Downloads: 1
Survey Paper | Neural Networks | Volume 15 Issue 5, May 2026 | Pages: 861 - 870 | India
A Survey on Graph Neural Networks for Crystal Property Prediction: Architectures, Expressivity, Uncertainty Quantification, Out-of-Distribution Generalization
Abstract: Graph Neural Networks (GNNs) have become the dominant machine learning paradigm for predicting physical and chemical properties of crystalline materials, offering prediction speeds several orders of magnitude faster than Density Functional Theory (DFT) while maintaining competitive accuracy. However, three fundamental limitations remain unresolved in existing systems: (1) the single-radius graph construction creates a provable representational ceiling bounded by the 1-Weisfeiler-Leman (1-WL) graph isomorphism test; (2) standard random-split evaluation systematically overestimates performance by 3-5? compared to structure-aware out-of-distribution (OOD) evaluation; and (3) no existing system provides statistically guaranteed prediction intervals. This survey presents a comprehensive review of GNN architectures for crystal property prediction, covering single-scale distance-based GNNs (CGCNN, SchNet, MEGNet), angle-aware line graph networks (ALIGNN, DimeNet), multi-scale and multi-view approaches (PMCGNN, PSCG-Net), and equivariant architectures. We provide the first formal expressivity separation proof for multi-radius crystal GNNs: Theorem 1 establishes that multi-scale GNNs are strictly more expressive than any single-radius GNN on periodic crystal graphs, supported by 10 empirically confirmed Weisfeiler-Leman collision pairs across TiO2, SnO2, and Fe2O3. We further survey uncertainty quantification methods- Monte Carlo Dropout, Deep Evidential Regression, deep ensembles, and conformal prediction- and propose the first application of split-conformal prediction to crystal property prediction, achieving 89.1% empirical coverage at a 90% target with ECE of 0.074 and Spearman ? of +0.382 (p < 10-53) under SOAP-LOCO structural OOD evaluation. Development-scale results demonstrate MAE of 0.033 eV/atom and R2 of 0.996, exceeding all paper-target metrics at 20,000 structures before full-scale training.
Keywords: Graph neural network, crystal property prediction, uncertainty quantification, conformal prediction, out-of-distribution generalization, Weisfeiler-Leman expressivity, materials science, deep evidential regression, Monte Carlo dropout, SOAP-LOCO
How to Cite?: Vinod Kulkarni, Matharishwa S, Bichitra Behera, Bharath S, Khushi Kalpesh Joshi, "A Survey on Graph Neural Networks for Crystal Property Prediction: Architectures, Expressivity, Uncertainty Quantification, Out-of-Distribution Generalization", Volume 15 Issue 5, May 2026, International Journal of Science and Research (IJSR), Pages: 861-870, https://www.ijsr.net/getabstract.php?paperid=SR26513224012, DOI: https://dx.dx.doi.org/10.21275/SR26513224012