Graph Representation Learning for Predicting Structural Properties of Crystalline Materials

Abstract

Accelerating the discovery of advanced materials requires efficient approaches for predicting structural and physical properties without relying exclusively on computationally expensive simulations or laboratory experiments. This study proposes a graph representation learning framework for predicting the formation energies and structural stability of crystalline materials. The proposed architecture represents crystal structures as periodic atomic graphs and combines message-passing neural networks with geometry-aware attention mechanisms to capture local atomic interactions and long-range structural dependencies. The illustrative experimental dataset contains 128,000 crystal structures spanning 64 chemical elements and multiple material families. The proposed model achieves a mean absolute error of 0.036 eV per atom for formation energy prediction, compared with 0.049 eV per atom for a conventional crystal graph neural network baseline. For binary stability classification, the framework achieves an AUROC of 0.941 and an F1-score of 0.893. Furthermore, the model reduces average property prediction time by approximately 73% compared with a selected computational screening workflow. These results suggest that geometry-aware graph learning can improve the efficiency of computational materials screening and accelerate the identification of promising crystalline compounds.