Bayesian optimization-driven machine learning for compressive strength prediction of hybrid fiber-reinforced recycled aggregate concrete

The use of recycled coarse aggregate (RCA) in concrete provides a sustainable approach to reducing construction and demolition waste, but predicting its compressive strength (CS) is challenging due to complex interactions among binder composition, supplementary cementitious materials (SCMs), fibers, and mix proportions. This study develops and compares seven Bayesian optimization (BO) based machine learning (ML) models to predict the CS of hybrid fiber-reinforced recycled aggregate concrete (HFRRAC). A dataset of 281 experimental samples with 14 input variables, including OPC, FA, GGBS, water content, aggregate proportions, and fiber characteristics, was used for model training and evaluation. Performance was assessed using R 2 , RMSE, MAE, and MAPE. The results show that ensemble and boosting methods outperform the conventional approaches evaluated in this study. On the held-out testing subset, the best-performing BO-GP-GBQR model achieved a training R 2 of 0.9959 and a testing R 2 of 0.9348, with a testing RMSE of 3.5675 MPa, MAE of 2.4883 MPa, and MAPE of 8.03%. SHAP analysis identified water content, OPC, and GGBS as the most influential parameters. The proposed framework enables accurate and interpretable prediction of 28-day CS and can support preliminary mixture screening for HFRRAC containing recycled coarse aggregate and hybrid fibers.