Unsupervised Machine Learning for Nuclear Safeguards

Machine learning is an area of active research that powers projects from search algorithms to self-driving cars. However, to date there has been very little work done on applying these techniques to nuclear safeguards problems. Current nuclear material accounting requirements for large throughput facilities require destructive analysis (DA) with sufficient precision to detect material loss or facility misuse. DA measurements can be expensive for the operator and regulator. The use of machine learning with cheaper non-destructive analysis (NDA) could lead to increased detection probability and timeliness. We present a unique interpretation of an unsupervised technique, the One-Class Support Vector Machine (OCSVM), to detect off-normal conditions at a facility. This technique is selected because it has a relatively small computational requirement and can be implemented at existing facilities with minimal investment in support equipment. Additionally, the OCSVM requires no knowledge of potential off-normal or facility misuse conditions. The model can be tuned to detect these unknown off-normal conditions while ignoring anticipated facility transients. In this work, we generated data from a nuclear pyroprocessing facility model to determine the effectiveness of the OCSVM. We used the data to train the OCSVM and evaluate its performance in distinguishing normal from off-normal conditions. Results will be presented in the full paper.