Machine Learning‐Based Performance Prediction and Mix Design Optimization of OGFC‐13 Mixtures

ABSTRACT

The valorization of steel slag, a major industrial solid waste, in open‐graded friction course (OGFC) mixtures holds significant promise for alleviating natural aggregate shortage and reducing environmental burden. However, the coupled influence of steel slag replacement rate and gradation parameters on the permeability coefficient and unconfined compressive strength (UCS) of OGFC‐13 remains insufficiently quantified, and mix design optimization still relies on empirical trial‐and‐error approaches. This study aims to develop a data‐driven framework for predicting the performance indices and optimizing the mix design of steel slag OGFC‐13. Based on 200 sets of laboratory experimental data comprising 40 gradations and 5 steel slag replacement rates, Pearson's correlation analysis was employed to screen four key design variables: steel slag replacement rate, and the passing percentages of the 13.2–9.5 mm, 2.36–0.075 mm, and < 0.075 mm sieve fractions. Four machine learning algorithms, namely M5P model tree, Random Forest (RF), Gradient Boosting (GB), and XGBoost, were constructed and compared for predicting UCS and permeability coefficient. The results demonstrate that the XGBoost model achieved relatively better prediction performance for UCS, with coefficients of determination ( R ² ) of 0.793 and 0.773 for the training and testing sets, respectively. Among the four models, the Random Forest model provided relatively better prediction performance for the permeability coefficient, with a testing R ² of 0.577 and a comparatively concentrated residual distribution. Based on the integrated scoring of XGBoost and Random Forest surrogate models, the optimal mix design achieving a balance between drainage performance and bearing performance was recommended as follows: steel slag replacement rate of 50%, 13.2–9.5 mm fraction of 43.8%, 2.36–0.075 mm fraction of 9.8%, and < 0.075 mm fraction of 3.6%. The predicted UCS and permeability coefficient of this mix are 5.59 MPa and 0.276 cm/s, respectively. The proposed data‐driven prediction and optimization approach can provide a useful reference for the mix design of steel slag OGFC‐13 within the investigated material and gradation ranges.