贝叶斯机器学习对裂变产额的不确定度评价  

Bayesian machine learning for the uncertainty evaluation of nuclear fission product yields

作　　者：易佳怡 乔春源 裴俊琛[2] 王子澳 陈永静[3] 舒能川[3] 葛智刚[3] 许甫荣[2] YI JiaYi;QIAO Chun Yuan;PEI JunChen;WANG ZiAo;CHEN YongJing;SHU NengChuan;GE ZhiGang;XU FuRong(Yuanpei College,Peking University1,Beijing 100871,China;State Key Laboratory of Nuclear Physics and Technology1,School of Physics,Peking University,Beijing 100871,China;China Nuclear Data Center,China Institute of Atomic Energy,Beijing 102413,China)

机构地区：[1]北京大学元培学院,北京100871 [2]北京大学物理学院,核物理与核技术国家重点实验室,北京100871 [3]中国原子能科学研究院,中国核数据中心,北京102413

出　　处：《中国科学：物理学、力学、天文学》2022年第5期119-125,共7页Scientia Sinica Physica,Mechanica & Astronomica

基　　金：国家自然科学基金(编号:11975032,11835001,11790320,11790323,11790325,11961141003);国家重点基础研究发展计划(编号:2018YFA0404403)资助项目。

摘　　要：先进核能研究需要更高精度的核裂变数据的支撑,但核裂变的微观理论描述和实验测量都非常困难.目前部分裂变实验数据存在精度不足、不完整、有分歧等问题.本文基于贝叶斯机器学习建立了对不够精确的核裂变产额分布的评价方法,以中子诱发^(235)U裂变为例,细致研究了误差传递对核裂变产额评价的影响.结果表明误差传递是一个综合效应,与临近的数据信息有关,这也说明机器学习包含了复杂的数据关联.核数据的不确定度评价对核装置的安全设计非常关键,也是进一步进行数据融合评价的必要基础.Providing highly accurate nuclear fission data is critical for next-generation nuclear energy production.However,both experimental measurement and reliable modeling of nuclear fission are very difficult.A wealth of raw experimental fission data exists,but these data are generally noisy,incomplete,and inconsistent.This work investigates the evaluation of imperfect fission yields based on Bayesian machine learning.As an illustrative example,the uncertainty propagation in the yield evaluation of neutron-induced fission of ^(235)U was studied in detail.The results show that uncertainty propagation is a combined effect closely related to knowledge of neighboring data.This also means that machine learning can take into account complex correlations in the data.Uncertainty quantification study is essential for the safety design of nuclear applications,and it is also a necessary step for data fusion evaluation.

关 键 词：核裂变 机器学习 裂变产额 误差传递 

分 类 号：O571.43[理学—粒子物理与原子核物理] TP181[理学—物理]