Hybrid Artificial Neural Network Models for Predicting Flexural Strength of FRP-Reinforced Concrete Beams

Artificial Neural Network (ANN) models rely on gradient descent algorithm which is sensitive to the values of initial weights and its occasional premature convergence to local minima. Metaheuristics are global optimization algorithms capable to capture the non-linear behavior of structural problems. The present study explores the application of novel hybrid algorithms in structural engineering by predicting the flexure strength of concrete beams incorporating fiber reinforced polymer (FRP) bars. Novel hybrid ANN based algorithms such as grey wolf optimizer (ANN-GWO), ant lion optimizer (ANN-ALO), Harris hawk optimizer (ANN-HHO), salp swarm (ANN-SSA), whale optimization (ANN-WOA) and particle swarm optimization (ANN-PSO) were trained and validated based on the experimental flexure strength of FRP concrete beams. Based on the results of eight distinct statistical indices and accuracy matrix, ANN-ALO model demonstrated the best performance across multiple evaluation metrics such as highest R2 values of 0.9487 and 0.9482 and the lowest RMSE values of 0.0478 and 0.0539 during the training and testing phases, respectively. Furthermore, Taylor diagrams also illustrated that the ANN-ALO model achieved the highest correlation coefficient (R > 0.94) during both the training and testing phases. Finally, predicted to experimental (P/E) ratios were computed for both ANN-ALO model and ACI code formulations for concrete beams incorporating fiber reinforced polymer bars. Resultantly, strong agreement was observed between the flexural capacity estimates given by the ANN-ALO model as majority of the P/E ratios were within ±10% of the ACI formulations. These findings confirmed the reliability, resource conservancy and time efficiency of novel hybrid ANN-ALO algorithm to model complex non-linear problems in the field of structural engineering.