Convolutional neural network-based liquefaction prediction model and interpretability analysis

Conventional methods for liquefaction discrimination are often semi-empirical, prone to human factors, with low success rates and balance. Moreover, current machine learning approaches lack ample sample support and have specific limitations. Through the integration of liquefaction datasets, 11 features were chosen: corrected standard penetration test blows, fines content, soil layer depth, groundwater table depth, total overburden stress, effective overburden stress, threshold acceleration, cyclic shear stress ratio, shear wave velocity, earthquake magnitude, and peak ground acceleration. A convolutional neural network (CNN) model was established. The Borderline SMOTE technique was introduced to address the issue of imbalanced datasets. The CNN model was compared against random forest, logistic regression, support vector machine, extreme gradient boosting models, and methods specified in Chinese codes. Furthermore, the SHapley Additive exPlanations (SHAP) algorithm was utilized to examine the influence trends of input features on the prediction outcomes. The results demonstrated that the CNN model attained an accuracy of 92.58%, outperforming all metrics of the four machine learning models and the method specified in Chinese codes. Examination of the SHAP results unveiled that soil layers with corrected standard penetration blows below 15 exhibited a higher liquefaction probability, whereas layers with cyclic stress ratios under 0.25 were less prone to liquefaction. The influence patterns of each factor align with current understanding, indicating the prediction model’s credibility and reliability.