A Source-Aware and Physically Interpretable Data-Driven Framework for Predicting Semi-Circular Bending (SCB) Fracture...

Fiber-reinforced asphalt mixtures improve cracking resistance through fiber bridging, pull-out, and crack-path deflection, but their semi-circular bending (SCB) fracture energy is affected by coupled mixture, testing, and fiber-related variables. This study developed a source-aware and physically interpretable data-driven framework for predicting SCB fracture energy using a literature-derived database containing 261 valid sample-level records from nine source groups. The database was constructed through semantic extraction, unit normalization, rule-based checking, manual verification, and source identifier (SourceID) tracking. Optimum asphalt content, air voids, test temperature, loading rate, fiber dosage, fiber length, diameter, elastic modulus, and tensile strength were used as input variables. Under sample-wise testing, the selected model achieved a coefficient of determination (R²) of 0.89, a root mean square error (RMSE) of 0.0470 kJ/m², and a mean absolute error (MAE) of 0.0247 kJ/m² for the full dataset, while the fiber-containing subset achieved R² = 0.94, RMSE = 0.0194 kJ/m², and MAE = 0.0103 kJ/m². Source-aware validation showed higher prediction errors, indicating that cross-source generalization remains more challenging than internal sample-wise prediction. SHapley Additive exPlanations (SHAP) analysis identified temperature, fiber dosage, and fiber mechanical descriptors as dominant contributors, consistent with temperature-dependent viscoelasticity, fiber bridging, and pull-out mechanisms. The dosage-response analysis was restricted to the observed fiber-dosage range of 0-0.678%, providing a bounded screening tool rather than an extrapolative design equation.