Domain Knowledge-Driven Machine Learning for Decision Support in Early Control of CP II-F Cement Strength

The cement compressive strength test requires from 3 to 28 days of curing, generating a lag in quality control. A predictive approach guided by domain knowledge is proposed to estimate the strength of CP II-F cement at 3, 7 and 28 days, from 849 samples from a Brazilian industrial plant. The methodology encompasses: (i) construction of attributes with strict temporal ordering (mitigating data leakage); (ii) coding of physicochemical indicators of the clinker; (iii) selection of compatible attributes for heterogeneous bases; and (iv) diagnosis of the informational limit at 3 days, suggesting operational adjustments. The 7- and 28-day models achieve R² at approximately 0.78 and mean absolute error of less than 1 MPa. The results indicate that the approach mitigates the gap between diagnosis and action, acting as decision support in the early screening of industrial batches.