Large-scale soaked CBR prediction using machine learning: insights from standard and modified proctor test data

This study presents a systematic evaluation of five predictive models for estimating soaked California Bearing Ratio (CBR) values using soil index and compaction parameters across two energy levels: standard Proctor (CBR-SP) and modified Proctor (CBR-MP). The models Linear Regression (LR), Multilinear Regression (MLR), Cubic Model (CB), M5P tree, and Artificial Neural Network (ANN) were developed and validated using a comprehensive database of 2535 soil samples, comprising 1205 CBR-SP and 1,330 CBR-MP measurements. Eight input parameters were considered: gravel content, sand content, fines content, liquid limit, plastic limit, plasticity index, optimum moisture content, and maximum dry density. Model performance was assessed using multiple statistical metrics, including coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), variance accounted for (VAF), mean absolute percentage error (MAPE), and weighted mean absolute percentage error (WMAPE). The ANN model demonstrated superior predictive capability for both CBR-SP (R2 = 0.981, RMSE = 1.848%) and CBR-MP (R2 = 0.981, RMSE = 3.421%), followed by the M5P tree model (CBR-SP: R2 = 0.944, RMSE = 3.014%; CBR-MP: R2 = 0.952, RMSE = 5.246%). Sensitivity analysis revealed that the maximum dry density and liquid limit were the most influential parameters for predicting CBR-SP and CBR-MP. The models better predicted CBR-MP values than CBR-SP, particularly in the higher CBR ranges. Additionally, correlations were established between soaked CBR and unconfined compressive strength (UCS). The novelty lies in the combination of large-scale data-driven modeling with sensitivity analysis to identify the most influential parameters affecting soil strength. This methodology offers practical applications in geoenvironmental engineering by facilitating rapid, cost-effective soil evaluation, supporting sustainable infrastructure design, and optimizing the reuse of marginal soils, thus contributing to environmentally responsible geotechnical practices. This modeling framework offers geotechnical engineers efficient tools for estimating soaked CBR values, addressing the limitations of traditional testing methods by reducing reliance on time-consuming, labor-intensive, and resource-heavy procedures. It enhances project efficiency, lowers costs, and ensures more reliable and consistent CBR estimations.