| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Novel Strategies for Rapid Acquisition and Processing of Large Datasets from Advanced Characterization Techniques
|
| Presentation Title |
Comparing Performance of U-Net Based Neural Networks for Automated Detection of Defects in TEM Images of Nuclear Materials |
| Author(s) |
Aiden Ochoa, Xinyuan Xu, Xing Wang |
| On-Site Speaker (Planned) |
Aiden Ochoa |
| Abstract Scope |
Nuclear materials undergo significant microstructural changes throughout their life. Transmission electron microscopy (TEM) is widely used for characterizing such evolution, though manual analysis of TEM images is time-intensive, potentially biased, and often irreproducible. These issues can be addressed by machine learning networks, enabling automatic and high-throughput analysis of microscale defects in nuclear materials. Here we systematically compared the performance of different U-Net-based networks on datasets of TEM images containing grain boundaries and helium bubbles generated by irradiation. We also investigate the impact of a large pre-trained backbone, known as EfficientNet, on the model’s prediction capabilities. Our results indicated that the basic U-Net was sufficient at capturing most studied defects with F1-scores higher than 0.76 for the grain boundary dataset and 0.87 for the helium bubble dataset. However, including a large pre-trained encoder improved the overall prediction quality, allowing for detecting some missed features and reducing human labelling flaws. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Nuclear Materials, Characterization |