Prediction of Rapid Chloride Penetration Resistance of Metakaolin-Based High Strength Concrete: Critical Comparison between Natural and Evolutionary Algorithm–Optimized Artificial Neural Networks

Abstract

This research presents a significant hybrid artificial neural network (ANN) model by comparing the performance of six optimization techniques: artificial bee colony (ABC), genetic algorithm (GA), grey wolf optimizer, Harris hawks optimization (HHO), particle swarm optimization, and salp swarm optimization (SSO) in optimizing ANNs for predicting chloride penetration in metakaolin concrete. Each model was trained and tested using 161 and 40 data sets, respectively, collected from the literature. The comparison revealed that the SSO_ANN model achieved superior performance, with a performance index of 1.8510 and an accuracy exceeding 97%. In contrast, the variance inflation factor analysis indicated problematic multicollinearity in the binder and fine aggregate variables. Consequently, the GA_ANN, HHO_ANN, and ABC_ANN models performed poorly, with a performance of less than 0.95. Furthermore, the score, uncertainty, and objective function criteria confirmed the SSO_ANN model’s superiority in predicting chloride penetration. Validation using an additional 20 data sets demonstrated the SSO_ANN model’s robust performance, achieving a score of 0.9879. Finally, shapley additive explanations (SHAP) analysis identified compressive strength, fine aggregate, and binder as the most significant features influencing chloride penetration prediction.