Prediction of the existing building-induced modifications of ground motion at layered sites using convolutional encoder-decoder neural networks (CEDNN)

Abstract

Previous studies have shown that existing buildings significantly affect ground motion at the foundation level due to seismic interaction between the building and the underlying site. The modification induced by the existing building is quantified by the transfer function, defined as the Fourier spectral ratio between the foundation motion and the free-field motion. This study focuses on evaluating the building-induced modification of ground motion using a convolutional encoder-decoder neural network (CEDNN) model. To this end, a series of numerical simulations were performed, including 2565 finite element models for the structure-soil system and the free-field condition. Using the simulation results as a database, the CEDNN model was developed to rapidly predict the transfer function. The predictive performance of the proposed model was then compared with that of other neural network models. The results indicate that the CEDNN model achieves high predictive accuracy, with mean absolute errors of 0.045 and a coefficient of determination of 0.967. Overall, the CEDNN model provides an efficient tool for predicting building-induced modifications of ground motion.