Benchmarking ML approaches for earthquake-induced soil liquefaction classification

Abstract

Earthquake-induced soil liquefaction represents a critical geotechnical challenge due to its nonlinear soil–seismic interactions and its impact on structural safety. Traditional empirical methods often rely on simplified assumptions, limiting their predictive capability. This study develops and compares six machine learning (ML) classifiers—namely, Support Vector Machine (SVM), Artificial Neural Network (ANN), k-Nearest Neighbor (kNN), Random Forest (RF), Decision Tree (DT), and Naïve Bayes (NB)—to evaluate liquefaction susceptibility using an original dataset of 461 soil layers obtained from borehole penetration tests in the Edremit region (Balıkesir, NW Turkey). The models were trained and validated using normalized geotechnical and seismic parameters, and their performance was assessed based on accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). Results demonstrate that SVM, ANN, and kNN consistently outperformed other models, achieving test accuracies above 93%, F1 scores exceeding 98%, and AUC values between 0.933 and 0.953. In contrast, DT and NB exhibited limited generalization (test accuracy of 84–88% and AUC of 0.78–0.82), while RF showed partial overfitting. In contrast, DT and NB exhibited weaker generalization, with test accuracies of 84% and 88% and AUC values of 0.78 and 0.82, respectively, while RF indicated partial overfitting. The findings confirm the superior capability of advanced ML models, particularly SVM, ANN, and kNN, in capturing complex nonlinear patterns in soil liquefaction. This study provides a robust framework and original dataset that enhance predictive reliability for seismic hazard assessment in earthquake-prone regions.