferson-Loveday, Richard
Ambrose, Stephen
Keywords: Aero-engine, Bearing chamber, Two-phase flow, Computational Fluid Dynamics, Quasi-Direct Numerical Simulation, Interfacial Turbulence, Neureal Network, Machine Learning
Subjects: T Technology > TA Engineering (General). Civil engineering (General) > TA 357 Fluid mechanics
Faculties/Schools: UK Campuses > Faculty of Engineering
Item ID: 71515
Depositing User: Bertolotti, Luc
Date Deposited: 31 Dec 2022 04:40
Last Modified: 31 Dec 2022 04:40

Actions (Archive Staff Only)

Edit View Edit View