Expansion of nuclear power as a means to meet increasing energy demands while addressing the issue of climatechangepresentschallengesfornonproliferationefforts, globally. Effective safeguards technologies have the potential to both improve current detection of anomalous operations, and increase monitoring efficiency to ensure that the future international expansion of nuclear electricity generationisresponsible. Physics-informedsystems, suchasdigitaltwins, arenecessarytoreproduce the complex behavior of nuclear reactors. Full high fidelity models can preserve the detailed physics well, but are unable to execute in real time and be useful in a digital twin. Mathematical surrogate models are a viable solution to this problem. Trained on data from full physics models, surrogates can maintain a degree of that fidelity, while drastically reducing computational costs to a level which can be useful in a digital twin. To what extent anomalous behavior can be detected with a surrogate model is an open area of research. An assessment of the impact of surrogate types, and volume/distribution of training data on the potential for anomaly detection is the subject of this research. Surrogate models and associated training data are generated using Idaho National Laboratory’s Multiphysics Object Oriented Simulation Environment (MOOSE). MOOSE contains an array of useful tools to create full physics models of virtually any reactor type and has a built-in surrogate model creation tool that can be directly trained on full physics model evaluations. Several surrogate model types, as well as sampling techniques are evaluated. Gaussian process, polynomial chaos, and nearest point surrogates are compared and trained using variably-sized samples with quadrature-based, Latin Hypercube, and Cartesian product distributions. Models are based on Idaho State University’s AGN-201 research reactor and surrogate model results are evaluated against both high fidelity physics models and real reactor data.
Year
2024
Abstract