Assessing the Inference Limits of Routine Cement Measurements for CEM I Performance Using Machine Learning: A Multi‐Producer Study – Extended Abstract

Abstract

Routine cement measurements support quality control, but their information content for performance inference across independent producers is uncertain. This study analyzes 476 records from 23 European producers collected over 27 years, with the main analysis restricted to 311 CEM I cements. Oxide chemistry, Blaine fineness, particle‐size descriptors, density, and derived Bogue and clinker‐modulus descriptors were evaluated using correlation analysis, additive and tree‐based models, feature attribution, and selected producer‐holdout tests. Fineness was the strongest isolated descriptor family, while chemistry contributed a comparable signal when descriptors were evaluated jointly. Strength class reached a balanced accuracy of 0.894 and water demand reached R ² = 0.713. For 28‐day strength on the PSD‐complete subset, the best cross‐validated model reached R ² = 0.638 and MAE = 2.71 MPa; an additive model achieved MAE = 2.80 MPa with 91% of predictions within a prespecified ±10% relative‐error band. Blaine and compact PSD representations gave comparable predictive performance within the studied descriptor space. The N/R early‐strength designation and transfer to unseen producers exposed the inference limit imposed by unresolved mineralogy, sulfate speciation, and process history.