Comparison of Liquefaction Potential Prediction Results Using Machine Learning Methods Based on CPT Data

Soil liquefaction is a significant geotechnical hazard that can lead to severe structural damage during seismic events. Traditional liquefaction assessment methods, such as those based on the Standard Penetration Test (SPT) and Cone Penetration Test (CPT), rely on empirical correlations but often struggle to capture the complex, nonlinear interactions between soil properties and seismic parameters. Recent advancements in machine learning (ML) offer data-driven approaches that can improve liquefaction prediction accuracy. This study evaluates and compares the performance of Random Forest (RF) and Artificial Neural Networks (ANNs) for liquefaction potential prediction using a dataset containing 480 field observations derived from CPT-based studies. The dataset was preprocessed using min-max normalization, and models were trained and optimized through hyperparameter tuning. Model performance was assessed using accuracy, precision, recall, F-measure, Cohen’s kappa, and AUC-ROC analysis. The results show that RF achieved the highest accuracy (89%), outperforming both ANN (86%) and the traditional CPT-based liquefaction assessment method (87%). Additionally, ROC-AUC values of 0.932 for RF and 0.872 for ANN indicate the superior classification capability of machine learning models. Feature importance analysis in RF revealed that cone tip resistance (qc), cyclic stress ratio (CSR), and peak ground acceleration (amax) are the most influential factors in liquefaction prediction. These findings demonstrate that machine learning techniques, particularly RF, provide more reliable liquefaction predictions compared to conventional empirical methods. The study highlights the potential of ML models in improving seismic risk assessments and guiding engineering decision-making processes.