An explainable AI framework for crack width and crack spacing prediction with serviceability-oriented design optimization of reinforced and prestressed concrete members

Abstract Reliable prediction of cracking behavior is essential for the serviceability assessment of reinforced and prestressed concrete members, where crack development depends on interacting material, geometric, reinforcement, prestressing, and loading parameters. This study presents an explainable and uncertainty-aware artificial-intelligence framework for predicting crack width and mean crack spacing while supporting serviceability-oriented reinforcement-detailing optimization. A quality-controlled experimental database comprising 19,863 observations from 30 independent experimental programs was transformed into physics-informed engineering features. Five ensemble-learning algorithms were evaluated using five-fold GroupKFold cross-validation to reduce information leakage between experimental programs. Model interpretation, predictive uncertainty, robustness assessment, and multi-objective optimization were incorporated using SHAP, split conformal prediction, Monte Carlo simulation, and NSGA-II, respectively. The optimized LightGBM and CatBoost models achieved R 2 values of 0.808 and 0.672 for crack-width and mean crack-spacing prediction, respectively. SHAP analysis identified reinforcement stress, normalized bending demand, reinforcement ratio, bending moment, and the stress-to-yield-strength ratio among the most influential predictors. Split conformal prediction achieved empirical coverage probabilities of 96.79% and 98.07%, while Monte Carlo simulation indicated that 6.96% of realizations exceeded the 0.30 mm crack-width limit under the investigated perturbation scenario. The representative optimization case reduced the predicted crack width and mean crack spacing by 33.63% and 37.80%, respectively, while increasing reinforcement demand by 46.72%. Within the scope of the compiled experimental database and the adopted validation framework, the proposed framework provides an interpretable and uncertainty-aware engineering decision-support tool for serviceability assessment and reinforcement-detailing optimization. Independent validation using external experimental datasets would further strengthen confidence in its broader engineering application.