Mechanics-informed transformer-GCN for structural dynamic response prediction

Abstract

Digital twins are currently a research hotspot, of which efficient computation and real-time interaction are two key issues. However, for digital twinning of civil infrastructures, traditional computing methods are time-consuming, which prohibit their application to intensive and large-scale simulations. This paper proposes a mechanics-informed transformer-graph convolutional network (MI-TGCN) method for computing structural linear dynamic responses. A novel neural network architecture is designed through combining the transformer and GCN, in which mode-superposition method is innovatively integrated into the multi-head attention mechanism of the transformer to predict structural dynamic responses. Moreover, the adjacency matrix of GCN is replaced by the structural stiffness matrix because of their similarity in topological representation, which further forces structural dynamic responses to conform to the deformation compatibility principle. A five-story frame structure under seismic loads is employed as the numerical example to demonstrate the effectiveness of the proposed method. The results show that the proposed method not only achieves much higher computational efficiency but also predicts structural dynamic responses accurately. The proposed method runs an order of magnitude faster than the commonly used finite element methods.