TimeMan Seminar - Thomas BILYK

Application of deep learning for the recognition of dislocation loops in a high entropy alloy

The characterization of nano-sized objects, detected by transmission electron microscopy (TEM), is essential to evaluate the mechanical behavior of materials. These objects, visible by phase contrast (nanobubbles) or diffraction contrast (dislocation loops or coherent precipitates), and often numerous on a micrograph, are prime candidates for automation. A large amount of data is produced and the problem of their analysis arises.

Manual analysis is meticulous, repetitive, prone to human error and time consuming. Semi-automatic methods, e.g. thresholding, are a first approach. They present several limitations such as an arbitrary choice of parameters that limit transferability, as well as particular post-processing steps to individualize superimposed objects. A second approach is provided by artificial intelligence, which is particularly well adapted to these problems of large numbers of data. Convolutional neural networks (CNN) have revolutionized the field of computer vision over the last decade. The idea is to highlight characteristic features in an image by correlating pixels with their surrounding pixels. The use of a Mask R-CNN allows to perform both detection and segmentation of the studied objects.

This type of approach will be presented in a first step, followed by an application to the segmentation of dislocation loops on STEM-BF micrographs of high-entropy alloys after ion irradiation. The training is based on the annotation of micrographs. After this, the Mask R-CNN can be used for the detection of dislocation loops). Finally, the analysis of the obtained masks allows individual and collective characterization of the objects. As will be discussed, the radial distribution function of the dislocation loops, supports in particular a low mobility of the loops after their formations. The approach is generalizable to other cases and further work is in progress.