Deep Learning segmentation with metal intrusion for quantitative microstructure analysis of hardened cement paste

Area of Science:

• Materials Science
• Civil Engineering
• Computational Science

Background:

• Quantitative microstructure analysis of hardened cement paste is vital for material performance but hindered by its complex, disordered structure.
• Traditional methods struggle to accurately differentiate and quantify various phases within the cement matrix.

Purpose of the Study:

• To develop and validate an advanced approach for enhanced microstructure analysis of cement composites.
• To improve the quantitative assessment of pore and solid phases in hardened cement paste.

Main Methods:

• Injected Field metal, a low-melting-point alloy, into cement samples under pressure to improve phase contrast in backscattered electron (BSE) imaging.
• Utilized deep learning segmentation models, specifically U-Net and LinkNet, to segment pore, unhydrated, and hydrated cement phases in BSE images.
• Performed quantitative size and shape analysis (area, diameter, solidity, circularity, aspect ratio) on segmented phases.

Main Results:

• The combined metal intrusion and deep learning approach enabled effective segmentation of different phases in hardened cement paste.
• Both U-Net and LinkNet achieved high segmentation accuracy, with mean IoU scores of 0.89 and 0.87, respectively.
• U-Net demonstrated superior performance, capturing finer details and more complex boundaries for precise phase analysis.

Conclusions:

• This novel methodology simplifies cement composite microstructure analysis, allowing for detailed quantitative assessment of individual phases.
• The accurate quantification of microstructure parameters is essential for understanding material behavior and optimizing cement composite performance.