Learning the Crystal Structure Genome for Property Classification
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 model to decode structure-property relationships in crystalline materials without explicitly considering chemical compositions. The focus is on classification of crystal systems, mechanical elasticity, electronic band gap, and phase stability. Our model utilizes a three-dimensional (3D) momentum space representation of structure from elastic x-ray scattering theory that exhibits rotation and permutation invariance. We perform novel ablation studies to help interpret the model performance by perturbing the physically meaningful input features (i.e., the diffraction patterns) instead of tuning the architecture of the learning model as in conventional ablation methods. We find that the spatial symmetry of the 3D diffraction patterns, which reflects crystalline symmetry operations, is more important than the diffraction intensities contained within for the model to make a successful classification. Our work showcases the potential of using statistical learning models to help understand materials physics, rather than performing predictive and generative tasks as in most materials informatics research. We also argue that learning the crystal structure genome in a chemistry-agnostic manner demonstrates that some crystal structures inherently host high propensities for optimal materials properties, which enables the decoupling of structure and composition for future codesign of multifunctionality.