Physics-Guided Machine Learning for Reliable Compressive Strength Prediction of Supplementary Cementitious Materials (SCM) Concrete

Traditional machine-learning approaches for predicting compressive strength of concrete containing supplementary cementitious materials are often evaluated using random data splitting, despite databases containing clusters of highly similar mixtures from the same study. This study examines how such dependence inflates performance and whether physics-guided learning improves generalization and consistency. A database of 1,451 unique mixtures from 137 studies was grouped by shared study origin or identical proportion vectors, yielding a study-disjoint hold-out of 286 mixtures and grouped nested cross-validation. Three unconstrained gradient-boosting baselines were compared with models using monotonic constraints on water-to-binder ratio (w/b) and cement content or an Abrams-law residual formulation. Random cross-validation produced R² of 0.84-0.86, compared with 0.46-0.48 under grouped validation, indicating an apparent R² inflation of approximately 0.38–0.40 under random splitting. Physics guidance did not materially improve accuracy but significantly improved physical consistency. Unconstrained models violated the expected strength-w/b relationship in 15.9-32.5% of grid steps and in every sampled context, whereas constrained models were violation-free; the best-extrapolating model was constrained (R² = 0.554 vs 0.427- 0.508). In contrast, the Abrams hybrid achieved the highest hold-out accuracy (R² = 0.683) but failed under extrapolation (R² = -0.026). SHAP attributed the largest effects to w/b, cement, and silica fume in physically expected directions, despite locally non-physical predictions. Prediction errors of 11.3–11.7 MPa RMSE in grouped cross-validation and 12.7–14.1 MPa on the hold-out suit these models to comparative screening of trial mixtures rather than strength compliance. Study grouped validation, physical-consistency testing, and extrapolation assessment are therefore essential for establishing the reliability of data-driven SCM concrete strength models.