Reliable chloride permeability prediction in SCM-blended concrete: A conformal approach for risk-aware mixture design

Predicting the long-term durability of concrete structures exposed to chloride ingress is crucial yet remains challenging due to the complex interplay of factors in concrete containing supplementary cementitious materials (SCMs). This study presents a novel machine learning framework for predicting chloride permeability with robustness and reliability, explicitly addressing inherent uncertainties to inform risk-based concrete design. By combining Bayesian-optimized ensemble learning (XGBoost and CatBoost) with conformal prediction, we develop models that not only achieve high point prediction accuracy but also provide rigorously quantified prediction intervals. Trained and validated on a curated dataset of 204 SCM-blended concrete mixtures, this framework demonstrates satisfactory performance in predicting rapid chloride permeability test (RCPT) results. Findings show that (i) both XGBoost and CatBoost deliver reasonable point prediction accuracy, with CatBoost achieving an R² score of 0.862 on the test set; (ii) conformal prediction generates well-calibrated prediction intervals, capturing true values within the predicted range for over 93% of cases; (iii) integration of conformal prediction improves classification performance, reaching F1 scores above 0.98 for permeability class determination; and (iv) SHAP analysis identifies silica fume content as most influential factor governing chloride permeability. To facilitate practical application, a user-friendly Streamlit web application was developed, which empowers engineers to make real-time, uncertainty-aware predictions, promoting data-driven decision-making in SCM-blended concrete design. This research should offer a powerful and interpretable methodology for enhancing concrete durability, minimizing life-cycle costs, and advancing sustainable construction practices.