Data-driven modeling of bond and anchorage behavior in steel bar-reinforced recycled aggregate concrete beams

A reliable bond between steel reinforcement and recycled aggregate concrete is essential for ensuring composite action and structural integrity. However, flexural pull-out test data that realistically capture this interfacial behavior remain limited, and existing bond strength models often neglect several key parameters. Research on anchorage length design is also relatively scarce. To address these limitations, this study adopts a data-driven framework to compile and analyze available experimental evidence, aiming to develop a reliable predictive model for peak bond strength and to propose an anchorage length formula with appropriate safety margins. A total of 158 test records from ten representative studies were collected under comparable specimen preparation and loading conditions to establish a baseline database. To mitigate data sparsity, advanced data augmentation techniques were employed to expand the dataset to 500 samples while preserving physical boundary constraints, statistical characteristics, and intrinsic correlations, thereby compensating for missing variables and enhancing data diversity. Based on the enriched dataset, five ensemble learning algorithms combined with interpretability tools were used to predict and interpret bond strength, and a reliability analysis was conducted to evaluate anchorage performance. The results indicate that bond strength is primarily governed by the ratio of bar diameter to bond length, cover thickness, recycled coarse aggregate content and quality, matrix strength, stirrup parameters, and recycled fine aggregate properties. All variables exhibit positive correlations except those related to recycled aggregates. Among the models, gradient boosting achieved the highest predictive accuracy without overfitting, while reliability analysis confirmed representative models provide adequate safety margins.