Machine learning approaches to predict the micromechanical properties of cementitious hydration phases from microstructural chemical maps. (30th December 2020)

• Record Type: Journal Article
• Authors: Ford, Emily; Kailas, Shankar; Maneparambil, Kailasnath; Neithalath, Narayanan
• Is Part Of: Construction & building materials. Volume 265(2021)
• Publication Date: 2020-12-30
• Subjects: Machine learning -- Nanoindentation -- Modulus -- Chemical mapping -- Microstructure -- Cement pastes -- Ultra-high performance concrete; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2020.120647
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Machine learning models to link chemical information to phase mechanical properties. Use easily obtainable imaging/chemical maps as inputs. Demonstrates the applicability (or lack thereof) of ML models to complex microstructures. Suggests potential options to overcome inaccuracies in multiple-material binders. Abstract: This paper demonstrates the use of normalized intensities of chemical species obtained from energy-dispersive X-ray spectroscopy (EDS) as inputs to machine learning (ML) models, in order to predict the nanoindentation moduli (M) of different phases in a cementitious matrix. Single and multi-component blends belonging to conventional and ultra-high performance (UHP) pastes are evaluated using a variety of ML models. It is shown that the relative intensities of Ca, Si, and Al can be used to accurately predict the phase moduli in well-hydrated pastes with limited microstructural complexities, using all the ML models investigated. When data sets belonging to multiple binders or those for UHP pastes consisting of multiple materials and low degrees of reaction are considered, the accuracy of ML predictions are found to be significantly lower. This is partly attributable to the presence of mixed phases with widely differing chemistry-property relationships, and the lack of data for higher stiffness phases that exaggerate the skew-sensitivity of ML models like ANN. Potential data augmentation strategies to tide over some of these effects are suggested.