Machine learning prediction of seismically induced foundation settlement on interbedded liquefiable deposits: Model development, interpretation, and the effect of dense granular columns
Soil-structure interaction (SSI) strongly influences the seismic performance of shallow-founded structures, particularly when liquefaction occurs in stratigraphically variable deposits. This study develops machine learning (ML) models to emulate three-dimensional nonlinear effective-stress finite element (FE) predictions of permanent foundation settlement at substantially lower computational cost. The models are trained using the DesignSafe PRJ-5802 database, comprising 1676 simulations of unmitigated sites and 1272 companion simulations with dense granular column (DGC) mitigation, representing 96 soil-foundation-structure configurations subjected to 59 recorded ground motions. Eight ML algorithms are evaluated using configuration-grouped 10-fold cross-validation, ensuring that no soil-foundation-structure configuration appears in both training and test sets. Gradient-boosted ensemble models provide the best performance; the Gradient Boosting model achieves an out-of-fold R² of 0.535 for natural-log settlement, a mean absolute error of 6.74 cm, and 68% of predictions within a factor of two of the FE simulations. The study also demonstrates that the exceptionally high R² = 0.9965 reported in the original analysis resulted from methodological errors, including prediction of foundation contact pressure rather than settlement and data leakage caused by noisy replicated records. Reproducing the original procedure yields R² = 0.9984, whereas eliminating leakage reduces it to 0.884. SHAP analysis identifies ground-surface slope, sand relative density beneath the footing, cumulative absolute velocity, and foundation contact pressure as the dominant predictors. DGC mitigation reduced permanent settlement in 87% of paired simulations, with a median settlement ratio of 0.74. An explicit applicability domain and lognormal prediction interval are provided to support reliability-based design applications.