Prediction of concrete compressive strength using machine learning: A mixture-grouped comparison of Random Forest, XGBoost, Hybrid, and Stacking models

Predictive evaluation of concrete mixtures can be inflated when observations from the same mix at different curing ages appear in both training and test sets. We re-examined the public UCI concrete compressive-strength dataset using a split at the mixture level. After removing 25 exact duplicate records, 1,005 observations represented 428 distinct mix compositions. Eighty-six complete mixes (206 observations) were held out; the other 342 mixes (799 observations) were used for grouped hyperparameter selection and model fitting. Random Forest (RF), XGBoost, an inverse-error weighted RF–XGBoost hybrid, and a ridge-based stacking model were compared with a linear ridge model and a median baseline. On the held-out mixes, XGBoost achieved R² = 0.898, MAE = 4.155 MPa, and RMSE = 5.745 MPa. Stacking achieved RMSE = 5.812 MPa and RF achieved 6.655 MPa. A group-bootstrap 95% interval for the XGBoost test RMSE was 4.822–6.732 MPa; the difference between stacking and XGBoost was small and uncertain (paired interval −0.123 to 0.243 MPa). Across 30 mixture-grouped partitions in which every model was re-tuned, mean test RMSE was 5.78 MPa for XGBoost and 5.77 MPa for stacking, whereas the identical procedure with record-wise partitions gave 3.94 MPa for XGBoost, quantifying the optimism of ignoring mixture identity. Errors were largest in the sparse ≥60 MPa stratum (MAE 8.167 MPa), where predictions were biased low. The results support nonlinear prediction within this dataset while showing that evaluation must keep related observations together. Independent laboratory datasets remain necessary to assess transfer across material sources and production settings.