Inference Limits of Routine Cement Characterization for CEM I Performance: Evidence From a Multi-Producer Dataset

Abstract

This study evaluates the prediction of 28-day compressive strength from routine cement characterization across independent producers. Industrial records from 23 European producers cover nearly three decades of testing in a single laboratory, encompassing 476 samples. The archive includes CEM I, CEM II, CEM III, and other blended cements, with the main analysis focused on CEM I. Chemistry and fineness descriptors were evaluated using linear and nonlinear models. Strength class and water demand were also examined as performance labels from standardized testing. Mean absolute errors in 28-day strength prediction were approximately 3 MPa, with linear models performing comparably to nonlinear models. Producer holdout tests assessed transfer to cements from producers excluded from training and yielded absolute errors of the same order. Equivalent alkali showed a consistent negative association with 28-day strength, with K${}_2$O carrying most of this signal. Together, these results highlight the value of industrial data collected under common laboratory conditions for interpretable 28-day strength prediction, with potential to inform cement design targeting lower production costs and CO${}_2$ emissions.

arXiv comment17 pages, 6 figures, 12 tables