Estimating the surface chloride concentration of marine concrete utilizing advanced hybrid machine learning models

Hybrid machine learning (ML) models exhibit enhanced accuracy relative to both ensemble and individual models. In this study, advanced hybrid ML approaches were employed to create a reliable model for estimating surface chloride concentration (CC) in marine concrete, eliminating the necessity for labour-intensive and expensive physical experiments. The study combined artificial neural networks (ANN) and support vector regression (SVR) with metaheuristic optimization algorithms, specifically grey wolf optimization (GWO) and gorilla troops optimization (GTO), resulting in the development of four innovative hybrid ML models. Additionally, the integration of partial dependence plots and SHapley Additive exPlanations (SHAP) values offered profound insights into the critical variables influencing surface chloride concentration, advancing both the interpretability and precision of the models. These data sets were sourced from a wide range of literature. All the models exhibited strong performance, with SVR-GWO proving to be the optimal selection. Notably, SVR-GWO demonstrated the highest coefficient of determination (R²) of 0.96, highlighting its exceptional forecasting accuracy relative to SVR-GTO (0.95), ANN-GWO (0.92), and ANN-GTO (0.90). Among the features examined, fine aggregate emerges as the most influential, followed by exposure type, exposure time, and chloride content. Additionally, a user interface has been designed to allow users to enter key inputs and seamlessly obtain CC predictions.