Materials property predictions have improved from advances in machine-learning algorithms, delivering materials discoveries and novel insights through data-driven models of structure-property relationships. Nearly all available models rely on featurization of materials composition, however, whether the exclusive use of structural knowledge in such models has the capacity to make comparable predictions remains unknown. Here we employ a deep neural network (DNN) model, deepKNet, to learn structure-property relationships in crystalline materials without explicit chemical compositions, focusing on classification of crystal systems, mechanical elasticity, electrical behavior, and phase stability. The deepKNet model utilizes a three-dimensional (3D) momentum space representation of structure from elastic X-ray scattering theory and simultaneously exhibits rotation and permutation invariance. We find that the spatial symmetry of the 3D point cloud, which reflects crystalline symmetry operations, is more important than the point intensities contained within, which correspond to various planar electron densities, for making a successful metal-insulator classification. In contrast, the intensities are more important for predicting bulk moduli. Phase stability, however, relies more upon chemical composition information, where our structure-based model exhibits limited predictability. We find learning the materials structural genome in the form of a chemistry-agnostic DNN demonstrates that some crystal structures inherently host high propensities for optimal materials properties, which enables the decoupling of structure and composition for future co-design of multifunctionality.
Download