Physics-guided interpretable machine learning for predicting recycled aggregate concrete strength

Le résumé fourni par la source

Recycled concrete aggregate (RCA) provides a sustainable alternative to natural aggregate, but variability in RCA quality limits reliable strength prediction. This study evaluated six machine learning algorithms using 520 experimental records from 52 published studies. Physics-guided feature engineering incorporated an Abrams-law water-to-cement ratio transformation, an RCA quality index, and two interaction terms. Under standard 10-fold cross-validation, AdaBoost achieved the best performance (R 2 = 0.902; RMSE = 5.041 MPa). However, study-grouped cross-validation revealed substantially weaker generalisation to unseen studies (R 2 = −0.101 and RMSE = 11.526 MPa), with similar deterioration for Random Forest and Gradient Boosting. These results indicate that the high row-level accuracy primarily reflects interpolation within the compiled database rather than robust cross-study generalisation. The best-performing MLP configuration achieved R 2 =0.768, while multiple linear regression achieved R 2 =0.525. SHAP analysis identified the water-to-cement ratio and RCA quality as the most influential factors within the available ratio-based feature space.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.