An Iterative Optimization Method Using Genetic Algorithms and Gaussian Process Based Regression in Nuclear Reactor Design Applications

Abstract

The optimization of a complex system involves the determination of optimum values for a set of design parameters. The optimization search happens in order to meet a specific set of objectives concerning the quantities of interest (QOI). Also, the design parameters are a subset of the input parameters and the QOIs are determined from the output parameters. Particularly, when the parameter space is large, optimization necessitates a significant number of executions of the simulator to obtain a desired solution in tolerance limits. When the simulations are expensive in terms of computation time, an emulator based on regression methods is useful for predictions. This work presents a novel methodology that uses an iterative hybrid global optimization method (GOM) using genetic algorithms (GA) and simulated annealing (SA) model coupled (HYBGASA) with a Gaussian process regression method based emulator (GPMEM) to optimize a set of input parameters based on a set of defined objectives in a nuclear reactor power system. Hereafter this iterative hybrid method comprising of HYBGASA and GPMEM would be called as the “IHGOM". In addition to optimization, IHGOM iteratively updates the trial data obtained from the neighborhood of the near optimal solution, used to train the GPMEM in order to reduce regression errors. The objective is to develop, model and analyze IHGOM, and apply it to an optimization problem in the design of a nuclear reactor. Development and analysis of IHGOM and its implementation in a nuclear reactor power system problem is a significant contribution to the optimization and the nuclear engineering communities.