Materials-guided concrete compressive strength prediction under public-to-plant dataset shift: Retrospective single-plant validation and conformal uncertainty

Reliable prediction of concrete compressive strength is important for mixture design, ready-mix quality control, material acceptance, and low-carbon binder use. Many machine-learning studies report high accuracy on public benchmarks, but benchmark accuracy alone does not show whether a model remains useful for a production plant with local raw material variability. This study develops a materials-guided prediction and reliability framework that combines material-ratio descriptors, raw material quality indicators, tree-ensemble learning, public-to-plant shift diagnosis, few-shot plant calibration, SHAP-based association analysis, response-surface checking, and split conformal prediction intervals. The framework was evaluated on the public UCI concrete compressive strength dataset and a private Wuhu Jianke ready-mix dataset containing 929 deduplicated 28-day records with 42 predictors. A clear public-to-plant mismatch was observed in the common feature space, with RBF-MMD 2 = 0.348 (permutation p = 0.0099 ), mean standardized Wasserstein distance = 1.030, and held-out domain-classifier AUC = 0.951 ± 0.011. After excluding cement grade and 28-day cement-strength descriptors, the Wuhu XGBoost model achieved RMSE = 1.590 ± 0.169 MPa over 15 repeated random hold-out splits and 1.829 ± 0.265 MPa under recipe-grouped validation. Repeated few-shot experiments showed that source-plus-target learning gave comparable mean accuracy and lower subset-to-subset variability when only 14 Wuhu labels were available, whereas target-only calibration became stronger at larger target-label budgets. Global split conformal intervals achieved near-nominal marginal coverage but under-covered high-strength samples; predicted-strength Mondrian calibration improved high-strength coverage to about 0.94 with wider intervals. The results support a cautious conclusion: concrete-strength prediction for plant use should be treated as a materials-guided, locally calibrated, and uncertainty-aware workflow, while broader generalization still requires chronological and independent-plant validation.