A hybrid parameter-driven analysis-based approach for predicting the interface bond strength between steel bar and machine-made gravel aggregate concrete

Given the progressive depletion of natural sand and gravel resources, the substitution of natural aggregates with machine-made gravel aggregate (MA) has emerged as an unavoidable trend. The disparities in surface morphology, mechanical properties, and chemical composition between MA and natural aggregates have the potential to modify the bond strength of machine-made gravel aggregate concrete (MGC) with steel bars. Based on this, a steel-MGC bond strength prediction model integrating machine learning (ML) algorithm and SHapley Additive exPlanations (SHAP) + Partial dependence plots (PDP) method is proposed, which combines the predictive ability of ML and the interpretable ability of SHAP+PDP method, compensates for the shortcomings of traditional prediction models and solves the black-box problem of ML, and improves the efficiency and accuracy of bond strength prediction. The model prediction performance is systematically verified based on a database of 493 experimental datasets and compared with traditional mathematical models. The results show that the ML model outperforms the traditional analysis methods in terms of prediction accuracy. Among them, the Random Forest model (RF) demonstrated an R2 value of 0.95 and a 20-index value of 0.92, which outperformed other models. In addition, the global and local robustness of the predictions was assessed through the SHAP+PDP interpretability technique while quantifying the importance of the parameters. The analysis indicates that the anchorage length-to-diameter ratio, concrete compressive strength, and steel bar yield strength are the primary factors influencing bond performance. The method proposes a new paradigm for interface characterization and provides theoretical support.