Heterogeneous Multi-objective Optimisation in Nuclear Power Design

Abstract

This thesis researches the deleterious effects of heterogeneous optimisation within nuclear power design. Heterogeneous optimisation is investigated in two forms: delaying a subset of objectives, and delaying a higher fidelity search. The large computational expense and extensive runtimes of nuclear power design optimisation demonstrate considerable scope for heterogeneous optimisation and the tools to accommodate it. By taking a stepped-back view of multi-objective optimisation, Chapter 1 outlines the nuclear power design optimisation background upon which this thesis develops. Chapter 2 goes on to develop the due diligence necessary to retain trust in the computational optimisation process and its achieved optima, acting as scaffolding to support the methods adopted throughout this thesis. Chapter 3 investigates the time-saving benefits of delaying a subset of objectives which can bottleneck computation times. By optimising a series of objectives for a simple nuclear fuel design, we show how one can bias subsequent sub-problems due to favouring of the initially optimised sub-problems. We go on to develop a novel method to overcome these biases whilst benefitting from accelerated computation times. Chapter 4 investigates the time-saving benefits of delaying a higher fidelity optimisation of a system. By ignoring aspects of reality for nuclear fuel design optimisation, we show how one can distort the objective space and any optima therein such that they are no longer representative of a real-world problem. We go on to develop a novel method to accommodate this faster optimisation method and recover from any resultant distortion to the objective space. Finally, Chapter 5 concludes this thesis by summarising the main findings of this research before suggesting future avenues of work. We close by discussing the implications of this research both within the field of nuclear power design optimisation and within the wider considerations of computational decision-making. As computational decision-making becomes increasingly involved in today's processes, it is increasingly incumbent on decision-makers to verify that the true problem is well-represented and without bias.