A deep learning approach for complex microstructure inference.

Area of Science:

• Materials Science
• Computer Science
• Metallurgy

Background:

• Automated microstructure inference is crucial for understanding material properties.
• Deep learning (DL) shows promise but lacks clear guidelines for materials science applications.
• Challenges include data requirements, methodology, and DL model transparency.

Purpose of the Study:

• To develop and validate a DL methodology for accurate microstructure segmentation.
• To provide guidelines for data quality, quantity, and training strategies in materials science.
• To enhance the interpretability of DL models in microstructure analysis.

Main Methods:

• A multidisciplinary approach focusing on specimen preparation and imaging.
• Training distinct U-Net architectures using 30-50 micrographs with EBSD-informed annotations.
• Investigating the impact of image context, pre-training, and data augmentation.

Main Results:

• Achieved 90% accuracy in segmenting lath-bainite in complex-phase steel, comparable to expert performance.
• Demonstrated the influence of image context, pre-training, and augmentation on segmentation accuracy.
• Network visualization confirmed plausible model decisions linked to grain boundary morphology.

Conclusions:

• The developed DL approach enables reliable and objective microstructure quantification.
• This methodology addresses key challenges hindering DL adoption in materials science.
• Findings pave the way for advanced materials development through accurate microstructure analysis.