基于卷积神经网络模型中梯度与特征分析的材料性能优化与预测机理研究-中国材料进展

文章信息/Info

Title:: Research on Optimization and Prediction Mechanism of Material Properties Based on Gradient and Feature Analysis in Convolution Neural Network

作者:: 曹卓1; 但雅波1; 李想1; 牛程程2; 董容智1; 钱松荣1; 胡建军1; 3; （1. 贵州大学机械工程学院，贵州贵阳 550025） (2. 贵州大学现代制造技术教育部重点实验室，贵州贵阳 550025) (3. 美国南卡罗来纳大学计算机科学与工程系，南卡罗莱纳哥伦比亚 29208）

Author(s):: CAO Zhuo1; DAN Yabo1; LI Xiang1; NIU Chengcheng2; DONG Rongzhi1; QIAN Songrong1; HU Jianjun1; 3; (1. School of Mechanical Engineering, Guizhou University, Guiyang 550025, China) (2. Key Laboratory of Advanced Manufacturing Technology, Ministry of Education, Guizhou University, Guiyang 550025, China) (3. Department of Computer Science and Engineering, University of South Carolina, Columbia SC 29208，USA)

Keywords:: materials informatics; convolutional neural network; formation energy; gradient analysis; feature extraction

摘要:: 材料信息学作为材料领域一种新的研究方法，引起了国内外广泛的关注。随着材料数据的快速增加，机器学习方法也越来越多地被应用在材料数据的分析中，并有望从大量的材料数据中获取具有指导性的材料学规律。采用卷积神经网络模型，使用从材料数据库中收集得到的4000多种材料的数据，对材料的形成能进行预测并得到了较为准确的预测结果。随后对材料特征矩阵的梯度进行分析，发现了梯度与材料性能间有一定的相关性，并可在梯度矩阵的指导下找到具有目标性能的材料特征矩阵分布。最后对卷积神经网络中识别出的特征模式进行了分析，进一步验证了卷积神经网络具有较好的材料性能预测能力。

Abstract:: As a new research mode in material science, material informatics has attracted wide attention. With the rapid increase of material data, machine learning methods are more and more used in the analysis of material data to obtain instructive physical and chemical laws from a large number of material data. This paper focuses on the convolutional neural network, using data from more than 4000 materials collected from the Material Project database to predict formation energy of materials, and the prediction results are accurate. Then, the gradient of feature map is analyzed, we observe that there are some certain correlations between gradient and material properties, and under the guidance of gradient matrix, the possible distribution of feature map with target properties can be found. Finally, the patterns recognized by the convolutional neural network are analyzed, which further verifies that the convolutional neural network can achieve excellent prediction results of material property.