Explainable Feature Selection and Hybrid Computational Intelligence Models for Predicting Chloride Penetration in Concrete

The Rapid Chloride Permeability Test (RCPT) is widely used to assess the durability of concrete against chloride ion ingress; however, this test is time-consuming, costly, and highly sensitive to laboratory conditions. Consequently, developing accurate and reliable predictive models has become an attractive alternative to experimental measurements. In this study, an advanced data-driven framework is proposed for predicting RCPT values by integrating hybrid machine learning models with robust feature selection strategies. A comprehensive database comprising 483 experimental samples was compiled from previous studies. In addition to the primary mix design parameters, a set of ratio-based and interaction features was generated to better capture complex nonlinear relationships among the variables. Feature selection was carried out using two complementary approaches, namely Q-learning–Decision Tree (Q-learning–DT) and NSGA-II–Decision Tree (NSGA-II–DT). The results demonstrate that NSGA-II–DT achieves superior performance by identifying only five key input variables while maintaining high predictive accuracy. Subsequently, four hybrid models—CatBoost–SWOA, NGBoost–WOA, TabNet–EVOA, and TabPFN–HEOA—were developed and evaluated using multiple statistical performance indices and Taylor diagrams. Among the developed models, CatBoost–SWOA exhibited the best predictive capability on the testing dataset, achieving an R 2 value of 0.966 with the lowest RMSE. Uncertainty analysis based on residual bootstrapping further confirmed the robustness and stability of this model, yielding the narrowest confidence intervals and the lowest uncertainty measures. In addition, SHAP-based interpretability analysis revealed that curing temperature and fly ash content are the most influential parameters governing RCPT values. The findings indicate that the proposed framework provides an accurate, reliable, and interpretable tool for RCPT prediction and offers a practical alternative to labor-intensive experimental testing in concrete durability assessment.