Exchange–Correlation Functional Screening and Development of a High Dimensional Neural Network Interatomic Potential for UO2

Uranium dioxide (UO2) undergoes defect formation and microstructural changes under reactor irradiation, making atomic-scale modeling crucial for predicting its long-term thermomechanical performance. In this work, the paramagnetic state of bulk UO2 is simulated using non-spin-polarized density functional theory (DFT) without Hubbard-U corrections to ensure a consistent, computationally tractable reference dataset. To bypass the high cost of direct ab initio molecular dynamics (MD), we developed a Behler–Parrinello high-dimensional neural network potential (HDNNP) optimized with an active-learning framework. The training database was constructed from variable-cell MD trajectories across various temperatures and pressures, calculated using the Quantum ESPRESSO package. The resulting HDNNP achieves outstanding accuracy, yielding energy and force test RMSEs of 1.46×10-4 eV/atom and 8.73×10-3 eV/Å, respectively. Implemented in LAMMPS, the potential accurately predicts the equilibrium lattice parameter and single-crystal elastic constants (C11, C12, C44) of UO2. The predicted elastic constants exhibit excellent agreement with experimental data and DFT+U benchmarks, with relative errors remaining within 2–17%. This highlights the model’s capability to serve as a high-fidelity, computationally efficient surrogate for large-scale simulations of the mechanical behavior of nuclear fuels.