Advances In Machine Learning For Safeguarding A Purex Reprocessing Facility

A longstanding goal of the IAEA has been to reduce the frequent sampling required at high throughput nuclear facilities. Advances in machine learning combined with Non-Destructive Analysis (NDA) measurements and process monitoring could reduce the sampling requirements for these high throughput facilities. Work from FY19 demonstrated promising results that showed machine learning approaches with NDA measurements could be competitive with traditional safeguards, such as statistical analysis on Destructive Analysis (DA) measurements. Building on previous work the machine learning approach has been improved and extended. The machine learning implementation has been updated to cover more realistic scenarios by including improved NDA features and to encompass an entire Material Balance Area (MBA). This work also discusses a new, joint approach that combines supervised and unsupervised anomaly detection. The joint approach would be useful where only a few high-consequence loss paths for a particular facility are known. Discussion will also be provided on practical details related to the machine learning approach, such as the quantity of training data that would be required. Finally, the work discusses the application of the current machine learning approach to other nuclear facilities, such as enrichment.