First-principles simulation of electronic properties of MoB/Si3N4 superlattices via machine learning

Abstract

MBene materials and Si3N4 materials have excellent properties, but there are still many technical bottlenecks in the preparation of high-performance and flexible structure electronic device materials. In this paper, MBene is combined with Si3N4 materials to form a superlattice structure, which is composed of six superlattice heterostru ctures. Firstly, the first-principles method was used to study the electronic properties of MBene/Si3N4 superlattice materials. The results show that with the increase of stacking times, the material structure becomes more stable, and the interlayer electrons are transferred from the MoB layer to the Si3N4 layer. Second, in the machine learning part, the Extreme Gradient Boosting (XGBoost) algorithm and the Convolutional Neural Network (CNN) algorithm were used to predict the density of states of six superlattice heterojunctions through ensemble learning, and it was found that the best prediction results were achieved when the input data with a combination of global descriptors and local descriptors were used. The lowest MAE is only 0.03, which not only indicates that the predicted density of states model is excellent, but also proves that the more refined the characterization of the surrounding environment of the atom, the more accurate the prediction of the density of states of the material and the higher the generalization performance of the prediction model.