Physics-informed graph neural network representation learning for crystal property prediction

Abstract Accurate crystal property prediction is essential for accelerating the discovery of materials. Crystal properties arise from the combined effects of various physicochemical attributes. Existing methods struggle to fully utilize attributes crucial for property prediction, limiting the representation capability of models. Systematically incorporating relevant attributes into crystal graphs for joint modeling with graph neural networks, without redundancy and while maintaining model universality, remains challenging. We propose AtomNet, a multi-edge graph neural network framework for crystal representation learning and property prediction. AtomNet initializes nodes by creating atomic descriptors from selected physicochemical attributes and enhances edge features with electronegativity differences, producing highly expressive crystal graph representations. Integrated gradients assess the contribution of attributes to property prediction. A distance decay weight function is adopted to prioritize core nodes over distant neighbors. Experimental on two standard benchmarks demonstrate AtomNet’s superior performance, validating the effectiveness of our proposed methods for accelerating new material discovery.