Development and Performance Testing of a Nuclear Material Testbed Digital Twin

The Dynamic Material Control program at Los Alamos National Laboratory was established to develop advanced nuclear material monitoring techniques for the LANL Plutonium Facility. As part of this effort, a nuclear material testbed was built to be a representative platform of a nuclear processing facility. The testbed consists of four mock gloveboxes, each instrumented with four 3 He neutron detectors. In addition to being a platform for testing measurement capabilities, the testbed is also being used to establish advanced data analysis methods, including a digital twin. The digital twin is trained on a set of measurements taken with a source of known strength in prescribed positions within the testbed gloveboxes. With these measurements, models are constructed which enable real-time continuous unattended testbed monitoring. A single-source template matching scheme is implemented to search the database of prescribed measurements for patterns similar to in situ measured count rates, and an iterative deconvolution algorithm enables multiple sources to be simultaneously identified and localized as they move throughout the testbed. In this work, we detail the benchmarking of the algorithms in the digital twin. We examine the pros and cons of experiment-trained data driven digital twin modeling applied to a non-idealized environment. We compare the performance of our data driven models to other methods, such as reduced order analytical modeling and high-fidelity Monte Carlo physics modeling. The performance analysis includes quantitative metrics like computational speed and accuracy, as well as qualitative metrics like ease of implementation and resilience to change.