Supervised Intelligent Prediction and Explanation of Soil Shear Creep Parameters Based on Data

The phenomenon of soft soil creep, characterized by its long-term deformation behavior, exerts a profound influence on the settlement behavior of soft soil foundations. The determination of the soil’s average viscosity coefficient (AVC) through laboratory-based shear creep testing provides a means to elucidate the soil’s creep behavior. However, the execution of such laboratory-based shear creep tests is not only time-consuming but also entails significant expenditure of labor and material resources. In order to efficiently and accurately predict soil creep parameters, this paper employs six well-established supervised machine learning (SML) methods—back propagation neural network (BPNN), bi-directional long and short-term memory network (BiLSTM), decision tree (DT), random forest (RF), least-squares boosting (LSBoost), and extreme gradient boosting (XGBoost)—to predict the soil’s AVC creep parameters using 103 sets of shear creep experimental data from existing literature. To enhance model performance, Bayesian optimization (BO) combined with the meta-heuristic algorithm particle swarm optimization (PSO) was utilized to determine the optimal hyperparameters for the SML model. The prediction efficacy of different models on creep parameters was assessed using the correlation coefficient, root mean square error (RMSE), mean absolute error (MAE), and A20 index. The results indicate that the BO-XGBoost model outperforms other models in predicting soil’s AVC, achieving R2, RMSE, and MAE values of 0.9556, 2.7401, and 2.2013 for the test set, respectively. Finally, the SHAP and SOBOL analyses were conducted to compare the importance of input features and their influence on prediction outcomes, thereby enhancing the reliability and accuracy of the model predictions.