Deep-coupling neural network and genetic algorithm based on Sobol-PR for reactor lightweight optimization

Abstract

We propose a deep-coupling neural network and genetic algorithm method based on Sobol-PR method for reactor lightweight shielding optimization. The Sobol method is first used to analyze the sensitivities between inputs and outputs of the neural network, and then these sensitivities are used to adjust the fitness function of the genetic algorithm dynamically. Meanwhile, two indicators (precision and recall rate) are introduced to facilitate the sample evaluation and selection, where the precision quantifies the prediction ability of the neural network, and the recall rate quantifies the optimization efficiency of the genetic algorithm. The deep coupling between the neural network and the genetic algorithm based on Sobol-PR method contributes to an integrated framework of “calculation-optimization-reconstruction-evaluation,” which is applied to the lightweight shielding design of a small helium-xenon-cooled reactor. It is found that the performance of the neural network and the genetic algorithm is improved, with the precision of the neural network reaches up to 99 % and the recall rate of the genetic algorithm reaches up to 84 %. Compared with the traditional method, the new method improves the ratio of ideal solutions by up to 3.8 times and the optimization depth by up to 3.2 times.