? 2022Continued growth in the nuclear industry is resulting in an increased burden for regulators due to the costs associated with implementing traditional safeguards. One such stakeholder, the International Atomic Energy Agency (IAEA), has published a research and development roadmap of suggested improvements that would help safeguards remain effective and efficient in the face of growing demands. These priorities include identifying, evaluating, and testing promising applications of robotics and machine learning/artificial intelligence (ML/AI) to improve safeguards. Nuclear material accountancy (NMA) is of particular interest because traditional approaches require time and resource intensive destructive assay (DA) measurements. One desirable improvement would be utilization of unattended measurements in order to reduce the burden of frequent DA measurements required at large throughput facilities, even though such unattended measurements would likely have higher uncertainty. The application of ML could improve detection of small changes, potentially due to material theft, in these higher uncertainty measurements collected at nuclear facilities to enable near real time anomaly detection. However, unique challenges arise when attempting to use deep neural networks with datasets containing errors that are characterized by models commonly associated with safeguards measurements. The purpose of this paper is to outline both traditional and ML based approaches to NMA and compare them through example. It can be shown that traditional methods for NMA often offers superior performance to the proposed ML pipeline, which consists supervised regression and unsupervised classification, when trained on datasets that contain measurement errors. ML approaches may still be competitive with traditional methods for NMA, however, special care must be taken to mitigate the impact of measurement error that disproportionately affects deep learning approaches.
展开▼
