Prediction Models of Newmark Sliding Displacement of Slopes Using Deep Neural Network and Mixed-effect Regression

Abstract

For assessing the seismic stability of slopes, the Newmark sliding displacement is an essential consideration. This study develops a deep neural network (DNN) prediction model and a hybrid prediction model with ground-motion intensity measures for the Newmark slope sliding displacements. The models are developed using a large number of ground motions and critical accelerations (ac) samples generated under a wide range of slope conditions. The DNN and hybrid models we proposed are more efficient and sufficient than the conventional regression models. In comparison with conventional regression models (SR08), they significantly reduce the standard deviations. For the first time, within- and between-ac residuals are distinguished in the hybrid model using a mix-effects regression algorithm. Further, the hybrid model that includes the effective friction angle improves prediction performance greatly. As a final step, the proposed models are applied in probabilistic slope displacement hazard analysis (PSSDHA) to demonstrate their effectiveness in engineering applications.