Probabilistic Predictive Model for Liquefaction Triggering in Layered Sites Improved with Dense Granular Columns

Abstract

This paper presents a probabilistic model for evaluating the liquefaction-triggering hazard in level, layered, and saturated granular soil profiles improved with dense granular columns (DGCs). The model is developed using the results of a comprehensive numerical parametric study, validated with a dynamic centrifuge test, and subsequently tested with the available case histories involving DGCs as a liquefaction countermeasure. The numerical database includes a total of 30,000, three-dimensional (3D), fully coupled, nonlinear, dynamic finite-element simulations with a statistically determined range of layer-, profile-, DGC-, and ground motion–specific input parameters. The criteria for the predicted degree of liquefaction (i.e., full, marginal, and no liquefaction) are based on the peak values of excess pore pressure ratio and shear strain anticipated within each soil layer. A machine learning approach that performs multinomial logistic regression along with variable selection and regularization is used to develop a set of functional forms for estimating the probabilities of full-, marginal-, and no-liquefaction in sites improved with DGCs. The proposed probabilistic model is the first of its kind that explicitly considers variations in the area replacement ratio (Ar), stiffness, and drainage capacity of the DGC; the thickness, depth, relative density, and hydraulic conductivity range of each layer; the evolutionary characteristics of ground motions; and the underlying uncertainty in the prediction of pore pressures and shear strains within each layer.