About this Abstract |
Meeting |
MS&T24: Materials Science & Technology
|
Symposium
|
Tackling Metallic Structural Materials Challenges for Advanced Nuclear Reactors
|
Presentation Title |
Performance Comparison of U-Net Based Machine Learning Architectures for Automated Analysis of TEM Images of Nuclear Materials |
Author(s) |
Aiden Ochoa, Xing Wang, Xinyuan Xu |
On-Site Speaker (Planned) |
Aiden Ochoa |
Abstract Scope |
Manual analysis of electron microscopy images is a time-intensive process that can be automated through the use of machine learning techniques. Compared to regular photography, difficulties in producing high quality images of irradiated materials often lead to smaller datasets that are more difficult to annotate. Additionally, previous studies have shown that the performance of machine-learning models may saturate with increasing dataset volume, implying further improvements may be achieved by improving model architecture. In this work, we aim to systematically compare the performance of different U-Net-based machine learning models for identifying grain boundaries in transmission electron microscopy images. We also investigate the impact of a large pre-trained network known as EfficientNet-B7 on the model’s performance. Our results indicate that the basic U-Net is generally sufficient at capturing global features. Nevertheless, the inclusion of a pre-trained encoder improves the crispness of prediction maps and contributes to the identification of mislabeled grain boundaries. |