A SHAP- and ALE-based interpretation framework for decoding complex influences on asphalt pavement cracking using LTPP data

While machine learning (ML) provides a powerful tool for correlating various factors and asphalt pavement cracking, its poor interpretability hinders a deeper mechanistic analysis of how these factors contribute to crack development. In this regard, this study proposed an interpretation framework for ML models, which consists of feature importance based on the SHapley Additive exPlanations (SHAP) method, factor influence analysis based on Accumulated Local Effects (ALE) and local conditional effects (LCE), and feature interaction analysis based on H-statistic and second-order ALE. Using the U.S. long-term pavement performance (LTPP) database, a gradient boosting regression tree (GBR) model was developed with maintenance, design, traffic, and climate factors as inputs and transverse, longitudinal, and alligator cracking as outputs. The generalization performance of all three cracking models exceeded 82%. Results indicated that transverse cracking is primarily triggered by low-temperature events, while longitudinal and alligator cracking are mainly caused by fatigue damage induced by traffic loads. Feature interaction analysis further revealed significant interactions in transverse cracking between the cumulative freezing index (CFI) and traffic load (CESAL) and in fatigue cracking between freeze‒thaw cycles (CFT) and precipitation (CTP). The outcome of this study is expected to bridge the gap between data-driven predictions and mechanistic understanding, providing interpretable insights for the design of durable asphalt pavements and the optimization of maintenance strategies.