| About this Abstract |
| Meeting |
Materials in Nuclear Energy Systems (MiNES) 2021
|
| Symposium
|
Materials in Nuclear Energy Systems (MiNES) 2021
|
| Presentation Title |
Explorations in Automated Cavity Detection Using an Expanded Machine Learning Training Data Domain |
| Author(s) |
Matthew Lynch, Ryan Jacobs, Steven Chen, Rett Graham, Dane D. Morgan, Kevin G. Field |
| On-Site Speaker (Planned) |
Matthew Lynch |
| Abstract Scope |
The quantification of cavities in post irradiated microscopy is key to understanding materials performance. However, manual detection is a time-consuming process. Recently, machine learning (ML) models have successfully detected and analyzed defects. These models perform at near human levels, but with increased speed and repeatability. A downfall of current models are they are mostly built around a single material/irradiation, meaning their use domain is narrow. Here, we explore the influence of expanding dataset sizes and domain for cavity features. This is accomplished via two expansions of the training data. Firstly, a 30,000+ instances experimental database has been developed to provide an expanded domain space. Secondly, using simplified physics and unirradiated micrographs, artificial data were automatically created and labeled at essentially no cost. Incremental and expanding model training using this combination revealed the correlation between training domain and test instances with model generality increasing with increased training domains. |
| Proceedings Inclusion? |
Undecided |