Continuous Representation of Chemical Environments for the Prediction of Material Properties

Machine learning (ML) methods are becoming increasingly popular for accelerating the design of new materials by predicting material properties with computational speeds orders of magnitudes faster than ab-initio methods. Previously, we developed a generalized crystal graph convolutional neural networks (CGCNN) framework to directly learn structure-property relations from the connectivity of atoms in crystals, providing an accurate and interpretable representation of crystalline materials. Despite its success in the prediction of crystal properties, it fails to extend to a broader range of materials like polymers and glasses, where connectivity alone cannot completely describe the system due to their amorphous nature. In this work, we develop a continuous representation of materials that captures arbitrary configurational and compositional features to predict their properties. We demonstrate the improvement of prediction performances compared with CGCNN on crystalline materials, as well as its application on amorphous materials. Finally, several examples illustrating how this method can be applied to the design of new materials will be presented.