Title:
Data Mining HeidelbergMaterials Database: Portland Cement – Limestone Systems Optimization with Machine Learning
Author(s):
Alexandre Ouzia and Mohsen Ben Haha
Publication:
Symposium Paper
Volume:
362
Issue:
Appears on pages(s):
998-1005
Keywords:
blended portland cements, data mining, limestone, machine learning, optimization
DOI:
10.14359/51742025
Date:
6/18/2024
Abstract:
This article reviews the challenges in the rational use of limestone and supplementary cementitious materials in the optimization of low carbon cement and concrete with machine learning (ML), and introduces preliminary results of the corresponding program of research at HeidelbergMaterials.
The mining of the Global R&D database showed that the main challenge was not the algorithm type—the general linear model performed as well as artificial networks—but the underlying dataset quality, the rational design of the experiment in the face of the high dimensionality of the problem, and the model testing methodology.
Preliminary results of show that a clinker ratio as low as 50% can be obtained at equal or better strength and workability performance. The surface area of limestone and aggregates was found to be as important as their weight proportion on rheology and early age properties. Regarding the predictors of early age strength, the best subset selection method identified no less than seven variables in addition to C3S and Blaine fineness. The prediction model thus identified a CEM I composition that could reach 50 MPa in one day, thus paving the way to higher SCM replacement levels.