Abstract
GFRP rebar is a desirable alternative to steel rebar especially in harsh environments with advantages of light weight, high strength, stable chemical properties, and corrosion resistance. However, there is a lack of understanding on the durability of GFRP rebar which significantly limits its engineering applications. Unfortunately, it is almost impossible to develop any simple theoretical model to predict the residual strength of GFRP rebar after environmental exposure. To effectively address this research gap, this paper investigates machine learning models to predict the residual tensile strength of GFRP rebar due to environmental degradation. Firstly, a database was built from the literature containing a total of 350 tensile testing results of GFRP rebar after experimental exposure. Six key influencing parameters were considered, including fiber content, bar diameter, resin type, exposure temperature, pH, and aging time. Secondly, machine learning models were trained and tested using the database, and the selected models included Decision Tree, Random Forest, Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Long and Short-Term Memory (LSTM) models. The LSTM model demonstrated the best performance in strength prediction, achieving an R(2) of 0.96 on the training set and 0.91 on the testing set. Thirdly, the influencing parameters were ranked using SHAP and Random Forest in terms of their impact on the residual tensile strength of GFRP rebar, and it was found that temperature has the most significant effect followed by fiber content, exposure time, pH and bar diameter, while resin type showed the least importance. Notably, SHAP analysis also showed that the coupling of different parameters also had impact on the residual strength, and the combined effect of fiber content with other parameters was the most prominent. This paper demonstrates the feasibility of machine learning models in the durability study of GFRP composite materials.