Machine-learned potential for fission gas diffusion in uranium oxide nuclear fuels

We report the first machine-learned interatomic potential for uranium dioxide with xenon gas, as well as a foundational machine-learned potential for uranium dioxide. Training datasets were constructed by leveraging a combination of density functional theory calculations with a Hubbard U correction and molecular dynamics simulations. Query-by-committee active learning procedures further automated the augmentation of training datasets. The efficacy of employing an equivari- ant message-passing neural network for iterative potential fitting was demonstrated by reproducing DFT+U -level forces and energies, despite the training datasets being much smaller than those for recently-reported uranium dioxide MLPs. We found our machine-learned potential for UO2 achieves strong agreement with experimentally-observed thermophysical and thermomechanical properties across a temperature range of 300 K to 3000 K. Our second machine-learned potential for uranium dioxide with xenon successfully replicates reference DFT+U incorporation energies of xenon into various lattice sites. We also employed this MLP to calculate migration barriers for xenon diffusion via tetravacancy defect cluster mechanisms. The superlative ability of these machine-learned poten- tials to capture the behavior of uranium dioxide with xenon inclusion across a range of temperatures and defect chemistries lays the foundation for larger-scale molecular dynamics simulations of fission gas transport through uranium oxide fuel matrices.