Structure and viscosity of liquid uranium-zirconium mixtures from machine-learning-based molecular dynamics

The thermophysical properties of liquid uranium-zirconium (U,Zr) mixtures play a crucial role in understanding the early stages of nuclear accident scenarios in pressurized-water reactors, as well as in other reactor designs. In this study, atomistic modeling leveraging ab initio calculations and a machine-learning interatomic potential framework are employed to investigate the static and dynamic properties of liquid (U,Zr) mixtures. More specifically, a spectral neighbor analysis potential trained on ab initio configurations is used to model the system, and its predictions are systematically compared and analyzed with experimental data from the literature and ab initio results. The model is then used to analyze an increase of the viscosity predicted for intermediate liquid (U,Zr) compositions, and its relationship to the internal structure of the mixtures. These insights contribute to the refinement of nuclear fuel models, enhance the accuracy of in-vessel corium retention simulations related to the mitigation of severe accident scenarios, and emphasize the possibility of using a combined ab initio and machine-learning interatomic potential approach to compute properties of such complex mixtures.