Study on the prediction of shear capacity of new and old concrete interfaces based on explainable machine learning algorithms

In prefabricated building structures, slip failure at the interface between new and old concrete can significantly impact the overall load-bearing performance of the structure. Given the complexity of the shear mechanism at the interface, traditional empirical formulas in current standards struggle to accurately capture the nonlinear relationships between various parameters and the shear capacity of the interface. Therefore, developing methods that can accurately assess and predict the shear capacity of the interface is crucial. This study constructed database of shear capacity with 300 specimens, selected key input parameters using Pearson correlation coefficients, and trained six machine learning models (Decision Trees, Random Forest, XGBoost, LightGBM, Artificial Neural Networks, Convolutional Neural Networks) using ten-fold cross-validation and hyperparameter tuning. Additionally, the study compared the predictions of five traditional empirical equations with those from machine learning models and quantified the predictive performance of models and equations using the Coefficient of Variation, the Root Mean Squared Error, the Mean Absolute Error, the Mean Absolute Percentage Error, and the Coefficient of Determination. Analysis was conducted on the machine learning models using SHAPley values and permutation feature importance techniques, with a detailed SHAP feature dependency analysis on the XGBoost model. The results demonstrate that the predictive capabilities of the machine learning models significantly surpass those of the existing traditional empirical equations, with the XGBoost model showing the best performance. Within the XGBoost model, the most contributing feature to the prediction of shear capacity was the defined clamping force, followed by the dimensions of the interface (length and width), the roughening method, and the depth of roughness.