Assessment of SPT-based Liquefaction Potential using Ensemble Learning and Feature Importance Approaches

Abstract Engineers must assess soil susceptibility to liquefaction, yet conventional evaluations are time‑consuming, costly, and affected by field-testing variability and semi‑empirical tools that introduce uncertainty. This study clarifies the importance of key parameters and proposes a streamlined machine‑learning approach using ensemble methods. A comprehensive reference dataset was compiled and used to train reliable machine‑learning models. Feature importance was examined with logistic regression and random forest, after which the data were split into training and test sets, predictors were scaled, and hyperparame0ters were tuned with GridSearchCV. Advanced models were then fitted, followed by ensemble approaches, including AdaBoost and voting classifiers. Based on feature importance results, the most influential features across all methods continue to be the Standard Penetration Test-derived parameters. The trained models were assessed, in which the AdaBoost provided the most accurate estimations by achieving precision, recall, F1_score, Jaccard index, and accuracy of 88%, 88%, 88%, 79%, and 88%, respectively. The Voting Classifier demonstrated superior performance over the AdaBoost Classifier in terms of lower false negative values and higher true positive values, which could be considered as a better predictor for high-risk cases.