Insights From Applied Machine Learning For Safeguarding A Purex Reprocessing Facility

This paper will present key conclusions and observations from the application of machine learning to improve safeguards at a large throughput PUREX reprocessing facility. Specifically, this work hypothesizes that under normal operation a facility can be well approximated by an arbitrary function. Under facility misuse conditions this arbitrary function would no longer represent the facility. Performance of the machine learning approach is compared to traditional safeguards statistical tests using customary metrics such as probability of detection. This work shows that the residuals between observed and expected facility behavior is a good metric for anomaly detection provided certain criteria are met. Additional discussion on variables that have large impacts on the performance of the machine learning approach, such as experimental setup and available training data, is also provided.