About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Accelerated Qualification Methods for Nuclear Reactor Structural Materials
|
| Presentation Title |
Building a Cohesive ML Ecosystem for Accelerated Nuclear Materials Qualification
|
| Author(s) |
Kevin Field, Vidit Agrawal, Gabriella Bruno, Christopher R. Field, Chan-Min Hsu, Ryan Jacobs, Ni Li, Matthew J. Lynch, Dane Morgan, John Olamfe, Xiaoning Qian, Lijun Qian, Yucheng Wang, Xing Wang, Siyuan Xu |
| On-Site Speaker (Planned) |
Kevin Field |
| Abstract Scope |
Convolutional Neural Networks (CNNs), such as Mask R-CNN, have demonstrated promise in accelerating and standardizing the quantification of radiation-induced cavities in nuclear materials. However, broader adoption is limited by three key challenges: variability in human labeling, CNN difficulty detecting small features, and limited domain awareness. This presentation introduces a machine learning (ML) ecosystem designed to address these barriers. First, we quantify human label variability through a combination of crowdsourcing and synthetic data, identifying consistent challenges with small feature recognition. We then demonstrate that synthetic data can effectively train CNNs, reducing reliance on human annotations. To enhance performance, we incorporate image super-resolution and two-modality image registration techniques, improving feature visibility and detection. Lastly, we present domain-aware methods using CNN-derived confidence scores and random forest classifiers to flag suboptimal image inputs. These advancements offer a path toward more robust, scalable, and interpretable ML-driven workflows for nuclear materials characterization and qualification. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Characterization, Machine Learning, Nuclear Materials |