Akademik Veri Yönetim Sistemi
Advanced Engineering Informatics, cilt.70, 2026 (SCI-Expanded, Scopus)
Accurate estimation of natural vibration frequencies plays a critical role in assessing the dynamic behavior and structural safety of large-scale hydraulic infrastructures such as arch dams. These frequencies are sensitive indicators of changes in stiffness, mass distribution, and boundary conditions, and serve as fundamental inputs for seismic safety evaluations, operational modal analysis, and structural health monitoring. Reliable calculation of natural vibration frequencies is a cornerstone of structural system design. While conventional machine learning approaches have accelerated this process with notable success, they lack mechanisms to inherently incorporate structural connectivity. Existing data-driven models typically treat input features independently, neglecting the relational structures among design parameters and foundation interactions. To address this limitation, this study proposes a regression framework that integrates K-Means clustering (optimized via silhouette analysis) and a multi-head Graph Attention Network (GAT). A graph-based structure was formulated by first clustering the dataset based on dam-foundation interaction type, dam height, and concrete grade, followed by constructing a connectivity graph to enable information propagation within each cluster using the attention mechanism in GAT. The methodology was validated using a comprehensive ANSYS generated dataset of Type-1 arch dams, spanning the first ten natural vibration modes. Comparative analysis demonstrated that the GAT model consistently outperforms a deep neural network (DNN) baseline, maintaining high and positive R2 values (≈0.82) across all modes, while the DNN exhibits negative or near-zero R2 for most modes. The clustering structure enhanced model interpretability, revealing coherent physical patterns and highlighting influential features via attention weights. These features collectively enhance both the accuracy and explainability of modal frequency prediction, establishing the cluster-GAT approach as a robust tool for data-driven structural health assessment and design optimization.