• Assessing shear strength parameters is important in asphalt pavement design. • Interpretable ML predicts Hveem-derived shear strength from Marshall test data. • Extra Trees provides the best results for cohesion and friction angle. • SHAP reveals key drivers: bitumen content, air voids, flow, fracture percentage. • SHAP-guided regression maintains high accuracy using four-parameter formulas. . For designing long-lasting asphalt pavements, it is important to accurately measure shear strength parameters. However, Hveem stabilometer testing, commonly used to determine cohesion and internal friction angle, is time-consuming and requires specialized equipment. In contrast, Marshall tests are low-cost, widely available, and generate extensive data. This study introduces an interpretable ensemble machine learning framework to predict Hveem-derived cohesion and internal friction angle based on Marshall test parameters. An experimental database of 108 asphalt mixture specimens was established, systematically varying mixture type, filler content, fracture percentage, and bitumen content. Multiple ensemble learning algorithms were evaluated under identical hyperparameter settings, with the Extra Trees model achieving the best performance for both target variables. Specifically, for cohesion, the model achieved R 2 =0.900 and RMSE=0.239 kPa, and for internal friction angle, it got R 2 =0.845 and RMSE=1.641°. SHAP-based interpretability was used to quantify the contribution of each feature and to analyze features interactions. The analysis revealed that bitumen content is the dominant predictor of cohesion, whereas air voids primarily govern internal friction angle. Guided by these insights, SHAP-based regression was applied by performing ordinary least squares on SHAP-selected features to derive four-parameter linear equations. These simplified formulas achieved accuracy comparable to ensemble models (R 2 =0.906 vs. 0.900 for cohesion; R 2 =0.887 vs. 0.845 for friction angle) while reducing complexity to four interpretable variables. Independent validation demonstrated that prediction errors fell within the standard uncertainty range of Hveem measurements (±5–10%).
Arabani et al. (Sun,) studied this question.