Predicting lattice thermal conductivity of semiconductors from atomic-information-enhanced CGCNN combined with transfer learning

Abstract

Rapid identification of lattice thermal conductivity of semiconductors from their crystal structure is required in the discovery of functional materials. A promising strategy is using a machine learning method based on a first-principles dataset, which, however, suffers from the dilemma of too little data available. In this work, the crystal graph convolutional neural networks (CGCNN) model was improved by enhancing the information of atomic descriptors (for short CGCNN-D), and the transfer learning (TL) method was combined to overcome the problem of small datasets. It is found that the CGCNN-D has improved predicting performance for both electronic bandgap with large data volume and thermal conductivity with small data volume, with the mean absolute error reducing 7% and 10%, respectively, indicating the importance of the improved atomic description. Applying TL with electronic bandgap as a proxy into the CGCNN-D further upgrades the prediction accuracy for thermal conductivity that has only 95 pieces of data, yielding 19% decrease in the mean absolute error as compared to the original CGCNN. The trained CGCNN-D-TL model was used to quickly estimate the thermal conductivities of thousands of semiconductors, and the materials identified with potentially high thermal conductivity were further screened by the optimized Slack model. Finally, the most promising BC2N was discovered and then confirmed by the first-principles calculations, which shows room-temperature thermal conductivities of 731, 594, and 500 W m−1 K–1 along the three principal axes of its lattice structure.