Data-driven modeling for interfacial behaviors between frozen soil and existing structures for applications of artificial ground freezing

When the artificial ground freezing technique is applied near existing underground structures, adfreezing behavior, characterized by ice bonding between the frozen soil and the existing structures, becomes a critical factor in assessing the stability of these structures. In this study, punch shear test data were employed to evaluate adfreezing behaviors at the frozen soil-structure interface under zero confinement conditions, representing critical states. Since machine learning (ML) algorithms have offered a powerful data-driven predictive modeling in geotechnical engineering, this study discussed the application of ML approaches to broaden the feasibility of the punch shear test for assessing the adfreezing behavior. Four ML algorithms, i.e., support vector regression (SVR), feedforward neural network (FNN), random forest (RF), and extreme gradient boosting (XGB), were adopted to develop predictive models based on the punch shear test results. To ensure optimal model performance, Bayesian optimization and five-fold cross-validation methods were employed to effectively train the ML models and identify the best hyperparameter combinations for each model. The predictive performance of these models was compared using three regression metrics: root-mean- square error (RMSE), mean absolute error (MAE), and determination coefficient (R2). The models were ranked based on their performance as follows: XGB > RF > FNN > SVR. Among them, the XGB model demonstrated the highest accuracy, with an RMSE of 0.0037, an MAE of 0.0015, and an R2 of 0.9999. The reliability and interpretability of the XGB model were further enhanced through post-hoc analysis estimating the prediction interval and SHAP values.