Student thesis: Phd


This thesis details the development of machine learning techniques for the benefit of safety analysis in the UK nuclear energy sector. The objective of this research was to develop machine learning models which can perform the same functionality as existing industry standard models but with reduced computational intensity. A key research objective was the production of surrogate machine learning models (SMLM). These models are produced with the intention of retaining the functional- ity of an existing model i.e. it produces similar outputs from the same inputs whilst reducing computational cost. A framework was developed for efficient generation and manipulation of data for SMLM development. The framework was then expanded to allow data exploration, visualisation, optimisation and evaluation of SMLMs. Through a visualisation and analysis of the data space, as well as a set of preliminary experiments, a prudent direc- tion for the research was determined. An experimental approach was taken to the development and optimisation of an SMLM for the purposes of this research. The experimental process involved the exploration of numerous configurations, model architectures and techniques. A key finding was the importance of retaining physical relationships within the input features and exploiting these relationships with models such as convolutional neural networks. Another key observation was the relative difficulty in producing a model capable of making accurate predictions at the extremes of the data space i.e. the very high and low points of the prediction continuum. This was noted to be likely caused by dataset bias. The direction of the research then moved towards improving the performance of the models produced during the first phase of the research. To this end, three existing machine learning techniques were adapted to the problem space. These techniques were: data augmentation, the development of a custom loss function and fine-tuning of pre-trained machine learning models. The first two of these techniques were found to improve model performance individually, but it was also found that they were best applied as part of a fine-tuning process of existing models.
Date of Award1 Aug 2024
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorGavin Brown (Supervisor) & Tingting Mu (Supervisor)


  • transfer learning
  • custom loss functions
  • data augmentation
  • convolutional neural networks
  • neural networks
  • surrogate models
  • machine learning
  • Nuclear
  • seismic models

Cite this