机器学习力场中截断半径的选择  

Selection of the cutoff radius in machine learning force fields

作　　者：林旭敏 袁子刚 武海斌[2] 刘端阳[2] LIN Xumin;YUAN Zigang;U Haibin;LIU Duanyang(College of Mathematics and Physics,Beijing University of Chemical Technology,Beijing 100029,China;State Key Laboratory of Superlattices and Microstructures,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China)

机构地区：[1]北京化工大学数理学院,北京100029 [2]中国科学院半导体研究所半导体超晶格国家重点实验室,北京100083

出　　处：《江苏海洋大学学报(自然科学版)》2025年第1期41-50,共10页Journal of Jiangsu Ocean University:Natural Science Edition

基　　金：中国科学院稳定支持基础研究领域青年团队计划项目(YSBR-053);中国科学院网络安全和信息化专项应用示范项目(CAS-WX2023PY-0101)。

摘　　要：基于从头算分子动力学的机器学习方法构建力场已成为一个高度活跃的研究领域,然而目前对机器学习力场中截断半径的选择并没有一个可供参考的依据。利用83种不同材料的1 600多种结构构型进行从头算分子动力学模拟,分析近邻原子移动前后作用在中心原子上的力的变化,从而确定机器学习力场中截断半径的选择。当近邻原子与中心原子的距离在7.5Å或10Å(1Å=0.1 nm)时,力的误差通常在10^(-3)eV/Å内;当原子间距为5Å时,力的误差会增加到10^(-2)eV/Å。故在机器学习力场中设定5Å的截断半径值就能够满足精度要求。此外,对于原子晶体来说,截断半径值可略大于5Å;而对于分子晶体来说,截断半径值可小于5Å;就金属晶体、离子晶体和合金而言,截断半径值为5Å即可。该研究首次补充了机器学习力场中截断半径相关的理论知识,将有利于机器学习力场的发展。Constructing a force field by machine learning method based on ab initio molecule dynamics has become a highly active research area,however,there is currently no reference for the selection of cutoff radius.The ab initio molecular dynamics simulation with over 1600 structural configurations across 83 different materials was carried out to analyze the variation in forces acting on the central atom before and after displacing a neighbor atom,in order to determine the selection of the cutoff radius in machine learning force fields.The force difference typically falls within the range of 10^(-3) eV/Å when the distance between the neighbor atom and the central atom was 7.5 Å or 10 Å.At a distance of 5 Å,this force difference increases to about 10^(-2) eV/Å.Therefore,setting a cutoff radius value of 5 Å in the machine learning force fields can meet the accuracy requirements.Moreover,for atomic crystals,a slightly larger cutoff radius value than 5 Å may be suitable,while for molecular crystals,a cutoff radius value less than 5 Å could be appropriate.In the case of metal crystals,ionic crystals and alloys,the value of 5 Å is deemed suitable.This is the first time to supplement the theoretical knowledge of the cutoff radius in the machine learning force fields,which will be conducive to the development of the machine learning force fields.

关 键 词：机器学习力场 从头算分子动力学 截断半径 力 

分 类 号：O469[理学—凝聚态物理]