ASUGNN: an asymmetric-unit-based graph neural network for crystal property prediction

Abstract

Material properties can often be derived directly from fundamental equations governing electron behavior. In this study, we present an open-source asymmetric-unit-based graph neural network designed to capture atomic patterns and their corresponding electron distributions. By coarse-graining sites belonging to conjugate subgroups and analyzing reciprocal space through powder X-ray diffraction patterns, our model predicts key physical properties, including formation energy, band gap, bulk modulus and metal/non-metal classification. Our method demonstrates exceptional predictive accuracy for properties calculated using density functional theory across the Materials Project dataset. Our approach is compared with state-of-the-art models and exhibits impressively low error rates in zero-shot predictions.