Applicability study of Bayesian optimization in core neutronic design using a toy model

Kuwagaki, Kazuki  ; Yokoyama, Kenji   

In Japan Atomic Energy Agency (JAEA), an innovative design approach named ARKADIA (Advanced Reactor Knowledge- and AI-aided Design Integration Approach through the whole plant lifecycle) for the advanced nuclear reactors is currently under development. One of the tasks in ARKADIA is to build a system that automatically optimizes core and fuel designs by conducting core neutronic and thermal-hydraulic calculations, fuel integrity evaluations, and plant dynamic analyses. This system will be implemented to automatically find an optimal design that minimizes (or maximizes) objective function defined by a core performance, while varying the core and fuel design parameters such as fuel pin diameter, core height and diameter. In this study, as the first step of the system development, we focused only on core neutronic design and conducted a study of automatic optimization. As the optimization algorithm, Bayesian optimization (BO), which is an effective method for optimization problems with expensive computation cost of objective function, was selected. The applicability of BO was studied based on single- and two-objective optimization examples of core neutronic design in a toy model. As a result, in the former, it was confirmed that BO can obtain the optimal solution, which well matches the reference solution calculated by a brute force calculation, with a small number of required calculation executions. Its usability on core neutronic designs, where the computation cost per case is large, was confirmed. In the latter, it was shown that BO can obtain a pareto solutions-set that shows good agreement with the reference solution.