Machine learning prediction of residual tensile strength of glass fiber reinforced polymer bars in alkaline environments

Comprehensive assessment of corrosion degradation of glass fiber-reinforced polymer (GFRP) bars in harsh environments is fundamental to maintain the service life of civil infrastructure. This study based on machine learning (ML) techniques, developed a method to quantify the tensile strength retention (TSR) of GFRP bars immersed in the alkaline environments. A database containing 136 specimens was established for this purpose. The input features included Fiber volume fraction (V_f), Matrix type (MT), FRP bar diameter (d), Tensile strength (TS), Surface characteristics of FRP bars (SC), Exposed time (t), Temperature (Temp), and the pH of the alkaline solution. The correlation coefficients among each input feature and the objective variable were computed to evaluate their relationships. Six representative ML models, including Random Forest (RF), Backpropagation Neural Network (BPNN), Artificial Neural Network (ANN), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT), and Categorical Boosting (CatBoost), were employed to forecast TSR. Among these, CatBoost delivered the strongest predictive performance (R²=0.982, RMSE=2.035, CoV=0.163, Avg=1.004 and IAE=0.016) and was thus adopted for further investigation of influential factors. The analysis ranked the pH as the paramount input feature, subsequent to Temp, TS, SC, MT, d, V_f, and t. By integrating CatBoost with physically interpretable techniques, the researchers devised a practical equation for calculating TSR, mitigating the typical black-box challenge in machine learning. This model demonstrated impressive predictive accuracy (RMSE=5.972, CoV=0.074, Avg=1.017 and IAE=0.054).