| About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
Materials Informatics to Accelerate Nuclear Materials Investigation
|
| Presentation Title |
Few-shot Machine Learning for Automated Analysis of TEM Images of Nuclear Materials |
| Author(s) |
Xing Wang, Xinyuan Xu, Zefeng Yu, Arthur Motta |
| On-Site Speaker (Planned) |
Xing Wang |
| Abstract Scope |
Transmission Electron Microscopy (TEM) is an essential tool for characterizing the effect of radiation damage on material microstructures. Automation of TEM image analysis could markedly expedite the characterization process. The recent use of Convolutional Neural Networks (CNN), relying on extensive training datasets, has automated many image analysis tasks. However, the scarcity of available TEM images of irradiated materials drives the need to learn efficiently from a reduced training dataset, i.e., few-shot learning. Using segmenting grains in TEM images of UO2 as an example, we demonstrate that selected CNN architectures, such as the UNet model, can successfully achieve few-shot learning for specific image analysis tasks. The performance of the UNet model in identifying UO2 grain boundaries was comparable to human experts using only about ten TEM images as the training dataset. This capability enables researchers to train custom models for automated grain morphology analysis in diverse materials, following the methodologies proposed. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Characterization, Nuclear Materials |