Interpretable data-driven modeling of strength evolution in cementstabilized recycled macadam incorporating aggregate physical properties

Abstract

Predicting the unconfined compressive strength (UCS) of cement-stabilized recycled macadam (CSRM) remains challenging because of limited experimental data, heterogeneous aggregate properties, and complex nonlinear relationships among material parameters. This study proposes an interpretable machine-learning framework integrating feature selection, outlier detection, data augmentation, performance prediction, hyperparameter optimization, and model interpretation. A database containing 230 experimental samples was compiled from 10 published studies, and recursive feature elimination and isolation forest were used to refine the feature space and identify anomalous samples. After outlier removal, 211 samples were retained, of which 15 were withheld as an independent validation set. A statistical feature matching loss-enhanced WGAN-GP (SML-WGAN-GP) was then developed to augment the limited development dataset. Compared with conventional WGAN-GP, SML-WGAN-GP reduced the feature-correlation deviation by 36.6%, while data augmentation substantially improved prediction accuracy and cross-validation stability across different machine-learning models. Among the optimized models, SSA-CatBoost achieved the best overall performance, with a test-set R2 of 0.963, MAE of 0.278 MPa, and RMSE of 0.389 MPa. Independent validation using 15 completely withheld experimental samples further yielded an R2 of 0.953, MAE of 0.265 MPa, and RMSE of 0.386 MPa. SHAP analysis identified curing age and cement content as the dominant predictors. Higher crushing values of natural and recycled aggregates were generally associated with lower predicted UCS, whereas recycled aggregate water absorption exhibited a nonlinear relationship with UCS, characterized by an initial increase followed by a decrease. These findings demonstrate that explicitly incorporating aggregate physical properties and data-quality optimization can improve the accuracy, stability, and interpretability of UCS prediction for CSRM.