Machine learning-based models for estimating liquefaction-induced building settlements

Engineers often estimate the amount of liquefaction-induced building settlements (LIBS) as a performance proxy to assess the potential of earthquake-induced damage to buildings. The first robust LIBS models were initially developed in 2017 and 2018 using traditional statistical approaches. More recently, machine learning techniques have started to be used in developing LIBS models. These recent efforts are a step forward in realizing the potential of machine learning in liquefaction engineering; however, they have often considered only one ML technique for a given dataset and typically used only held-out test sets for model assessment. In this study, five ML-based LIBS models with varying flexibility (i.e., ridge regression, partial least square regression — PLSR, random forest, gradient boosting decision tree — GBDT, and support vector regression) are developed using a LIBS database generated by soil–structure numerical simulations of different buildings and soil profiles shaken by ground motions with varying intensity measures. The motivation for considering models with different flexibility is to include different bias–variance trade-offs. Feature selection with different ML techniques indicates that cumulative absolute velocity, spectral acceleration at one second, contact pressure, foundation width, the thickness of the liquefiable layer, and a shearing liquefaction index are important features in estimating LIBS. The developed ML-based models are assessed considering prediction accuracy in test sets, performance against centrifuge tests and case histories, and trends. The assessment indicates that the random forest, GBDT, and SVR models perform best, providing standard deviation reductions up to 40% relative to a multi-linear regression. Specifically, the random forest and GBDT models exhibit a root mean square error (RMSE) of 0.29 and a coefficient of determination (R2) of 0.93 on test sets, demonstrating a notable improvement compared to a traditional multi-linear regression model, which yields an RMSE of 0.47 and an R2 of 0.82. Moreover, random forest and GBDT, alongside SVR, show a good performance in centrifuge tests and case histories. Finally, given the scarcity of LIBS models, this study also contributes to treating epistemic uncertainties in estimating LIBS, which is ultimately beneficial for performance-based assessments. •Liquefaction-induced building settlements (LIBS) have significantly affected infrastructure in recent earthquakes.•Most