An ML/AI Approach to Identifying Gaps in a priori
Understanding of Nuclear Facility Design and Operations

Monitoring and characterization of nuclear facilities is an essential activity of nuclear nonproliferation, materials control, and safeguards. Such inferences are best supported by extensive a priori knowledge of facility design and operations, but that knowledge is not always correct, current, and complete. We present a technique for identifying discrepancies between a priori understanding and actual conditions on the ground by comparing the output of computational models of facility activities with sensor data gathered on-site. The technique leverages a novel unsupervised machine learning algorithm to provide a near-real-time rating of the discrepancy between expected and observed behavior. The algorithm is validated against a comprehensive anomaly detection benchmark, including 14 other unsupervised anomaly detection methods on ten datasets. We present promising results from applying the proposed technique to a prototype ML/AI system deployed at two testbed facilities. The results show the algorithm's effectiveness in identifying and explaining real-world discrepancies in support of monitoring and characterization activities.