Predictive modeling and parametric influence of fiber reinforced polymer-timber interfacial bond behavior via ensemble learning models and shapley additive explanations

The interfacial bond strength (IBS) between fiber-reinforced polymer (FRP) composites and timber is known to be susceptible to premature failure through debonding at the FRP-timber interface. While the existing research on this topic has predominantly focused on experimental and analytical approaches, the application of more robust machine learning techniques for predicting FRP-timber IBS remains limited. To address this gap, the current study investigates the use of four ensemble learning algorithms for predicting the IBS of FRP-timber interfaces. A comprehensive dataset of 358 FRP-timber samples was curated, encompassing influential material and geometric properties. The performance analysis of the machine learning models over the training and test datasets revealed that the CatBoost algorithm exhibited the strongest ability to understand the relationship between the FRP-timber properties and the IBS. CatBoost yielded the highest accuracy, with coefficient of determination (R 2 ) values of 0.921 and 0.954 for the training and test data, respectively. Further insights were gained through Shapley Additive Explanations (SHAP) analysis, which identified the FRP tensile strength and FRP bond length as the most significant parameters influencing the IBS. The findings of this study demonstrate the potential of CatBoost in accurately predicting the IBS of FRP-timber composites based on geometric and material properties.