| Abstract Scope |
Uranium mononitride (UN) is one of the ceramic nuclear fuel alternatives to oxide fuels considered for light water reactor and advanced reactor designs, but challenges remain before it can be qualified. For instance, accurate prediction of fission gas swelling is essential for safety analysis. Fuel performance in UN is further complicated by an accelerated swelling rate at high temperatures observed experimentally, sometimes referred to as breakaway swelling. In this work, we present a framework that bridges uncertainties from atomistic calculations to fuel performance predictions using neural networks and Bayesian calibration.
Density functional theory (DFT) is employed to compute the energetics of key defects and defect clusters in UN. This dataset informs a cluster dynamics model (Centipede) that computes self- and fission gas diffusivities under irradiation, as functions of temperature and nitrogen partial pressure. These outputs are in turn used to parameterize the fuel performance code BISON, that simulates fission gas swelling (and other fission gas properties) in specific reactor setups. Because this workflow is computationally expensive (and intractable for uncertainty quantification purposes), we train a neural network surrogate model to replace the full modeling chain.
We then use a Bayesian calibration strategy, using experimental measurements from separate effects testing and post-irradiation examinations, to compute the probability distributions of various meaningful parameters, including DFT energetics and microstructure properties such as the dislocation line density. Finally, the forward propagation of these distributions leads to a mechanistic, data-informed, and uncertainty-aware prediction of fission gas swelling in UN fuel. |