Interpretable ML Predicts Asphalt Concrete Splitting Strength

Predicting the splitting strength (ST) of asphalt concrete is critical for material design in civil engineering. This research introduces an interpretable machine learning framework to achieve this, leveraging a database of 296 samples with 14 input variables related to asphalt, aggregate, and fiber characteristics. Six different machine learning models, including TabPFN, ANN, SVR, RF, XGBoost, and LightGBM, were developed and rigorously compared. TabPFN emerged as the top performer, demonstrating superior predictive capability. Crucially, SHAP (SHapley Additive exPlanations) analysis was employed to provide interpretability, identifying nine dominant variables that account for 92% of the total contribution. The analysis also quantified favorable parameter ranges for improving splitting strength, such as specific aggregate sizes and fiber content. A graphical user interface (GUI) was developed to enhance the framework's practical applicability.