Machine Learning-Based Predictors for Maximum Pile Bending Moment of the Soil-Pile-Superstructure System in Liquefiable Soils

Abstract

Accurate and reliable prediction of the maximum pile bending moment of soil-pile-superstructure system (SPSS) in liquefiable soils is essential for the seismic design and damage assessment of structures. The maximum pile bending moment can be obtained using experiments that consider only a limited number of factors or time-consuming finite element simulations. In this case, this study aims to develop a machine learning-based model for predicting the maximum pile bending moment of SPSS in liquefiable soils with significantly improved time efficiency. To train the model, a large number of three-dimensional finite element models considering soil and structural uncertainties are dynamically analyzed under different ground motions to establish the database. In addition, six ML regression algorithms are used to develop the predictive model: multiple linear regression (MLR), lasso regression (Lasso), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), and extreme gradient boosting (XGBoost). The results show that the prediction models based on the boosting algorithm (especially XGBoost) are the best to make accurate predictions. The feature importance analysis results show that ground motion features (especially peak ground velocity, PGV) have the greatest effect than soil- and structure-related features on the maximum pile bending moment of SPSS in liquefiable soils.

Keywords: Machine learning, liquefaction, Soil-pile-superstructure system, Modeling uncertainties, Maximum pile bending moment prediction