EXPLAINABLE SUPPORT VECTOR REGRESSION FRAMEWORK FOR PREDICTING MAXIMUM DRY DENSITY FROM SOIL INDEX AND GRADATION PROPERTIES
Keywords:
Maximum dry density, soil compaction, support vector regression, machine learning, SHAP, partial dependence plots, geotechnical engineering, soil index propertiesAbstract
Maximum dry density (MDD) is a critical parameter in geotechnical engineering, governing the strength, stiffness, and long-term performance of compacted earth structures. Conventional laboratory determination of MDD through Proctor testing is time-consuming and labor-intensive, particularly when multiple soil sources or stabilization scenarios must be evaluated. This study presents a robust and interpretable machine learning framework based on Support Vector Regression (SVR) with a radial basis function (RBF) kernel for predicting MDD directly from routinely measured soil index and gradation properties. A dataset comprising 486 soil samples was compiled, incorporating gravel content, sand content, fines content, liquid limit, plastic limit, and optimum moisture content as input variables.
Model development involved systematic data preprocessing, feature standardization, hyperparameter optimization, and rigorous validation using training–validation–testing splits and 20-fold cross-validation. Model performance was evaluated using multiple statistical indicators, including R², RMSE, MAE, MAPE, and CVRMSE, alongside graphical diagnostics such as learning curves, residual analysis, and percentage error distributions. The SVR model achieved strong predictive accuracy, with R² values ranging from 0.86 to 0.92 and MAPE values close to 2%, demonstrating excellent generalization capability and minimal bias.
To enhance transparency and engineering interpretability, SHapley Additive exPlanations (SHAP) and partial dependence plots were employed. These analyses identified optimum moisture content as the dominant predictor of MDD, followed by fines and gravel content, consistent with established soil compaction theory. Nonlinear trends revealed by SHAP and PDPs further confirmed the physical realism of the learned relationships.
Overall, the proposed SVR-based, explainable framework provides an accurate and interpretable surrogate for laboratory compaction testing, offering significant potential to improve efficiency and decision-making in geotechnical design and construction practice.
