Dynamic modeling and substructuring analysis leveraging long short-term memory neural network

Existing studies that use neural networks for civil engineering problems mostly established neural network models for the entire structure, and seldomly created dynamic interaction between neural network models and other models. This paper focuses on the dynamic substructuring modeling of structures installed with energy dissipation devices (EDDs) that are modeled by long short-term memory neural network (LSTM) models. Numerical studies at the device level show that LSTM models can be effectively trained to represent different physical models of EDDs with high precision. Investigations of different training schemes indicate that using band-limited white noise inputs to generate the training dataset is the optimal option. Procedures to establish coupling models for the inter-story-damped and base-isolated structures in the MATLAB Simulink platform are presented, where the LSTM model of EDDs is coupled with the physical model of the main structure through the state-space substructuring method. Seismic analyses of these two types of structures are conducted, showing that the coupling model achieves high performance and could relax computational or programming requirements that EDDs’ physical models may have. The present work reveals that the proposed hybrid structural modeling procedure utilizing the LSTM model is a flexible and practical option for structural analysis.