基于神经网络和决策树算法的裂变产物核(n,2n)反应截面研究  

Study of(n,2n)Reaction Cross Section of Fission Product Based on Neural Network and Decision Tree Model

作　　者：孙小东 韦子豪 王端 续瑞瑞 田源 陶曦 张英逊[1] 张玥 张智 葛智刚[1] 王记民 王俊辰 夏候琼 舒能川[1] SUN Xiaodong;WEI Zihao;WANG Duan;XU Ruirui;TIAN Yuan;TAO Xi;ZHANG Yingxun;ZHANG Yue;ZHANG Zhi;GE Zhigang;WANG Jimin;WANG Junchen;XIA Houqiong;SHU Nengchuan(Department of Nuclear Physics,China Institute of Atomic Energy,Beijing 102413,China;Reactor Engineering Technology Research Institute,China Institute of Atomic Energy,Beijing 102413,China;Graduate Department of Nuclear Industry,Beijing 102413,China)

机构地区：[1]中国原子能科学研究院核物理研究所,北京102413 [2]中国原子能科学研究院反应堆工程技术研究所,北京102413 [3]核工业研究生部,北京102413

出　　处：《原子能科学技术》2023年第4期798-804,共7页Atomic Energy Science and Technology

基　　金：国家自然科学基金(12005303,1187050492);中核集团青年科技创新团队项目。

摘　　要：为了大规模预言缺少实验测量的裂变产物核反应截面数据,在整理现有(n, 2n)反应截面5 294个实验数据的基础上,分析相关的物理特征建立实验数据集,分别构建和训练反向传播神经网络和极致梯度提升树模型学习数据。神经网络的隐藏层包含两个子网络,分别由2层各128个神经元构成。极致梯度提升树模型集成了16棵决策树。结果表明,虽然(n, 2n)反应截面实验测量数据大多集中分布在中子入射能量14 MeV附近,且相互之间存在分歧,本工作机器学习模型均可较好描述反应截面的实验测量数据,具有较好的预言能力,对于缺少实验数据的情况同样与评价数据库基本符合。人工神经网络模型在测试集中预测结果与实验数据平均相对偏差小于10%的数据占比超过85%。机器学习方法能为核数据评价研究提供参考。The neutron induced nuclear reaction cross section of fission products is related with the neutron flux and the reactor burnup,which plays an important role in accurately designing nuclear engineering.To predict(n,2n)[JP2]reaction cross sections especially those without experimental data,the relevant features were analyzed and the experimental data set were established on the basis of sorting out the experimental data recorded in EXFOR library.This work includes 5294(n,2n)cross section measured results,among which a lot of experimental data concentrate around 14 MeV incident energy.Moreover,there may be divergence between measurements due to system error and negligence error.Faced with these real and defective data,researchers discovered laws behind them and established the compound nucleus reaction models.However this means a heavy workload.It would be surprising if machine learning could reach a quantitative level close to the evaluation results.The 8 features include the proton number Z,the mass number A,the single nucleon separation energy of both proton and neutron,the Casten factor,the level density,the pairing correction,and the incident energy.The back propagation artificial neural network(ANN)and decision tree(DT)models were built to learn the experimental data set,respectively,adopting PyTorch and XGBOOST toolboxes.Draw lessons from the variational auto-encoder network,the 2 sub-networks with the same internal structure,which contains 128 neurons in 2 hidden layers,were designed to learn the mean and variance respectively.The boosting model integrates 16 decision trees.The training set includes 4000 uniform and randomly selected data,while the remaining data constitute the test set.The results show that both ANN and XGBOOST models describe the experimental cross section data well,moreover model gives a smooth and continuous curve,indicating a certain predictive ability.For the case of lack of experimental data,the predictions are also basically consistent with the evaluation nuclear reaction data librar

关 键 词：反向传播神经网络 极致梯度提升树 (n 2n)反应截面 核数据 

分 类 号：O571.55[理学—粒子物理与原子核物理]