High-throughput analysis of dislocation loops in irradiated metals using Mask R-CNN.

Area of Science:

• Materials Science
• Computational Materials Science
• Microscopy

Background:

• In situ transmission electron microscopy (TEM) generates extensive data on material defect evolution under extreme conditions.
• Computer vision models, like Mask R-CNN, offer efficient image segmentation capabilities.

Purpose of the Study:

• To develop and validate a workflow for labeling, segmenting, and analyzing irradiation-induced defects in TEM images using Mask R-CNN.
• To assess the model's accuracy and generalization capabilities for quantitative defect analysis in various metallic materials.

Main Methods:

• Implementation of a Mask R-CNN-based workflow for defect segmentation in bright-field TEM videos.
• Training and testing of different Mask R-CNN model architectures and hyperparameters on austenitic stainless steel 316L.
• Validation of defect analysis metrics, including areal density, foreground fraction, particle count, and size distribution.

Main Results:

• The best Mask R-CNN model achieved 83.6% accuracy in predicting defect areal density in 316L.
• Estimated physical metrics (foreground fraction, particle number, size) showed relative errors below 5%.
• The model successfully differentiated defect evolution patterns in 16Cr-37Fe-13Mn-34Ni alloy and pure Cr.

Conclusions:

• The proposed workflow enables consistent, real-time analysis of defects during in situ TEM experiments.
• The method provides quantitative data crucial for refining mesoscale models of material behavior.
• Mask R-CNN is a powerful tool for accelerating defect analysis in materials science research.