Explainable machine learning-based investigation for influential parameters towards interfacial capacity of spliced glass fiber reinforced polymer bars

Splicing glass fiber reinforced polymer (GFRP) bars is commonly encountered in practical applications due to limitations associated with rebar length. In turn, this affects the bond stress maintaining the equilibrium between concrete and reinforcement. Also, there is a lack of estimating the bond strength of spliced GFRP-reinforced beams employing machine learning approaches and interpreting the effective parameters. Thus, this paper predicts interfacial capacity of spliced GFRP-reinforced beams using machine learning (ML) models including ordinary least square regression (OLS), artificial neural network (ANN), and extreme gradient boosting algorithm (XGBoost), and compares those models against the established models. A database compromising 120 samples was collected from relevant experimental studies conducted on beam tests. XGBoost model gave the best performance out of the remaining ML models with an R2 of 0.97 for the training set and an R2 value of 0.93 for the testing set, while the comparative models failed in comparison. Also, a refined investigation of the parameters that influence bond strength was carried out through an explainable artificial intelligence (xAI) approach, specifically SHapley Additive exPlanations (SHAP) by employing XGBoost model. Based on SHAP analysis, bar tensile strength, splice length of the rebar, and bar diameter were the most influential features on the bond resistance. Accordingly, a proposed model to estimate the bond strength of spliced GFRP bars as a function of most contributing parameters was developed and proposed to compute the bond strength in case of spliced GFRP bars. As a result, the proposed equation demonstrated an acceptable performance to predict in comparison with the existing empirical models.