| About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
Materials Informatics to Accelerate Nuclear Materials Investigation
|
| Presentation Title |
Synthetic Data Driven Materials Informatics Methods for Nuclear Materials Characterization |
| Author(s) |
Kevin Field, Matthew J. Lynch, Gabriella Bruno, Ryan Jacobs, Nicholas Clancy, Dane D. Morgan |
| On-Site Speaker (Planned) |
Kevin Field |
| Abstract Scope |
Nuclear materials characterization relies on electron microscopy-based techniques. Image formation using electrons can be readily simulated which lends towards development of synthetic image datasets. Here, the applications of synthetic data to both develop convolutional neural network (CNN)-based machine learning models and to better identify the plausible human bias and variability in nuclear materials datasets will be discussed. Specifically, we will demonstrate the ability to rapidly generate a synthetic data trained CNN model (YOLOv7 architecture) for the quantification of radiation-induced cavities with accuracies on par with more tedious human-labelled datasets. The discussion on human bias and error will also focus on cavities and swelling but leverage the synthetic data methodology and a novel crowd-sourcing workflow. The collected dataset shows human capabilities such as small feature detection degrade when cavities are below 4-5% of the image width and we will cast these in terms of viable uncertainty bounds in literature swelling data. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Nuclear Materials, Computational Materials Science & Engineering |