Abstract Estimating the level of earthquake-induced liquefaction settlements in shallow foundations is essential for assessing seismic risk and designing effective soil improvement strategies. Reliable predicting methods of vertical deformations and tilting in buildings founded on shallow liquefiable soils enable the development of seismic-resilient and cost-effective foundation solutions while supporting informed decision-making in earthquake insurance and mitigation planning. This study presents a comparative analysis of machine learning and deep learning models for classifying liquefaction-induced building settlements using a systematically compiled database of documented case studies. The dataset integrates building characteristics, geotechnical parameters, and seismic intensity indicators. Different kinds of resampling methods working at data level, cost-sensitive learning strategies, and algorithm-level advancements were studied in detail to treat extremely skewed class distribution. A SHAP-based feature selection method was applied as an extension to identify the most influential predictors, optimizing model efficiency without sacrificing predictive performance. Moreover, to improve classification reliability, dynamic threshold tuning and ensemble learning with weighted voting was employed. The result points out that data-driven feature selection and threshold optimization can better classify seismic damage, thus offering geotechnical earthquake engineering applications a methodology that is both interpretable and computationally efficient.
Sarğın et al. (Thu,) studied this question.