Machine learning techniques for data analysis in materials science

Abstract

This work applies Graph Neural Networks (GNNs), a class of deep learning methods, to predict physical properties and obtain optimal cathode materials for batteries. Two GNNs are selected: Crystal Graph Convolutional Neural Networks (CGCNN) and the more recent Geometric-Information-Enhanced Crystal Graph Network (GeoCGNN). Both networks are trained on a selected open-source ab initio Density Functional Theory (DFT) dataset for solid-state materials to predict the formation energy and then calculate the redox potential. Numerical results show the inference of the best trained model ran on combinatorial space of interest to discovery the optimal one via multi-objectives method. This approach allows to detect the optimum faster than physics-based computational approaches.