Machine learning for high-fidelity prediction of cement hydration kinetics in blended systems. (October 2021)

• Record Type: Journal Article
• Authors: Cook, Rachel; Han, Taihao; Childers, Alaina; Ryckman, Cambria; Khayat, Kamal; Ma, Hongyan; Huang, Jie; Kumar, Aditya
• Is Part Of: Materials & design. Volume 208(2021)
• Publication Date: 2021-10
• Subjects: Machine learning -- Random forests -- Portland cement -- Hydration -- Mineral additives; Materials -- Periodicals; Engineering design -- Periodicals; Matériaux -- Périodiques; Conception technique -- Périodiques; Electronic journals; 620.11
• DOI: 10.1016/j.matdes.2021.109920
• Languages: English
• ISSNs: 0264-1275

Abstract

Graphical abstract: Highlights: This work contains successful prediction and optimization of Portland cement systems. Novel predictions of heat-evolution profiles were achieved via machine learning (ML). This work offers an original dataset, which contains results for 300+ unique entries. The database considers mixture design and physiochemical features as attributes. This work can be expanded to formulate mixture design based on user kinetic-criteria. Abstract: The production of ordinary Portland cement (OPC), the most broadly utilized man-made material, has been scrutinized due to its contributions to global anthropogenic CO2 emissions. Thus — to mitigate CO2 emissions — mineral additives have been promulgated as partial replacements for OPC. However, additives — depending on their physiochemical characteristics — can exert varying effects on OPC's hydration kinetics. Therefore — in regards to more complex systems — it is infeasible for semi-empirical kinetic models to reveal the underlying nonlinear composition-property (i.e., reactivity) relationships. In the past decade or so, machine learning (ML) has arisen as a promising, holistic approach to predict the properties of heterogeneous materials, even without an across-the-board comprehension of the underlying composition-properties correlations. This paper describes the use of a Random Forests (RF) model to enable high-fidelity predictions of time-dependent hydration kinetics of OPC-based systems — more specifically [OPC + mineral additive(s)] systems — using the system's physiochemical attributes as inputs. Results show that the RF model can also be used to formulate mixture designs that satisfy user-imposed kinetics-related criteria. Lastly, the presented results can be expanded to formulate mixture designs that satisfy target kinetic criteria, even without knowledge of the underlying kinetic mechanisms.