Tracking Material Transfers At A Nuclear Facility With Physics-informed Machine Learning And Data Fusion

Operations at nuclear fuel cycle facilities are complex sequences of numerous subtasks that may include material production, destruction, or chemical processing; packing and unpacking; and transfer into, out of, or between areas or buildings within the facility. Identifying these tasks and tracking temporal patterns may indicate and help characterize long term operations at the facility. We have developed methods to selectively identify transfers of radioactive material using a network of infrasound and NaI detectors deployed around a multiuse research reactor and radiochemical processing facility. The testbed in question focuses on the collocated High Flux Isotope Reactor and Radiochemical Engineering Development Center at Oak Ridge National Laboratory, which have been instrumented with multiple measurement modalities under the Multi-Informatics for Nuclear Operation Scenarios venture. Using real infrasound and gamma-ray spectral data collected around these facilities, we have developed a neural network to robustly identify passing vehicles and differentiate them from other sources of background infrasound signals such as wind, planes, and heavy machinery. Furthermore, we have studied the physics of neutron-emitting nuclear material in a heavily shielded cask to develop a physics-informed feature construction procedure that allows specific identification of shielded material using gamma-ray spectral data when typical analysis procedures such as peak analysis are not applicable. These two data streams are fused to uniquely identify vehicle-based nuclear material transfers, which are then input into a sequence model based on long short-term memory recurrent neural networks. In our initial results, trained on synthetic data and transferred to real data collected at the testbed, these recurrent models are able to find sequences of material transfer detections throughout the multisensor network that are indicative of material transfers and differentiate these patterns from spurious false alarms typical of complex, multiuse facilities.