Assessment of Crack Resistance and Influencing Factors of Ultra-Thin Asphalt Overlay Mixtures Using SCB Testing

Asphalt ultra-thin overlays have emerged as an efficient and cost-effective pavement preservation technology. However, because of their reduced thickness, their service life is often limited by cracking. Based on semi-circular bending (SCB) test data, this study developed three types of regression models: multiple linear regression (MLR), random forest regression (RFR), and artificial neural network (ANN) regression. The multifactor experimental matrix covered three test temperatures, three nominal maximum aggregate sizes, two binder types, three aging states, and four long-term aging durations. The results show that fracture energy (G_f), flexibility index (FI), and cracking resistance index (CRI) respond differently to individual factors. G_f and FI decreased as temperature increased, whereas CRI increased, indicating that CRI may be misleading at intermediate temperatures and should be interpreted cautiously. The aggregate size exhibits a non-monotonic relationship with the three indicators, reaching its maximum at a size of 8. Short-term aging increased the G_f, and the gain was more pronounced at higher aging temperature, reaching a maximum increase of up to 1.54 times. However, it also accelerated the subsequent G_f reduction during long-term aging, after the decline rate decreased. Higher short-term aging temperatures caused more pronounced deterioration in FI and CRI. Meanwhile, both indices gradually decreased with increasing long-term aging duration, and short-term aging further intensified this reduction. In model comparisons, ANN showed the most stable generalization. The factorwise trends were verified using both the Beta values and SHAP values, demonstrating the feasibility of the models for qualitative discrimination and trend interpretation. The integrated importance ranking identified long-term aging duration and temperature as the dominant factors. External validation further confirmed ANN as the most reliable model for predicting crack resistance evolution. These findings support crack-resistance prediction and provide guidance for optimizing the design of ultra-thin overlay asphalt mixtures.