Hybrid ANN models for durability of GFRP rebars in alkaline concrete environment using three swarm-based optimization algorithms

This study investigates the non-linearity of GFRP degradation in terms of tensile strength retention (TSR) encompassing five input variables (i.e., diameter of GFRP rebar, volume fraction of fibers, pH of surrounding solution, temperature of surrounding solution, and duration of conditioning). Additionally, the degradation models were extrapolated to calculate environmental reduction factor (C_E). Based on high non-linear capabilities of artificial neural network (ANN) models, hybrid ANN models were created by deploying particle swarm optimization (PSO), grey wolf optimization (GWO) and marine predators algorithm (MPA) to predict the TSR of GFRP rebars conditioned at variable temperatures and durations. Trial and error procedure was adopted to select the optimal hyperparameter of the hybrid models. The developed models were subjected to statistical, accuracy, sensitivity and uncertainty analyses. ANN-GWO model yielded the highest correlation (R ≈ 0.89 and 0.88, MAE = 0.0558 % and 0.0914 %, and RMSE = 7.86 % and 11.67 %) for the training and testing datasets, respectively. Furthermore, Taylor diagram and accuracy matrix revealed that the performance of the models followed the order; ANN-GWO (R_overall = 0.885) > ANN-PSO (R_overall = 0.806) > ANN-MPA (R_overall = 0.821). For GFRP rebars, made of E-glass fibers (volume fraction = 0.62) impregnated in vinyl ester resin conditioned at 28 °C with surrounding pH of 13, the values of C_E for bar sizes 9.5 mm, 12.7 mm and 15.9 mm were recorded as 0.75, 0.87 and 0.95, respectively. It is highly recommended to consider the effect of bar size as an important variable in the design of GFRP reinforced concrete structures and keep C_E ≈ 0.75 at 28 °C in alkaline environment.