Time-dependent compressibility characteristics of soil: Experimental and integrated machine learning framework

Abstract

The mechanical performance of clay as a geomaterial is influenced by its time-dependent stress-strain characteristics. It is a complex phenomenon triggered by the viscous nature of adsorbed pore water and the internal arrangement of soil structure. This research emphasizes the significance of creep deformation characteristics of soil in both its natural and reconstituted states, utilizing the Elasto-Viscoplastic Swelling (EVPS) model. Various predictive machine learning models were presented to determine the creep parameters. To optimize model performance, Extreme Gradient Boosting (XGBoost) and Bayesian optimization were used. The findings observed that incorporating admixtures significantly influences the compressibility characteristics of clay, leading to a reduction in time-dependent parameters such as the creep coefficient and creep strain limit. Notably, the study finds that creep effects are most pronounced during the second cycle, with a gradual reduction occurring in subsequent cycles. Among the predictive models, the Random Forest exhibited the highest predictive accuracy, while Linear Regression and Support Vector Regression showed comparatively lower performance. Additionally, SHapley Additive exPlanations (SHAP) were utilized to explore feature importance and revealed that approximately 74.24% of the total mean absolute SHAP values were attributed to water content and plasticity behavior, highlighting a pivotal role in prediction, followed by applied stress. Moreover, the study also highlights integrated approaches of the EVPS model and machine learning modeling for accurately determining creep behavior. © 2025 Elsevier Ltd.