SS-XGBoost: A Machine Learning Framework for Predicting Newmark Sliding Displacements of Slopes

Abstract

Estimation of Newmark sliding displacement plays an important role for evaluating seismic stability of slopes. Current empirical models generally utilize predefined functional forms and relatively large model uncertainty is involved. On the other hand, machine learning method typically has superior capacity in processing comprehensive data sets in a nonparametric way. In this study, a machine learning framework is proposed to predict Newmark sliding displacements using the extreme gradient boosting model (XGBoost) and the Next Generation Attenuation (NGA)-West2 database, where the subset simulation (SS) is coupled with the K-fold cross validation (CV) technique for the first time to tune hyperparameters of the XGBoost model. The framework can achieve excellent generalization capability in predicting displacements and prevent data overfitting by using optimized hyperparameters. The developed data-driven Newmark displacement models can better satisfy both sufficiency and efficiency criteria, and produce considerably smaller standard deviations compared with traditional empirical models. Application of the models in probabilistic seismic slope displacement hazard analysis is also demonstrated. The proposed SS-XGBoost framework has great potential in developing data-driven prediction models for a wide range of engineering applications.