Discrepancies Between Reported and Actual Predictive Performance of Machine-Learning Models for Concrete Compressive Strength: A Validation Study Using Experimental Data

Abstract Recent studies employing machine learning (ML) for predicting compressive strength of concrete have reported high levels of accuracy, attracting research interest. However, many of these models are trained on publicly available or multisource datasets, raising concerns regarding their generalization ability. The present study evaluates the predictive performance of ML models and traditional regression (TR) models proposed in the previous literature, using compressive strength data obtained from laboratory tests. The results show that many of the selected ML models failed to reproduce the high performance reported in their original studies and did not exhibit consistent superiority over TR models. A substantial deviation in prediction values was also observed among the ML models. Despite this, a specific ML model that directly incorporated recycled aggregate properties produced predictions closest to the experimental results, confirming the potential of ML when relevant material characteristics are properly addressed. This study also highlights that many previous works are limited by the closed science practices, stemming from nondisclosure of datasets and inaccessible source code, which restricts research reproducibility. This study discusses potential factors contributing to increased prediction uncertainty in ML models and emphasizes the need for rigorous validation and methodological transparency in future data-driven model development.