Precisely Predicting Neutronics Parameters of Nuclear Reactor

Minxiao Zhong, Yuqi Feng, Qing Li, Yanan Sun

Abstract: The prediction of neutronics parameters is an indispensable step in designing the nuclear reactor core, but the methods for prediction have long been hindered by the tradeoff of efficiency and precision. During past years, many researchers have proposed machine learning models to improve the efficiency. However, there is lack of precision, because the models cannot precisely understand the complex non-linear relationship among various physics attributes. To address this problem, we propose a vertical thermal coding approach. It vertically classifies and stacks a variety of physics attributes in the reactor, and reduces the mutual interference of different physics information in the same plane. Moreover, we also propose an end-to-end convolutional neural network (CNN) model, along with a flexible combination of feature extraction blocks, to represent the reactor scheme more accurately. The experimental results demonstrate our CNN achieves superior precision across three typical neutronics datasets, i.e., critical boron concentration, power peaking factor, and nuclear enthalpy rise factor. In addition, ablation studies demonstrate that the proposed encoding strategy significantly enhances the representation of a reactor scheme, comparing with other encoding methods in literature. The supplementary materials and source code are available at https://github.com/ArmorCVU/Prediction.