Machine learning-based prediction of bond performance of FRP composite bars in concrete for marine composite structures

Abstract

The bond between Fibre Reinforced Polymer (FRP) bars and concrete degrades under seawater, compromising the structural integrity of FRP-reinforced concrete structures in marine environments. Accurate modelling of this bond behaviour is important for ensuring the reliability of such structures. The objective of this study is to develop and evaluate advanced tree-based machine learning (ML) models, including Extreme Gradient Boosting (XGBoost), M5P, and Random Forest, to accurately predict the bond strength retention and failure modes of FRP-reinforced concrete exposed to seawater. A database of 658 experimental results was collected, considering 14 influential parameters, and used to train and test the models. Despite the inherent variability in durability results, the developed models achieved satisfactory predictive accuracy. Feature contribution analysis identified concrete compressive strength as the most significant factor, followed by conditioning duration and bar surface condition. Lesser contributions came from concrete type, conditioning temperature, bar tensile strength, concrete cover, bar elastic modulus, bar diameter, and fibre type, with minimal impact from sustained load, resin type, bond length, and test type. Compared to Fib Bulletin 40 predictions, the ML models showed good accuracy within the range of available conditioning durations. However, accuracy diminished for marginal durations like 365 days due to limited data, indicating lower extrapolation capability and the need for longer-duration experimental results to enhance predictive performance.