神经网络方法分析U同位素链裂变核反应截面  

Analyzing Fission Cross Section of U Isotope by Neural Network Method

作　　者：田源[1] 续瑞瑞[1] 陶曦[1] 王记民[1] 张玥 孙小东 张智 王俊辰 夏候琼 TIAN Yuan;XU Ruirui;TAO Xi;WANG Jimin;ZHANG Yue;SUN Xiaodong;ZHANG Zhi;WANG Junchen;XIA Houqiong(China Nuclear Data Center,Key Laboratory of Nuclear Data,China Institute of Atomic Energy,Beijing 102413,China)

机构地区：[1]中国原子能科学研究院核数据重点实验室、中国核数据中心,北京102413

出　　处：《原子能科学技术》2023年第4期805-811,共7页Atomic Energy Science and Technology

基　　金：国家自然科学基金(1187050492,12005303)。

摘　　要：本文利用前馈神经网络方法分析U同位素链实验测量裂变核反应截面数据。采用包含4个输入量、1个输出量和3层隐藏层的前馈神经网络,对U同位素链实验测量裂变截面数据进行训练,并利用贝叶斯算法对网络中的超参数进行优化,最终得到整个铀同位素链随入射中子能量变化的裂变截面数据。神经网络方法产生的裂变截面数据能很好地再现裂变截面的阶梯结构,与实验和评价数据的结果十分接近。Neutron-induced fission nuclear reaction data are critical for understanding nuclear physics,engineering,and technology.The accuracy of fission cross section of key fuel nuclides and sub-actinide nuclei is increasingly important with the development of new nuclear energy system concepts.Machine learning is a powerful tool for data analysis and modeling,and can be used to extract features from large amounts of data without requiring encoding for a specific task.Neural networks,a subset of machine learning,are particularly effective for mapping input to output for tasks with a sufficient number of target-valued data.In a recent study,researchers applied machine learning methods to analyze fission nuclear reaction cross section data.The team collected fission cross section data for the uranium isotope chain from the experimental nuclear reaction data(EXFOR)library and five sets of frequently used evaluation data from the evaluated nuclear data file(ENDF)library.The neutron-induced fission cross section data of 233-239 U in the energy range from 2 to 20 MeV were selected as the training data set.To obtain fission cross sections for a given neutron number,proton number,and corresponding incident neutron energy,the feedforward neural network(FNN)was trained using fission cross section data from experimental measurement of the uranium isotope chain to achieve maximum agreement with experimental and evaluation data.The results show that the fission cross section data generated by the machine learning method can reproduce the step structure of the fission cross section well and is very close to the results of experimental and evaluation data.This provides a research basis for large-scale analysis of fission cross section data of easy-to-fission nuclei.The application of machine learning to nuclear physics and engineering is still in its early stages,and there are several challenges that need to be addressed.One major challenge is the lack of high-quality data sets,which are essential for training and testing machine learn

关 键 词：裂变截面数据 前馈神经网络 超参数优化 

分 类 号：O571.432[理学—粒子物理与原子核物理]