Study-Disjoint Machine Learning for Concrete Compressive Strength

Abstract

This study developed and evaluated machine-learning models for predicting 28-day cylinder compressive strength of concrete using a literature-derived dataset of 1,456 mixtures collected from 136 source studies. Eight recorded mixture quantities were used as the original inputs. Source-study identity and cross-study duplicate relationships were incorporated into the validation design to prevent linked observations from appearing on both sides of a train-test split. Seven prediction approaches were assessed using repeated nested grouped cross-validation, with model selection and hyperparameter tuning performed only within the training partitions. ExtraTrees gave the lowest mean outer-fold error among the evaluated model families, with an MAE of 9.758 MPa, RMSE of 12.884 MPa, and study-macro-MAE of 10.392 MPa. Conventional random validation produced substantially more optimistic performance than study-disjoint validation; for ExtraTrees, R² decreased from 0.844 to 0.471 and MAE increased from 4.783 to 9.807 MPa. Applicability-domain assessment showed higher errors for observations classified outside the represented training domain. Permutation importance identified the water-to-binder ratio as the strongest model-specific predictor. Group-aware prediction intervals achieved approximately 91% empirical coverage for a nominal 90% level. The results show that study-disjoint validation provides a more rigorous estimate of model transferability across literature-derived concrete datasets than conventional random splitting.