基于深度卷积神经网络的α-Fe晶界能预测  

Prediction of grain boundary energy of α-Fe based on deep convolutional neural network

作　　者：陈村 李六六 彭蕾 时靖谊[1,2] CHEN Cun;LI Liu-Liu;PENG Lei;SHI Jing-Yi(State Key Laboratory of Particle Detection and Electronics,University of Science and Technology of China,Hefei 230026,China;School of Nuclear Science and Technology,University of Science and Technology of China,Hefei 230027,China)

机构地区：[1]中国科学技术大学核探测与核电子学国家重点实验室,合肥230026 [2]中国科学技术大学核科学技术学院,合肥230027

出　　处：《原子与分子物理学报》2022年第3期126-134,共9页Journal of Atomic and Molecular Physics

基　　金：国家自然科学基金(11805131)。

摘　　要：本文提出了一种预测晶界能以研究α-Fe晶界性质的深度学习方法.在分子动力学生成的α-Fe对称倾斜晶界的基础上,通过中心对称参数和原子密度信息构造出晶界特征—积累中心对称参数,提出了数据增强和按倾斜角分层抽样的方法,建立了预测晶界能的卷积神经网络模型.测试集结果表明,预测晶界能的平均相对误差小于1.75%,平均每个晶界的预测用时在0.002 s以内.该方法在一定范围内具有较高的准确性和鲁棒性,提供了研究晶界的微观结构特征与宏观性能之间关联的途径.A deep learning method for predicting grain boundary energy was proposed to study the grain boundary properties of α-Fe in this paper. Based on the α-Fe symmetric tilt grain boundary generated by molecular dynamics, we constructed the accumulated centro-symmetry parameters by using the centro-symmetry parameters and atomic density information. The methods of data augmentation and stratified sampling according to the tilt angle were proposed to establish the convolutional neural network model which can predict the grain boundary energy. The test results show that the average relative error of grain boundary prediction is less than 1.75%, and the average prediction time for each grain boundary is about 0.002 s. In a certain range, this method can maintain high accuracy and robustness, and provides an approach to study the relationship between the grain boundary structural characteristics and properties.

关 键 词：晶界 卷积神经网络 中心对称参数 晶界能 

分 类 号：O483[理学—固体物理]