Interpretable machine learning framework for predicting cement adhesive bond strength in NSM FRP systems using differential evolution and SHAP analysis

This study presents a robust and interpretable machine learning framework for predicting the average bond strength ( τ avg ) of cement-based adhesives in Near-Surface Mounted (NSM) FRP systems. Four state-of-the-art ensemble algorithms were evaluated on a dataset of 150 experimental tests using a rigorous nested cross-validation protocol and Differential Evolution for hyperparameter optimization. A key aspect of the framework is the explicit engineering of interaction terms to model physical synergies. The optimized LightGBM model emerged as the superior predictor, achieving a mean R² of 0.42 on unseen test data while fitting the training data with an R² of 0.79. This contrast reflects the high heterogeneity of the aggregated dataset. A SHAP analysis revealed that engineered interaction terms were the most influential predictors across all models, with the interplay between groove depth and surface treatment ( d g · Treatment) consistently highly ranked. For the best-performing LightGBM model specifically, the synergy between surface treatment and FRP tensile strength ( f frp ) proved most dominant. These findings highlight that the predictive power of primary variables is significantly enhanced when their complex interdependencies are explicitly modeled. This work provides a transparent, methodologically rigorous framework that serves as both a predictive tool and an instrument for scientific insight into NSM FRP bond mechanics.