A machine learning framework for predicting physical properties in configuration space of gate alloys

Abstract

In materials design, it is meaningful to fast and accurately predict material properties required for a specific application based on specific structures. In the present study, a framework for predicting formation energy, bulk modulus, shear modulus and Debye temperature in a large configuration space of gate alloys was proposed, by integrating two approaches, density functional theory (DFT) and machine learning (ML). DFT calculations were conducted to obtain the physical properties, as labels of dataset used in ML, for 3051 configurations enumerated. For features of ML modeling, a comparative analysis was carried out on four structure descriptors as well as their combinations with element descriptors. The predictive performance with the optimal features combination was compared with that from crystal graph convolutional neural networks (CGCNN). The predictive ability of the framework was further validated by additional data collected from Materials Project platform. Finally, the effectiveness of transfer learning is sufficiently demonstrated in the extrapolation case of gate alloys system.