Forecasting interfacial bond strength in FRP-reinforced concrete using soft computing techniques

Fiber-reinforced polymer (FRP) retrofits have become widely used to strengthen concrete and masonry structures because of their properties including lightweight, high strength, high elastic modulus, and corrosion resistance. However, debonding failure at the FRP-substrate interface remains a challenge that undermines the effectiveness of these systems. To advance predictive modeling of interfacial bond strength (IBS), this study aims to develop machine learning (ML) approaches using a comprehensive database of experimental shear pull-out test results. The collected database consists of 855 samples encompassing a wide range of parameters, including: concrete substrate width (100–500 mm), concrete compressive strength (16–74.67 MPa), FRP material Young's modulus (23.9–425 GPa), FRP thickness (0.083–2 mm), FRP width (10–150 mm), and bonded FRP length (20–400 mm). The interfacial pull-out forces ranged from 2.4 kN to 54.79 kN. The dataset was divided into 571 training samples and 284 testing samples. Various ML algorithms, including Random Forest, Random Tree, Stochastic-Random Forest, Stochastic-Random Tree, Bagging-Random Forest, and Bagging-Random Tree, are evaluated and compared based on their performance. The stochastic-random forest technique demonstrates the highest accuracy with correlation coefficients of 0.9949 and 0.9779 for the training and testing datasets, respectively. It also achieves the lowest mean absolute error of 0.41 and 1.45, and the lowest root mean squared error of 1.07 and 2.10 for the training and testing datasets, respectively. The Nash-Sutcliffe efficiency values are 0.99 and 0.96 for the training and testing datasets, and the comparative measure gives a value of 2.22. This research provides an optimized ML framework for reliably forecasting FRP-substrate bond strength to support the design of strengthened structures.