Prediction of nonlinear seismic response of underground structures in single- and multi-layered soil profiles using a deep gated recurrent network

Abstract

A deep learning method consisting of the gated recurrent (GRU) neural network is proposed to predict the seismic responses of underground structures in single-layered soil and multi-layered soil. The proposed neural network employed ground motion as the input, and the acceleration time history of underground structure acquired by a finite element method (FEM) to be the output. The distribution of the normalized errors and the error of the peak value of the predicted acceleration time history are adopted to evaluate the prediction performance and extrapolating ability of the proposed GRU architecture. The GRU network showed better prediction performance in the seismic response of a tunnel in a homogeneous sand layer than in a clay layer, as sand has a higher shear modulus representing lower nonlinearity of the soil-structure system. The peak acceleration of a two-story three-span subway station in multi-layered soil exhibited similar prediction accuracy to the tunnel in single-layered clay. The proposed GRU neural network generally displayed satisfactory prediction performances on the acceleration responses of underground structures in various soil conditions.