Irradiation in nuclear fuels generates vacancy clusters whose coupled impact with temperature on the mechanical integrity of uranium dioxide (UO 2 ) remains difficult to quantify across relevant length and time scales. Here we develop an on-the-fly machine-learning force field (MLFF) for UO 2 with vacancy-defect clusters, trained via on-the-fly machine learning on a high-throughput density functional theory (DFT) dataset and deployed for massively parallel molecular dynamics simulations in LAMMPS. The MLFF reproduces key elastic properties of UO 2 in close agreement with molecular dynamics simulation results and available experiments, enabling efficient simulations over 300–1500 K and 1–5% vacancy concentrations. We find that vacancy clusters and thermal perturbations act cooperatively to produce an approximately linear degradation of the bulk modulus and shear modulus, with the degradation rate increasing systematically with both vacancy concentration and cluster size. Beyond mechanical softening, rising temperature and vacancy concentration expose the limitations of the force field under high-entropy configurations, resulting in inadequate description of vacancy cluster evolution and growth by the MLFF. These results provide valuable insights for the further development and optimization of MLFF, establishing a predictive framework for the performance evaluation of uranium-based nuclear fuels.
Zhou et al. (Sun,) studied this question.