A physics-informed graph neural network approach to quantify spatial-temporal impact forces in seismically pounding concrete structures with variable separation gaps

A Physics-Informed Graph Neural Network (PI-GNN) framework is proposed to quantify spatial-temporal impact forces in seismically pounding concrete structures with variable separation gaps. High-fidelity nonlinear dynamic simulations were conducted using LS-DYNA to replicate the complex interactions between adjacent concrete elements under seismic loading, generating a comprehensive dataset that captures diverse gap configurations, material nonlinearities, and dynamic excitation scenarios. The PI-GNN model is implemented in Python using the PyTorch Geometric library. It leverages graph convolutional networks (GCNs) enhanced with attention mechanisms to extract spatial correlations among structural nodes, while a recurrent neural network (RNN) module is employed to capture the temporal evolution of impact forces. To enforce physical consistency, the governing equations of motion and contact mechanics are embedded into the loss function as penalty terms, ensuring that the network predictions adhere to established mechanical principles. Model performance is validated against benchmark finite element analyses performed in ANSYS Mechanical, demonstrating notable improvements in both computational efficiency and prediction accuracy relative to conventional methods. Sensitivity analyses further illustrate the influence of variable separation gaps on the pounding response, offering valuable insights into the structural dynamics under seismic excitation. This work not only advances the state-of-the-art in computational structural dynamics but also establishes a practical pathway for integrating physics-informed deep learning techniques into seismic risk assessment and structural health monitoring applications.