Accurate in-situ appraisal of concrete quality becomes difficult when supplementary cementitious materials (SCMs) and high-range water-reducing admixtures weaken standard non-destructive testing (NDT) correlations. This study builds a compact but information-rich multi-output dataset ( \(n=147\) ) that pairs Ultrasonic Pulse Velocity (UPV) and Rebound Number (RN) with mix descriptors and physically motivated ratios. Six heterogeneous base learners—Ridge, RBF-SVR, Random Forest, Gradient Boosting, GPU-XGBoost, and a shallow MLP—undergo nested cross-validation and then combine through convex stacking with weights proportional to each model’s global mean absolute SHAP value. This SHAP-weighted ensemble aligns predictive accuracy with interpretability in a single, transparent scheme for simultaneous UPV and RN prediction. The approach attains a cross-validated RMSE of 1.58 (mean across targets) and \(R^{2}=0.81\) , outperforming the best single learner by about 11%. Residual diagnostics show near-unbiased fits with no visible heteroscedasticity. SHAP analysis highlights water–cement ratio, binder content, and water–binder ratio as dominant drivers, providing engineer-friendly explanations and what-if insight. This work introduces a novel SHAP-weighted stacked ensemble for the joint prediction of UPV and Rebound Number, where model weights are assigned based on global SHAP importance rather than validation accuracy alone. Unlike conventional approaches that model each NDT output independently, the proposed multi-output framework integrates engineered binder–water ratios and second-order interaction terms, offering a unique, interpretable and data-efficient solution for SCM-rich concretes.
