A Machine Learning-Based Hybrid Seismic Analysis of Base Isolated Buildings

The machine learning-based hybrid seismic analysis developed recently by the authors is validated in this paper using the full-scale experimental data of a shake-table test program performed at the Hyogo Engineering Research Center (E-Defense) of Miki, Japan. The experimental datasets contain the time history responses of two base-isolated buildings, each with the lead rubber bearings (LRB) and the triple pendulum bearings (TPB). First, the machine learning models (MLMs) based on the recurrent neural network (RNN) are designed to predict the complex load-deformation relationship of seismic isolators at the isolation layer. The experimental data are used to train and test the surrogate MLMs. Then, the developed MLMs are incorporated into the time integration method to perform the hybrid seismic analyses of base-isolated buildings subjected to the earthquake ground motions. The hysteresis curves of the isolation layer and the roof displacement time-histories are calculated and compared with the experimental records. The hybrid seismic analysis could capture the experimental responses. The proposed hybrid seismic analysis that takes advantage of both analytical methods and machine learning-based methods was found to be capable of obtaining with good accuracy responses of seismically isolated buildings.