Machine learning prediction of stability, topological properties and band gap of topological insulators in tetradymites

Abstract Quickly design of excellent TI materials is an important issue that needs to be solved urgently. This reports machine learning aided models with molecular descriptors to predict the electronic structure properties of TIs. Specifically, 18 simple and physically meaningful kinds of descriptors are defined to characterize 243 tetradymites insulators. It is shown that the artificial neural network with the 4-22-1 structure has the best external prediction performance with an RMSE value of 0.046. The RUSBoosted Trees (Accuracy = 95.8%, AUC = 0.87) and Quadratic SVM (Accuracy = 85.4%, AUC = 0.90) exhibit the best accuracy among 23 ML classification models. This work reveals that ML model in combination with descriptors can serve as an excellent strategy for fast prediction new quantum materials without time-consuming quantum mechanical studies.