A generalizable gated graph recurrent unit (Graph-GRU) network for nonlinear response prediction of cross-structures

Accurate seismic response prediction is essential for structural safety and resilience in civil engineering. Recently, artificial intelligence has emerged as a powerful tool for efficiently modeling the response of highly nonlinear structures. However, existing models struggle to generalize across diverse structural systems, which remains a bottleneck in deep learning-enabled surrogate modeling of nonlinear structures. This paper introduces a graph gated recurrent unit network (Graph-GRU) designed to achieve generalized nonlinear structural response prediction across different structures under unseen earthquakes. The core innovation lies in the specific design of the network by integrating both seismic excitations and structural characteristics into the GRU hidden state to learn the dynamic properties of different structures and achieve the generalizability to unseen structures. Here, the structural characteristics are featured using a graph convolutional network based on the structural graph with arbitrary degrees-of-freedom. Three pooling strategies including max, average, and attention pooling are considered to calculate the global structural feature vector. Additionally, the proposed approach is compared to the state-of-the-art deep learning models. The generalizability performance of the proposed Graph-GRU network is validated across 40 unseen reinforced concrete (RC) frames with varying design parameters of story heights and floor mass distributions. Results demonstrate that the proposed Graph-GRU is capable of predicting nonlinear responses of diverse unseen structures, effectively addressing the major generalizability challenge of existing methods. • Proposed a Graph-GRU network for generalized nonlinear structural response prediction. • Integrated structural graphs and seismic inputs through hidden state propagation. • Graph convolutional network is utilized to capture global structural dynamic features. • Explored max, average, and attention pooling methods for global structural feature fusion. • Achieved superior accuracy and generalization on 40 unseen RC frame structures.