Performance evaluation and machine learning-based prediction of black cotton soil stabilization with supplementary cementitious binders

Abstract Expansive black cotton (BC) soils pose significant challenges for pavement subgrades due to their high plasticity, low strength, and moisture sensitivity. This study investigates the stabilization of BC soil using cement, lime, fly ash, and ground granulated blast furnace slag (GGBS), combined with machine learning (ML) for predictive modeling of strength and bearing capacity. Laboratory experiments evaluated Atterberg limits, unconfined compressive strength (UCS), and California Bearing Ratio (CBR) across varying dosages and curing periods. The untreated soil exhibited poor performance (LL = 58%, PI = 31%, UCS ≈ 1.2 kg/cm², CBR ≈ 8.8%). Cement showed the greatest strength enhancement, with 8% cement achieving ~ 19 kg/cm² UCS and ~ 19% CBR after 28 days. Lime was most effective in improving subgrade performance, with 9% lime yielding ~ 7.7 kg/cm² UCS and the highest CBR of ~ 27–28%. GGBS at 30% provided ~ 8.9 kg/cm² UCS and ~ 9% CBR, while fly ash (40%) achieved only ~ 3 kg/cm² UCS and 6% CBR. To complement the experimental program, ML algorithms—Decision Tree, Random Forest, and XGBoost—were developed to predict UCS and CBR. Random Forest delivered the best accuracy for UCS (R² = 0.99, RMSE = 0.25, MAE = 0.13), while XGBoost excelled for CBR prediction (R² = 0.99, RMSE = 0.20, MAE = 0.11). SHAP analysis identified cement dosage and curing time as dominant factors for UCS, and lime as the most influential for CBR. The integration of laboratory data with ML models establishes a robust framework for optimizing stabilizer blends, reducing experimental effort, and promoting sustainable, low-carbon pavement design.