Experimental investigation and explainable machine learning for predicting the soaked CBR of lime–silica fume stabilized clay

Low-bearing-capacity clay subgrades often require stabilization before use in pavement systems. Quicklime (QL) is widely used for soil improvement, while silica fume (SF) provides a silica-rich industrial by-product for lime-based stabilization. However, the soaked California Bearing Ratio (CBR) of QL-SF stabilized clay depends on interacting mixture, curing, plasticity, and compaction variables. This study presents an integrated experimental, metaheuristic-optimized, and explainable machine learning framework for evaluating the soaked CBR of QL-SF stabilized Mershon kaolin soil. The primary dataset comprised 210 specimens representing 14 mixtures, three curing ages, and 42 mix–curing groups; 40 additional specimens were used in a five-cycle wetting–drying assessment. The combined treatment outperformed the single-additive mixtures, and QL3SF7 achieved the highest 28-day soaked CBR of 42%, approximately 14 times that of the untreated clay. After five cycles, QL3SF7 retained 97.0% of the soaked CBR of its age-matched control and showed the smallest dimensional and mass changes. Five models, including MLR, GB, CatBoost, XGBoost, and BP, were evaluated, and the nonlinear models were tuned using the Puma Optimizer (PO). Nested grouped cross-validation prevented replicate leakage and assessed unseen experimental groups. CatBoost-PO performed best, with grouped out-of-fold R 2 = 0.917 and RMSE of 2.796. A reduced model using only SF content, QL content, and curing days retained high accuracy, with R 2 = 0.930 and an RMSE of 2.582. Grouped out-of-fold SHAP analysis identified curing days and QL content as the dominant predictors and showed that the modeled effects of SF and QL varied with curing days and plasticity index, respectively. For geotechnical engineers, the proposed framework supports preliminary mix screening within the investigated material range and reduces reliance on trial-and-error laboratory testing but does not replace laboratory confirmation.