Prediction of band gap of transition metal sulfide with Janus structure by deep learning atomic feature representation method

Abstract

With the development of artificial intelligence, machine learning (ML) is more and more widely used in material computing. To apply ML to the prediction of material properties, the first thing to do is to obtain effective material feature representation. In this paper, an atomic feature representation method is used to study a low-dimensional, densely distributed atomic eigenvector, which is applied to the band gap prediction in material design. According to the types and numbers of atoms in the chemical formula of material, the Transformer Encoder is used as a model structure, and a large number of material chemical formula data are trained to extract the features of the training elements. Through the clustering analysis of the atomic feature vectors of the main group elements, it is found that the element features can be used to distinguish the element categories. The Principal Component Analysis of the atomic eigenvector of the main group element shows that the projection of the atomic eigenvector on the first principal component reflects the outermost electron number corresponding to the element. It illustrates the effectiveness of atomic eigenvector extracted by using the transformer model. Subsequently, the atomic feature representation method is used to represent the material characteristics. Three ML methods named Random Forest (RF), Kernel Ridge Regression (KRR) and Support Vector Regression (SVR) are used to predict the band gap of the two-dimensional transition metal chalcogenide compound MXY (M represents transition metal, X and Y refer to the different chalcogenide elements) with Janus structure. The hyperparameters of ML model are determined by searching for parameters. To obtain stable results, the ML model is tested by 5-fold cross-validation. The results obtained from the three ML models show that the average absolute error of the prediction using atomic feature vectors based on deep learning is smaller than that obtained from the traditional Magpie method and the Atom2Vec method. For the atomic eigenvector method proposed in this paper, the prediction accuracy of the KRR model is better than that of the results obtained from the Magpie method and Atom2Vec method. It shows that the atomic feature vector proposed in this paper has a certain correlation between the features, and is a low-dimensional and densely distributed feature vector. Visual analysis and numerical experiments of material property prediction show that the atomic feature representation method based on deep learning extraction proposed in this paper can effectively characterize the material features and can be applied to the tasks of material band gap prediction.