3D Reconstruction of Polycrystalline Microstructure of Iron-Based Superconductors using Deep Learning

Image analysis to identify the constituent phases from microstructural images is an important topic in understanding the mechanism of functional materials and in applying to informatics techniques. Classic automatic thresholding methods are well known as a phase segmentation tool of functional polycrystalline materials and quantification of microstructural factors. On the other hand, polycrystalline materials often contain characteristic microstructural defects and secondary phases, and those electron microscopy images are affected by artifacts that occur during sample preparation and observation, making it challenging to achieve high phase segmentation accuracy. In this study, we focused on semantic segmentation by deep learning, which is successfully applied to medical images [1] and automatic driving [2]. Secondary electron images from the iron-based polycrystalline superconductor Ba122[3] were acquired in three dimensions by 3D FIB-SEM [4] and segmented by a deep learning model. A high IoU value over 90% was obtained [5, 6] for the images captured from the same sample and observation condition. Moreover, ty tuning the learning conditions and adopting data augmentation, we have succeeded in significantly improving the segmentation accuracy of electron microscopy images of samples with different chemical compositions and observed under different conditions than the learned images.<br/><br/><b>References</b><br/>[1] O. Ronneberger <i>et al</i>.<i>, Proc</i>.<i> MICCAI</i> (2015).<br/>[2] M. Cordts <i>et al.</i>, <i>In IEEE Conference on CVPR</i> (2016).<br/>[3] S. Tokuta and A. Yamamoto, <i>APL Mat.</i> 7, 111107 (2019)<br/>[4] Y. Shimada <i>et al</i>., <i>J. Alloys Compd</i>., <b>923</b>, 166358(2022)<br/>[5] Y. Hirabayashi <i>et al.</i>, Materials research Meeting. A4-PR15-04 (2021)<br/>[6] A. Yamamoto<i>.</i>, European Conference on Applied Superconductivity. 4-MO-FM2-01l (2023)