Prediction of thermodynamic properties of lanthanide/transition metal alloys by deep learning

Abstract

The utilization of machine learning, especially deep learning, in solving materials science issues bring an opportunity to accelerate the development process of new materials and draw the attention of researchers all over the world. In this work, we present our study on applying deep neural networks to represent and predict thermodynamic quantities including formation energy, convex hull distance, and to recognize potential thermodynamical stabile materials. We employ our novel material descriptor, named orbital field matrix (OFM), to determine the feature vectors for materials. The OFM descriptors were developed based on the information of valence electron configuration and the Voronoi analysis of the atomic structures of materials. Our experiments show that deep neural networks can accurately predict formation energy and convex hull distance with the mean absolute error around 0,124 eV/atom and 0,105 eV/atom, respectively. In addition, the classification neural network can yield an accuracy of 92% in distinguishing the stable and unstable materials.