Data-driven prediction of unconfined compressive strength in stabilized soils using machine learning

Accurately predicting the unconfined compressive strength (UCS) of Stabilized soil is essential for advancing sustainable geotechnical engineering practices. This study proposes a hybrid machine learning and interpretability framework to estimate UCS based on a comprehensive set of 702 samples comprising soil properties and mixture compositions. Four models, including a back propagation artificial neural network, random forest, extreme gradient boosting, and deep forest, were developed and evaluated using multiple performance metrics. The DF model demonstrated superior accuracy and generalization, achieving the highest coefficient of determination and the lowest error indicators. To enhance model transparency, SHAP analysis was employed to quantify global and local feature contributions, revealing that the NaOH and GGBS contents were the most influential variables. Further interpretation using partial dependence plots, individual conditional expectation, and accumulated local effects techniques confirmed the nonlinear but saturating effects of these activators on UCS. The proposed framework improves prediction accuracy and offers mechanistic insights into the role of key parameters, providing a robust decision-support tool for mix design optimization in stabilized soils.