Comparing and assessing the thermophysical and structural predictions of an empirical and a machine-learning...

Liquid uranium-zirconium (U,Zr) mixtures play a crucial role in the context of nuclear accident scenarios, particularly in the early stages of pressurized-water reactor accidents. In this study, we compare the thermophysical and structural predictions of two interatomic potentials (IAP) for this system, namely a modified-embedded atom model semi-empirical IAP and a spectral neighbor analysis potential (SNAP). Simulations are performed across a temperature range of 2050-2800 K at zero pressure, spanning the full composition range of liquid (U,Zr) mixtures, using simulation cells of 2000 atoms. The predictions of both potentials are benchmarked against experimental data andab initiomolecular dynamics results available in the literature. These models are employed to investigate the relationship between the viscosity, density, and structural properties of liquid (U,Zr) mixtures, and compare their respective predictions. The modified-embedded atom model (MEAM) potential predicts a significant viscosity anomaly at a molar fraction of 70% Zr, where the viscosity is up to approximately 14 times larger than the SNAP prediction at 2200 K, and 7 times larger at 2500 K. The density at this composition is also overestimated by the MEAM potential by 8% relative to the SNAP prediction. A thorough structural analysis leveraging radial distribution functions and structure factors, Voronoi tessellation, average degree of five-fold local symmetry, and common neighbor analysis supports these findings and attributes them to the formation of an icosahedral short-range order, with perfect icosahedral environments 8 times more prevalent at 70% Zr and 2200 K in the MEAM than in the SNAP. In contrast, this structural ordering is not observed with theab initioand SNAP-based computations. Since viscosity is a direct input to corium pool simulation tools such as PROCOR, an overestimation of this magnitude could significantly affect predictions of melt flow and stratification behavior in accident progression simulations. We finally analyze those differences and suggest that the semi-empirical MEAM potential may overestimate effects of short-range ordering in the liquid phase. These new results can play a role in the refinement of nuclear fuel models, improving the available recommendations for in-vessel corium retention simulations aimed at mitigating severe accident scenarios.