Predictive modeling of Cbr and compressibility in lime stabilized lateritic soil using machine learning and Pchip data augmentation

This study investigated the compressibility and strength enhancement of A-2-6(0) lateritic soil, common in tropical regions, using quicklime stabilization. Integrating laboratory experiments with machine learning (ML) predictive modeling, the research aimed to support field-scale applications. Lateritic soil samples were treated with quicklime at concentrations from 0 to 30% by dry weight, and standard ASTM tests, including California Bearing Ratio (CBR), were performed. Untreated soil yielded a CBR of 25.7%. The addition of 10% quicklime proved optimal, achieving a peak CBR of 34.5%, representing a 34.24% improvement. Beyond this threshold, CBR values declined, indicating an optimal stabilization point. To enhance predictive modeling, Piecewise Cubic Hermite Interpolation (PCHIP) augmented the dataset from 9 to 61 points. Several ML models, including Linear Regression, Polynomial Regression, Support Vector Regression (SVR), Decision Tree, and Random Forest, were trained using various soil and lime parameters. Polynomial Regression (degree 3) exhibited the highest accuracy (R² = 1.0000, RMSE = 0.0167), with Linear Regression also performing exceptionally well (R² = 0.9999), while SVR achieved R² = 0.9997 (RMSE = 0.0427, MAE = 0.0217). Random Forest (R² = 0.9821, RMSE = 0.3417) and Decision Tree (R² = 0.9740, RMSE = 0.4119) exhibited strong but lower accuracy. The study’s novelty lies in its integration of PCHIP for data augmentation and interpretable ML for precise CBR prediction. This framework offers a rapid, cost-effective method for field decision-making, confirming 10% lime as the optimal content for static load applications in variable natural lateritic soils.