基于数据增广的小样本材料属性预测  

Prediction of small sample material properties based on data augmentation

作　　者：李丹 董赞强[2] 杨涛[1] 佟强 谭胖 刘秀磊[1,3] LI Dan;DONG Zanqiang;YANG Tao;TONG Qiang;TAN Pang;LIU Xiulei(Beijing Advanced Innovation Center for Materials Genome Engineering,Beijing Information Science&Technology University,Beijing 100101,China;School of Intelligent Engineering,Zhengzhou University of Aeronautics,Zhengzhou 450046,China;Laboratory of Data Science and Information Studies,Beijing Information Science&Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学北京材料基因工程高精尖创新中心,北京100101 [2]郑州航空工业管理学院智能工程学院,郑州450046 [3]北京信息科技大学数据科学与情报分析研究所,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2023年第5期39-44,共6页Journal of Beijing Information Science and Technology University

基　　金：国家重点研发计划(2021YFB2600600);促进高校分类发展-重点研究培育项目(2121YJPY225);河南省科技攻关(高新技术)项目(182102210106)。

摘　　要：如何提升数据可用性是机器学习应用在材料属性预测任务中的一大挑战。对此,提出了一种基于数据增广的小样本材料属性预测方法。针对公开发布的Matbench屈服强度属性预测任务,仅使用312个数据样本,应用全连接神经网络对属性进行预测。首先,去除降低模型性能的元素,提高数据质量;然后,使用线性插值扩充数据集;最后,对神经网络进行建模和训练,预测屈服强度属性。相比于最优基线模型MODNet,该方法的平均绝对误差降低了4.97%。How to improve data availability is a major challenge for machine learning applications in material property prediction tasks.In this regard,a small sample material property prediction method based on data augmentation was proposed.For the publicly released Matbench yield strength property prediction task,only 312 data samples were used by applying a fully connected neural network to predict the property.Firstly,elements that reduce model performance were removed to improve data quality.Then,data augmentation was used to expand the dataset.Finally,the neural network was modeled and trained to predict yield strength property.Compared with the optimal baseline MODNet,the mean absolute error of the proposed method is reduced by 4.97%.

关 键 词：机器学习 属性预测 屈服强度 数据增广 神经网络 

分 类 号：TP304[自动化与计算机技术—计算机系统结构]