A Machine Learning–Based Predictive Model for Liquefaction Ejecta Manifestation in Stratigraphically Variable Sites

This paper presents a machine learning–based probabilistic predictive model for estimating the potential severity of free-field surface ejecta in stratigraphically variable and interbedded deposits as a measure of liquefaction manifestation and damage. The random forest (RF) model is developed using the results of an extensive numerical parametric study, which was calibrated and validated with triaxial and centrifuge experiments, as well as observations from liquefaction case histories. The numerical database includes 19,760 fully coupled, nonlinear, dynamic, effective stress, finite-element analyses of three-dimensional (3D) free-field soil columns subject to one-dimensional vertically propagating horizontal shear waves. The database was designed using quasi–Monte Carlo sampling of a large number of soil and ground motion parameters. The ejecta potential index (EPI) was used as a proxy for the potential and magnitude of sediment ejecta. The machine learning–based predictive model was calibrated and refined with data from the 2010–2011 Christchurch, New Zealand, earthquake liquefaction case history database. The model’s predictive performance was further evaluated with the available global liquefaction manifestation case history database. The proposed machine learning model enables prediction of liquefaction ejecta severity in a wide range of soil deposits. Further validation is needed with new well-documented case histories globally. Nevertheless, the proposed model is the first of its kind to incorporate the EPI proxy probabilistically, while capturing key hydromechanical processes that drive sediment ejection in interbedded deposits based on an extensive set of numerical and field data.