Machine-learning-based predictive models for estimating seismically-induced slope displacements

Abstract

Engineers often use semiempirical models, which estimate the amount of seismically-induced slope displacements (D), to evaluate the seismic performance of earth structures and natural slopes. These procedures often use as inputs slope properties, earthquake parameters, and ground motion intensity measures (IMs). In this study, we propose a new set of machine learning (ML) based models to estimate D using the NGA-West2 shallow crustal ground motion database. We consider both the classification of negligible D and its estimation. The selection of features to explain D (which is based on LASSO, Forward Selection, and Random Forest) suggests that the most efficient features are the slope's yield coefficient (ky), its fundamental period (Ts), the earthquake magnitude (Mw), the peak ground velocity (PGV), and the degraded spectral acceleration at 1.3 Ts. Moreover, the feature selection suggests that there is no significant gain in accuracy beyond five features. We formulate 19 different models, considering various ML-based algorithms such as Generalized Linear Models (GLM), Partial Least Square Regressions (PLSR), Principal Component Regressions (PCR), Bagging and Boosting, Random Forest, Polynomial-based regressions, Multi-order regressions, and Kernel-based models. We assess the performance of the proposed models by evaluating test errors, their predictive performance in case histories, and comparisons against existing models. Based on the assessments, we recommend 6 ML-based models to estimate D in engineering practice.