Neutronic analysis and optimization of the advanced high temperature reactor fuel design using machine learning

Abstract

The Liquid Salt Cooled Reactor (LSCR) is a graphite moderated and liquid salt cooled reactor The LSCR is designed with the intention of providing superior economics and safety compared to the current commercial light water reactors. Due to estimated similar capital cost and a significantly higher thermal efficiency the LSCR is expected to have a lower levelized unit electricity cost despite a potentially much higher fuel cycle cost. The overall objective of this thesis is to develop a fuel design that minimizes the fuel cycle cost and facilitates that the LSCR is more economical than a typical LWR over a typical reactor lifetime. The approach to developing an optimal design is to take the existing baseline design and optimize over the most significant parameters identified within the design. Rather than trying to find a radically new geometric design the focus is to identify advanced sampling, modeling, and optimization techniques that are ideal for finding the LSCR fuel assembly design that minimizes the fuel cycle costs. Once the limiting factor is determined further changes in the design can be intelligently recommended. The fuel design features double heterogeneous geometry configured with two layers of TRISO fuel in a carbon matrix pressed on both sides of an amorphous carbon slab. The reactor neutronics is analyzed using CSAS6 and TRITON sequences of SCALE6.1 with the KENO Monte Carlo neutron transport code. Since the continuous energy depletion calculations are prohibitively expensive, one objective of this study is to implement a methodology to replace them with significantly faster multigroup calculations while preserving adequate accuracy. Impact of the double-heterogeneous geometry may be accounted for by calculating MCDancoff factors to be used for the multi-group approximation. Automated calculations are performed with the Least Squares regression method and artificial neural networks in order to generate functional representations of average temperatures, MCDancoff Factors, cycle lengths, and burnups as functions of the parameters over the design space. These surrogate models are used to allow quick sampling of the fuel design phase space thus avoiding major computational requirements and facilitating the use of optimization algorithms. Due to the concave non-linear multivariable design space, a global heuristic optimization algorithm is necessary. Branch calculations are performed to provide reactivity coefficient constraints on design optimization.