An Uncertainty-Aware Explainable AI Framework for Sustainable Concrete Mixture Design and Decision Support

Sustainable concrete mixture design requires reliable strength prediction, robust generalization, quantified uncertainty, and transparent environmental and economic assessment. This study presents an integrated, uncertainty-aware explainable AI framework combining family-aware validation, SHAP interpretation, applicability diagnostics, and multi-objective optimization. Following systematic auditing, 2623 records were retained from an initial dataset of 2700. Historical nested group-aware XGBoost validation achieved R2 = 0.857 ± 0.039 and RMSE = 5.386 ± 0.635 MPa. External evaluation employed 102 overlap-screened records from an independently sourced dataset of 427 observations, yielding R2 = −1.015 and RMSE = 11.115 MPa for the frozen model. External prediction-interval coverage reached 36.27% and 47.06% at nominal 90% and 95% levels, highlighting transferability and uncertainty-calibration challenges. A reproducible NSGA-II procedure incorporating absolute-volume constraints generated 200 candidate mixtures, accompanied by 35 sensitivity runs examining optimization and material assumptions. Carbon estimates employed documented emission-factor proxies, while cost assessment incorporated explicit procurement assumptions. The laboratory assessment of six concrete mixtures, supported by stamped reports documenting mixture compositions and specimen-level compressive strengths, yielded RMSE = 5.765 MPa. By integrating predictive modeling, interpretability, uncertainty, and engineering constraints, the framework provides a transparent basis for preliminary strength–carbon–cost screening and risk-informed sustainable concrete mixture design.