| About this Abstract |
| Meeting |
MS&T21: Materials Science & Technology
|
| Symposium
|
Advanced Characterization of Materials for Nuclear, Radiation, and Extreme Environments
|
| Presentation Title |
Deep Learning Pipeline for Cavity Segmentation in Transmission Electron Microscopy |
| Author(s) |
Chun Yin Wong, Xing Wang, Zhe Fan, Karren L. More, Sergei V. Kalinin, Maxim Ziatdinov |
| On-Site Speaker (Planned) |
Chun Yin Wong |
| Abstract Scope |
A physics-informed deep learning (DL) pipeline is proposed to perform cavity segmentation in transmission electron microscopy (TEM) images of irradiated concentrated solid solution alloy (CSA) systems. Irradiation-induced cavities threaten the mechanical integrity of structural materials in nuclear reactors and CSAs, particularly high entropy alloys, have shown to resist cavity growth. The challenge of measuring large numbers of cavities can be overcome by automated analyses yet traditional machine learning (ML) methods are inadequate. The proposed DL pipeline outperforms traditional ML algorithms by incorporating ensemble learning in the pixelwise identification of cavities. Cavity dynamics are incorporated via computer vision techniques—Laplacian of Gaussian and Hough Transform—to measure the cavities. The DL pipeline achieved an intersection-over-union of 0.80 and overlapping cavities were individually identified. The cavities were identified instantaneously and only five images were labeled for training. The success of the DL pipeline paves the way for automated microscopy experiments in the future. |