A machine learning-based hybrid seismic analysis of a lead rubber bearing isolated building specimen

Abstract

This paper presents the first experimental validation of the hybrid seismic analysis, previously developed by the authors, by simulating the structural behavior of a full-scale base-isolated building specimen tested at the E-Defense. The two-dimensional structural behavior of the lead rubber bearings (LRBs) is predicted by a machine learning model (MLM) previously developed by the authors, whereas analytical models describe the response of the superstructure. A Python script has been developed to incorporate both analytical models and MLMs into an explicit time integration method to perform three-dimensional hybrid seismic analyses. Moreover, conventional time history analyses have been carried out using a bilinear model for the isolation devices. Experimental, ML-based, and conventional analysis responses of 14 shake table motions are processed and compared in terms of hysteresis curves of the isolation layer, roof displacement and peak floor acceleration. Overall, hybrid seismic analysis showed better results than conventional one in terms of capturing the energy dissipation capacity of the isolation layer (maximum a20-index of 0.859 for the drift and 0.720 for the shear with respect to experiment). Similarly, roof displacement was predicted with a maximum a20-index value of 0.832. Besides, the maximum average computation time of hybrid seismic analysis was 1.78 s per integration time-step. Therefore, the efficiency of the hybrid seismic analysis is experimentally proven in this paper in terms of both accuracy and applicability. This machine learning-based analysis method is strongly recommended to simulate time-history response of complex structures.