Predicting concrete-to-concrete interfacial shear strength using explainable machine learning

This study advances shear strength prediction at the concrete-to-concrete interfaces by integrating ensemble learning with model explainability to enhance accuracy, trust, and interpretability. Fourteen models were developed employing seven ensemble algorithms across two scenarios with varying input features, including engineered features inspired by code-specified equations. A comprehensive dataset of 446 test samples encompassing 11 features, curated under the supervision of American Concrete Institute (ACI) Committee 445, was used for training and evaluation. Among the tested approaches, models employing engineered features achieved superior performance, with Categorical Boosting (CatBoost) emerging as the best algorithm. The CatBoost model achieved a mean absolute error (MAE) of 0.798 MPa, root-mean-square-error (RMSE) of 1.308 MPa and coefficient of determination ( R 2 ) of 0.750, representing a ≈ 26% improvement over code-specified equations. Furthermore, the ensemble model outperformed Neural Additive Models by 6% when trained on the same dataset. Feature importance analysis revealed that reinforcement crossing the interface, including reinforcement ratio and yield strength, accounted for approximately 43% of prediction influence, followed by the combined effects of friction coefficient, reinforcement inclination, and compressive strength. SHapley Additive exPlanations (SHAP) analysis provided actionable engineering insights, uncovering critical design parameters. These findings demonstrate the potential of explainable ensemble learning to support data-driven design, improve interface performance, and guide future research in optimizing shear transfer mechanisms.