Automated Grain Boundary Detection for Bright-Field Transmission Electron Microscopy Images via U-Net

Area of Science:

• Materials Science
• Electron Microscopy
• Machine Learning

Background:

• Quantifying microstructures is vital for understanding material properties.
• Identifying grain boundaries in bright-field transmission electron micrographs is difficult due to image contrast.
• Manual tracing is often required, limiting efficiency and scalability.

Purpose of the Study:

• To develop and validate a machine learning-based approach for automated grain boundary detection in bright-field transmission electron micrographs.
• To overcome the limitations of conventional edge detection algorithms and manual analysis.
• To provide a versatile tool for microstructure characterization.

Main Methods:

• Utilized a U-Net convolutional neural network trained on bright-field images and manual tracings.
• Implemented targeted postprocessing algorithms to refine boundary detection and preserve fine features.
• Validated the method by comparing identified microstructural markers and grain size distributions with manual analysis.

Main Results:

• The machine learning pipeline accurately estimates grain boundary positions, distinguishing them from intragrain contrast.
• Low-contrast grain boundaries were successfully identified.
• Statistical tests confirmed significant overlap between machine learning-derived and manually derived grain size distributions.
• The method demonstrated versatility by successfully analyzing new microstructures of platinum and palladium.

Conclusions:

• The developed machine learning approach offers an effective and automated solution for grain boundary detection in bright-field transmission electron micrographs.
• This technique enhances the accuracy and efficiency of microstructure quantification.
• The approach shows promise for broad application in materials science research.