Doctoral Dissertations

Date of Award


Degree Type


Degree Name

Doctor of Philosophy


Nuclear Engineering

Major Professor

John W. Hines

Committee Members

Vlad Sobes, Ondrej Chvala, Germina Ilas


The design and optimization of nuclear systems can be a difficult task, often with prohibitively large design spaces, as well as both competing and complex objectives and constraints. When faced with such an optimization, the task of designing an algorithm for this optimization falls to engineers who must apply engineering knowledge and experience to reduce the scope of the optimization to a manageable size. When sufficient computational resources are available, unsupervised optimization can be used.

The optimization of the Fast Neutron Source (FNS) at the University of Tennessee is presented as an example for the methodologies developed in this work. The FNS will be a platform for subcritical nuclear experiments that will reduce specific nuclear data uncertainties of next-generation reactor designs. It features a coupled fast-thermal design with interchangeable components around an experimental volume where a neutron spectrum, derived from a next-generation reactor design, will be produced.

Two complete genetic algorithm optimizations of an FNS experiment targeting a sodium fast reactor neutron spectrum are presented. The first optimization is a standard implementation of a genetic algorithm. The second utilizes neural network based surrogate models to produce better FNS designs. In this second optimization, the surrogate models are trained during the execution of the algorithm and gradually learn to replace the expensive objective functions. The second optimization outperformed by increasing the total neutron flux 24\%, increased the maximum similarity of the neutron flux spectrum, as measured by representativity, from 0.978 to 0.995 and producing configurations which were more sensitive to material insertions by +124 pcm and -217 pcm. In addition to the genetic algorithm optimizations, a second optimization methodology using directly calculated derivatives is presented.

The methods explored in this work show how complex nuclear systems can be optimized using both gradient informed and uninformed methods. These methods are augmented using both neural network surrogate models and directly calculated derivatives, which allow for better optimization outcomes. These methods are applied to the optimization of several variations of FNS experiments and are shown to produce a more robust suite of potential designs given similar computational resources.


Microsoft word reports the word count of the abstract as being 344 words, which is below the 350 word limit.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."