Automated analysis of grain morphology in TEM images using convolutional neural network with CHAC algorithm

The shape and size of grains significantly impact the properties of polycrystalline materials. In particular, high temperature and radiation exposure in nuclear reactors can lead to considerable grain growth of UO₂, thereby substantially modifying the fuel performance. Transmission electron microscopy (TEM) has been used to characterize the grain morphology of polycrystalline materials, but the manual analysis of TEM images is a time-consuming and labor-intensive process, which cannot meet the increasing demand for high-throughput data analytics. This study presents an automated approach we developed for characterizing grain morphology recorded in bright field TEM images during ion irradiations performed in situ. Our approach combines a machine learning model for detecting grain boundaries and a computer vision algorithm named CHAC for selecting well-labeled grains for statistical analysis. Using TEM images acquired from in-situ ion irradiation experiments on nanocrystalline UO₂, we demonstrate that this automated approach can achieve comparable results to human experts while significantly reducing the analysis time. Moreover, the machine learning model functions as a "few-shot" model, requiring only a modest number of training images to perform effectively on a specific task. Consequently, researchers in need can efficiently train their own models following the procedures described in this study to automate grain morphology analysis of their own TEM images.