Uranium nitrides (UN, ${\mathrm{U}}_2{\mathrm{N}}_3$, and ${\mathrm{UN}}_2$) are candidate materials for advanced nuclear fuels and corrosion-resistant coatings, yet achieving a unified atomic-scale model that accurately captures their behavior across stoichiometries and under extreme conditions remains challenging. In this work, we present a unified machine learning interatomic potential (MLIP) for the entire U–N system, developed within the neuroevolution potential framework. The potential is trained on a comprehensive density functional theory (DFT) dataset constructed through an iterative active learning procedure that effectively samples configuration space using farthest point sampling and random network distillation. The resulting MLIP accurately reproduces structural, elastic, vibrational, and defect properties of all major uranium nitride phases, showing close agreement with DFT and experimental reference data. It reliably predicts finite-temperature thermodynamics, defect migration barriers, nitrogen diffusion kinetics, and the stability of nonstoichiometric compounds. This transferable and efficient potential enables large-scale molecular dynamics simulations of key processes in uranium nitrides, providing a robust and efficient computational tool for microstructural evolution and phase stability in next-generation nitride-based nuclear systems.

Close

PrevNext

View figure in article

• Density functional theory
• Machine learning
• Molecular dynamics