Bayesian neural network (BNN) machine learning (ML) surrogate modelling methods for simulating turbulent flow within nuclear power plant (NPP) t-junction pipes
File(s)
Author(s)
Yew Hoe, Wong
Type
Thesis
Abstract
The integration of nuclear power plants (NPPs) into modern smart grids poses new challenges, particularly in ensuring component reliability under increasingly dynamic operating conditions. One area of concern is thermal fatigue in T-junctions within the primary circuit, which is a critical issue for nuclear safety that remains poorly addressed by current modelling practices. Existing approaches often focus on isolated components or employ low-fidelity models, limiting their ability to capture the complex flow-induced thermal stresses that drive thermal fatigue. This research addresses these limitations through a three-part investigation. First, mesh discretisation error within the computational models are thoroughly studies and estimated. This provides significant insights into how model accuracy can be improved, while concurrently providing model uncertainty estimates. The second would be developing a high-fidelity simulation framework to model turbulent flow in various T-junction configurations. Using Power Spectral Density (PSD), Proper Orthogonal Decomposition (POD), and Spectral POD (SPOD), the study identifies dominant flow structures responsible for thermal fatigue. These analyses provide detailed insights into unsteady flow behaviours that are often overlooked in traditional models. Results reveal underlying flow structures within turbulent flow that potentially contributes to thermal fatigue phenomena. Third, machine learning techniques are explored to recover turbulent flow features in these geometries. Several neural network architectures such as deep neural networks (DNNs), physics-informed neural networks (PINNs), Bayesian neural networks (BNNs), and Bayesian PINNs (BPINNs), are evaluated as surrogate models for flow field reconstruction. Results show that BNNs achieve the best performance, providing both accurate predictions and quantified uncertainty. This positions it as a valuable component in the broader effort to enhance predictive capabilities in nuclear thermal fatigue analyses.
Version
Open Access
Date Issued
2025-05-13
Date Awarded
01/10/2025
License URL
Advisor
Bluck, Michael
Eaton, Matthew
Lampunio, Lisa
Duan, Yu
Sponsor
Singapore Nuclear Research and Safety Institute
Publisher Department
Department of Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)