Prediction of Site Amplification of Shallow Bedrock Sites Using Deep Neural Network Model

Abstract

Site amplification models are widely used with ground prediction equations to estimate ground motion intensity measures. The time-averaged shear wave velocity of top 30 m (V_S30) is the primary site proxy in site amplification models. A large number of models have been developed for a range of site conditions. However, the simplified nature of all models produce large residuals compared with the computed responses. The prediction accuracy of the models can be greatly enhanced through use of machine learning technique. In this study, the outputs of nonlinear one-dimensional site response analyses are used to train the deep neural network (DNN) model. The linear and nonlinear components are separately trained. The comparisons highlight that the DNN model successfully captures the amplification characteristics of the shallow bedrock sites and produces significantly lower residual compared with the available simulation based model.