Machine learning-based prediction of crack width and bond-dependent coefficient (kb) in GFRP-reinforced concrete beams

Glass fiber reinforced polymer (GFRP)-reinforced concrete (RC) flexural elements are typically designed to fulfill the serviceability criteria, including deflection and crack width. Crack width control in RC structures enhances durability, increases service life, and improves aesthetic appearance. The design provisions of ACI 440.11-22 and CSA S806-12 incorporate a bond-dependent coefficient, k b , into equations controlling crack width to account for the bond between GFRP bars and concrete. The primary objective of this research is to apply eight machine learning (ML) models, including extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), adaptive boosting (AdaBoost), gradient boosting (GB), random forest (RF), K nearest neighbors (KNN), Ridge, and Lasso, to enhance crack width prediction accuracy in GFRP-RC beams. The study also considers the importance of the k b coefficient in estimating the crack width and investigates the key factors that primarily impact its value. The results indicated that existing design provisions generally overestimate crack width, whereas ML models demonstrate a noticeably closer alignment with experimental data. Additionally, AdaBoost stands out as the most accurate predictor of crack width. SHapley Additive exPlanations (SHAP) analysis highlights that GFRP bar strain, bar spacing, and concrete cover significantly impact the k b coefficient.