Interpretable machine learning for predicting splitting strength of asphalt concrete: insights from SHAP analysis

Predicting the splitting strength (ST) of asphalt concrete is essential for assessing pavement durability, yet traditional experimental methods are time-consuming and costly. This study investigates whether machine learning (ML) can provide a more efficient, data-driven alternative. The authors compiled a database of 296 asphalt concrete samples, incorporating 14 variables related to asphalt properties, aggregate gradation, and fiber characteristics. They compared six ML models: TabPFN, ANN, SVR, RF, XGBoost, and LightGBM. While five models underwent hyperparameter optimization using the NSGA-II genetic algorithm, the TabPFN model was applied using its default configuration to test its efficiency as a foundation model for tabular data.

All six models demonstrated satisfactory predictive capabilities, but TabPFN emerged as the superior model, achieving the highest R-squared (0.88) and the lowest error metrics (RMSE of 0.28, MAE of 0.21). The study utilized SHAP (SHapley Additive exPlanations) to demystify the model's decision-making process. The analysis revealed that nine variables accounted for 92% of the total contribution to ST predictions. Specifically, the passing rates of 9.5 mm and 4.75 mm aggregates, fiber type, asphalt content, and ductility were identified as the most critical factors. The authors further quantified favorable parameter ranges—such as keeping asphalt content below 5.4 wt.% and air voids below 3.6%—to optimize splitting strength.

To bridge the gap between complex ML models and practical engineering, the researchers developed a Graphical User Interface (GUI). This tool integrates the trained TabPFN model with SHAP-based explanations, allowing engineers to input mixture design parameters and receive both an ST prediction and an interpretation of how those specific variables influenced the result. This framework aims to support more transparent and accessible data-driven mixture design in civil engineering.