Liquefaction susceptibility prediction using ML-based voting ensemble classifier

An accurate assessment of soil liquefaction susceptibility is critical in the design of earthquake-resistant structures. A large dataset consisting of 749 field cone penetration test (CPT) and standard penetration test (SPT) results is integrated into an ensemble of machine learning (ML) models. This study presents a novel approach for predicting liquefaction susceptibility using the voting ensemble-based ML model. This ensemble architecture considered in the study combines different classification algorithms, viz., logistic regression (LoR), decision tree classifier (DTC), random forest classifier (RFC), XGBoost classifier (XGBC), and gradient boosting model (GBC). Among the different ML models examined, the RFC is found to be the most reliable prediction tool. However, individual ensemble models are prone to overfitting; hence, voting ensemble models are considered to improve prediction accuracy. Out of 62 possible combinations, a voting ensemble consisting of LoR, XGBC, and GBC is found to be the best-performing model. Based on the sensitivity analysis carried out using the best-performing voting ensemble, the penetration resistance is found to significantly influence the model performance, followed by peak ground acceleration. Additionally, a separate analysis performed on 250 data points of individual field tests (i.e., CPT and SPT) showed that ML models fed on the CPT dataset provide better accuracy compared to the SPT dataset. Finally, the accuracy of the proposed voting ensemble model is compared with the traditional approach used to perform liquefaction susceptibility. The findings of this study hold significant implications for liquefaction susceptibility prediction, implying that the CPT data available at a given site can enhance decision-making and design processes using the proposed ML models.