GRAPH NEURAL NETWORKS FOR STRUCTURAL DYNAMICS: ADVANCES IN PHYSICS-INFORMED LEARNING, SURROGATE MODELING, AND DAMAGE DETECTION

Abstract

Graph Neural Networks (GNNs) have emerged as a promising approach for addressing the computational challenges associated with structural dynamics, particularly where complex spatial relationships, irregular geometries, and large-scale simulations are involved. This review examines recent advances in GNN applications for structural dynamics, with emphasis on message-passing architectures, physics-informed learning, surrogate modeling, dynamic response prediction, and damage detection. More than 40 peer-reviewed studies are synthesized to evaluate architectural developments, computational efficiency, predictive performance, and generalization across structural systems. The reviewed studies indicate that GNN-based models can improve prediction accuracy by approximately 8–12% compared with conventional surrogate approaches while achieving computational speedups of up to 550–1200 times relative to high-fidelity finite element simulations. Hybrid architectures integrating topology-aware message passing with global routing demonstrate improved capability for capturing local structural interactions and long-range dependencies. Physics-informed GNN frameworks further enhance physical consistency and reduce dependence on large labeled datasets, while transfer learning supports prediction across previously unseen structural configurations. Despite these advances, challenges remain in temporal dynamics, material nonlinearity, interpretability, uncertainty quantification, data requirements, and large-scale deployment. Future research should focus on physics-guided architectures, probabilistic GNNs, temporal graph learning, transfer learning, and lightweight edge-deployment frameworks to support reliable real-time structural monitoring and engineering decision-making.