An interpretable model for predicting the performances of asphalt mixtures comprising the chemistry composition of steel slag

At present, the influence of important chemical compositions variables of aggregates on the prediction performances of pavement asphalt mixtures in machine learning (ML) has been rarely studied, and the opacity of the prediction process limits its application in practical engineering. Hence, an interpretable model considering the chemistry composition of steel slag (SS_CC) is proposed to resolve the aforementioned problems. The K-fold cross-validation, statistical indicators and Shapley additive exPlanations (SHAP) model are utilized to construct, evaluate and explain models. The results are as follows: the predicted pavement performances of the XGBoost model is superior to that of the LightGBM, CatBoost and RF models. The R² with SS_CC in predicted pavement performances have significant improved compared with the situation without SS_CC. In addition, the more important characteristic variables in pavement performances are per cent passing sieve 2.36 mm (P_2.36), penetration (P), calcium oxide (CaO), steel slag replacement ratio of natural aggregate (SRNA), softening point (SP) and asphalt content (AC), respectively. The favorable influence tendencies and range of the interaction between variables explain the impact on pavement performances in models in line with engineering practice. It has great potential for providing reliable visual optimization design for different pavement performances in engineering application.