Generic Models for Predicting Coseismic Displacements of Earth Slopes Based on Numerical Analysis and Machine Learning Algorithm

Generic models for estimating the earthquake-induced displacement of Earth slopes are developed based on a numerical approach. A number of 14,112 slope models with different configurations of slope geometry and soil property parameters are developed to represent generic Earth slopes. Thousands of slope dynamic analyses are then conducted in FLAC to estimate the coseismic slope displacements. Based on the displacements calculated, 18 ground-motion intensity measures (IMs) and eight slope variables are considered as candidate predictor variables to develop predictive displacement models using the light gradient boosting machine (LightGBM). Comparative results indicate that yield acceleration ($K_y$) and Arias intensity (AI) are the most efficient scalar variables in regressing the displacements. Based on the efficiency, sufficiency, and computability criteria, the vector IMs of (AI, peak ground velocity) and (AI, peak ground acceleration), together with $K_y$ and initial shear modulus, are regarded as the preferable predictor variables, respectively. Two sets of predictive displacement models are thus proposed using the preferable variables via the LightGBM- and polynomial-based approaches, respectively. The aleatory variability in predicting the slope displacement for the polynomial models is approximately 15%–30% larger than that of the LightGBM models, indicating that the predictive performance of the LightGBM models is superior to the polynomial models.