基于卷积神经网络的衍射图空间群识别研究  

Study on diffraction pattern space group recognition based on convolutional neural network

作　　者：石竞琛 王文杰 刘霏凝 赵瑞[1] SHI Jing-chen;WANG Wen-jie;LIU Fei-ning;ZHAO Rui(College of Mathematics and Computer,Jilin Normal University,Siping 136000,China)

机构地区：[1]吉林师范大学数学与计算机学院,吉林四平136000

出　　处：《白城师范学院学报》2023年第2期7-13,共7页Journal of Baicheng Normal University

基　　金：吉林省科技发展计划项目(YDZJ202201ZYTS629).

摘　　要：X射线衍射(XRD)图谱数据的采集和分析是新材料开发周期中必不可少的步骤之一,常规实验表征很难实现大批量的测试和快速鉴别.文章基于DenseNet设计了一个衍射图空间群识别的神经网络模型SE-DenseNet.SE-Dense Net在简化了网络结构的同时,通过增加注意力机制(Squeeze and Excitation,SE),并采用新的激活函数来提高网络模型的性能.研究表明,在具有32337个样本包含20类空间群的数据集上,SE-Dense Net的准确率为81.73%,较基础对照模型提高了4.9%.研究发现,尽管数据集的不平衡性是限制神经网络模型预测准确度的主要原因之一,但SE-DenseNet的性能足以在短时间对大量衍射图数据产生准确的预测,并提供有意义的参考.The acquisition and analysis of X-ray diffraction(XRD)pattern data is one of the essential steps in the development cycle of new materials,and conventional experimental characterisation makes it dif-ficult to test and rapidly identify in large quantities.This article designs a neural network model SE-DenseNet for space group recognition of diffraction patterns based on the DenseNet.While simplifying the network struc-ture,SE-DenseNet improves the performance of the network model by adding an attention mechanism(Squeeze and Excitation,SE)and employing a new activation function.The study showed that on a dataset with 32237 samples containing 20 space groups,SE-DenseNet was 81.73%accurate,improving by 4.9%over the baseline model.It was found that although the unbalanced nature of the dataset was one of the main reasons limiting the prediction accuracy of the neural network model,SE-DenseNet performed well enough to produce accurate pre-dictions for large amounts of diffraction pattern data in a short time and to provide a meaningful references.

关 键 词：卷积神经网络 SE-DenseNet X射线衍射图 空间群 识别 

分 类 号：O436.1[机械工程—光学工程]