Introducing vacancy defects by rapid quenching in FCC metal for deep learning micrograph dataset

Abstract

Accurate detection of crystal defects from microstructure micrographs is essential for evaluating the mechanical and functional properties of metallic materials. Emerging deep learning techniques have great potential for application in crystal defects detection, but the excellent detection accuracy requires domain knowledge and micrograph datasets. Here we target to detect and distinguish nanoscale dislocation loops (DLs) and stacking fault tetrahedrons (SFTs) in transmission electronic microscope (TEM) micrographs. The micrograph datasets of crystal defects derive from copper prepared by rapid quenching and post-annealing. Applying standard U-Net architecture, the maximum Mean Intersection over Union (mIOUmax) of 3-class model is 54.29%, while it is 73.81% and 74.67% of DL 2-class and SFT 2-class model. Results suggest that excellent detection accuracy of deep learning model could be achieved by learning features from specific defect datasets. These findings provide a new strategy for accurate detection of crystal defects in the field of material engineering.