Accurate identification and characterization of defects in materials enabled by deep learning image analysis

Transmission electron microscopy (TEM) allows detailed study of materials at the atomic scale, revealing defects and structures. Many objects such as voids, second-phase precipitates, stacking fault tetrahedra, and radiation-induced segregation on grain boundaries or dislocation loops etc can be seen in a TEM micrograph, leading to complex and numerous data. Manual analysis is error-prone and time-intensive. While semi-automatic methods have limitations, artificial intelligence offers a tailored solution for analyzing these complex images. For two years, we have been working on a workflow designed for TEM analysis using deep learning tools. The new recruit will create a workflow that incorporates previous work and new advancements. Two main areas are addressed: first, using computer vision and deep learning like Yolo and Mask R-CNN for TEM image analysis to detect, track, follow and classify objects; and second, employing TEM tomography to generate 3D visualizations from multiple angled images, a task made efficient by our novel segmentation approach using a convolutional auto-encoder. Finally, great attention will be given to the database design of the experimental information.