Machine learning approaches to admixture design for clay-based cements

Abstract

Replacement of 30% of ordinary Portland cement (OPC) by metakaolin (MK) reduces the CO2 intensity but negatively impacts the workability. A critical challenge facing adoption of this next-generation infrastructure material is developing admixture systems that impart workability similar to unblended OPC while retaining the advantages in strength and environmental stability conferred by MK. Hierarchical machine learning is a highly-supervised methodology that integrates physical and statistical modelling to understand and optimize complex systems. Here it is applied to designing admixture formulations for OPC-MK blends, providing exceedingly rapid admixture development as well as formulations tailored to specific materials. Elucidating how MK impacts workability of these systems was addressed by screening the effects of superplasticizers, viscosity-modifying admixtures, and water-reducing admixtures on pore solution properties, OPC rheology and the colloidal properties of MK suspensions. Changes in slump spread of 70% OPC/30% MK blends as a function of admixture formulation were fit using regression methods. Increases in slump spread were found to be a strong function of pore solution viscosity, effects of superplasticizer on MK zeta potential and electrosteric interactions, and coupling between pore solution viscosity and osmolality with MK zeta potential and electrosteric interactions, respectively. Work toward designing new admixtures that optimize these interactions will also be pursued.