TY - JOUR
T1 - Machine learning on microstructural chemical maps to classify component phases in cement pastes
AU - Ford, Emily
AU - Maneparambil, Kailasnath
AU - Neithalath, Narayanan
N1 - Funding Information:
This study was partly supported by U.S. National Science Foundation (CMMI: 1727445). EF acknowledges the Dean’s Fellowship from the Ira A. Fulton Schools of Engineering at Arizona State University.
Publisher Copyright:
© 2021 The Authors. Published by Pouyan Press.
PY - 2021
Y1 - 2021
N2 - This paper implements machine learning (ML) classification algorithms on microstructural chemical maps to predict the constituent phases. Intensities of chemical species (Ca, Al, Si, etc.), and in some cases the nanomechanical properties measured at the corresponding points, form the input to the ML model, which predicts the phase label (LD or HD C-S-H, clinker etc.) belonging to that location. Artificial neural networks (ANN) and forest ensemble methods are used for classification. Confusion matrices and receiver-operator characteristic (ROC) curves are used to analyze the classification efficiency. It is shown that, for complex microstructures such as those of ultra-high performance (UHP) pastes, the classifier performs well when nanomechanical information augments the chemical intensity data. For simpler systems such as well-hydrated plain cement pastes, the classifier accurately predicts the phase label from the intensities of Ca, Al, and Si alone. The work enables fast-and-efficient phase identification and property forecasting from microstructural chemical maps.
AB - This paper implements machine learning (ML) classification algorithms on microstructural chemical maps to predict the constituent phases. Intensities of chemical species (Ca, Al, Si, etc.), and in some cases the nanomechanical properties measured at the corresponding points, form the input to the ML model, which predicts the phase label (LD or HD C-S-H, clinker etc.) belonging to that location. Artificial neural networks (ANN) and forest ensemble methods are used for classification. Confusion matrices and receiver-operator characteristic (ROC) curves are used to analyze the classification efficiency. It is shown that, for complex microstructures such as those of ultra-high performance (UHP) pastes, the classifier performs well when nanomechanical information augments the chemical intensity data. For simpler systems such as well-hydrated plain cement pastes, the classifier accurately predicts the phase label from the intensities of Ca, Al, and Si alone. The work enables fast-and-efficient phase identification and property forecasting from microstructural chemical maps.
KW - Cement pastes
KW - Chemical mapping
KW - Machine learning
KW - Microstructure
KW - Nanoindentation
KW - Ultra-high performance concrete
UR - http://www.scopus.com/inward/record.url?scp=85118784249&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85118784249&partnerID=8YFLogxK
U2 - 10.22115/SCCE.2021.302400.1357
DO - 10.22115/SCCE.2021.302400.1357
M3 - Article
AN - SCOPUS:85118784249
SN - 2588-2872
VL - 5
SP - 1
EP - 20
JO - Journal of Soft Computing in Civil Engineering
JF - Journal of Soft Computing in Civil Engineering
IS - 4
ER -