作者:石竞琛，刘霏凝，王文杰，赵瑞

Authors:SHI Jingchen,LIU Feining,WANG Wenjie,ZHAO Rui
摘要:针对目前大多数机器学习模型预测材料性质时需要大量的先验知识以及特征向量筛选困难的问题, 基于电子轨道矩阵和元素周期表法两种描述符,通过特征融合的方式,设计了一种卷积神经网络模型 OPCNN(Or- bital of electron and Periodic table CNN) 。实验数据表明 ,OPCNN 与其他预测模型相比,在带隙、生成热以及形成能 数据集上都有着更好的性能 ,平均绝对误差分别为 0. 26 eV、0. 037 KJ/mol 和 0. 073 eV/atom,且 R2 都达到了 91% 以 上 。OPCNN 在保证了预测准确性的同时对先验知识的要求更低,只需要元素周期表中的信息即可预测材料性质, 特征融合的思想可以让特征设计更加灵活,有利于新材料体系快速和准确的预测。
Abstract:Aiming at the problem that most machine learning models need a lot of prior knowledge and manual selection of feature vectors in the prediction of material properties, a convolutional neural network model OPCNN (Orbital of Electron and Periodic table CNN) is established by feature fusion based on two descriptors, electronic orbit matrix and periodic table method. The experimental data show that compared with other prediction models, OPCNN has better performance on the bandgap, heat of formation and formation energy datasets, with the mean absolute error of 0. 26 eV, 0. 037 KJ/mol and 0. 073 eV/atom, respectively, and the R2 is more than 91% . OPCNN has lower requirements for prior knowledge while ensuring the accuracy of prediction. It only needs the information in the periodic table to predict the material properties. The idea of feature fusion can make feature design more flexible, which is conducive to the rapid and accurate prediction of new material systems.

PDF全文下载地址:

可免费Download/下载PDF全文