Multi-output time history prediction for seismic responses of structures with uncertain parameters via deep learning

Abstract

Conventionally, the design and assessments of structures would involve a large number of structural analyses. This process is usually realized by refined finite element models, which is extremely time-consuming. Recent advancements in the data-driven deep learning (DL) models have opened a viable avenue for forecasting time history responses. Nevertheless, most existing DL surrogate models treat the load as the sole input parameter, neglecting the impact of structural properties. Due to this limitation, it becomes impossible to achieve uncertainty quantification of response that considers stochastic structural parameters. Under this circumstance, two methods are proposed to incorporate structural parameters and ground motions (GMs) into input sources for predicting seismic responses in the present study. In this way, uncertainties within both structural parameters and GMs could be considered during uncertainty quantification of seismic response. The proposed methods’ feasibility was verified through a numerical case study involving a typical reinforced concrete frame structure. Furthermore, two benchmark methods that exclude structural parameters as input sources were employed to compare. Additionally, the application of seismic reliability analysis on the basis of these methods is elucidated. The results show that the proposed methods not only enhance the prediction precision and robustness of surrogate models compared with two benchmark methods but also achieve uncertainty quantification of seismic response considering uncertainties in structural parameters and GMs.