Assessment of machine-learned vs. classical/quantum atomistic simulations for predicting palladium-hydrogen materials

Palladium-hydrogen materials are used in a wide range of hydrogen storage, purification, sensing, and catalysis technologies; however, accurately predicting the structural and kinetic behavior of palladium hydrides remains computationally challenging. In this work, we present the first systematic assessment of Meta’s Universal Model for Atoms (UMA), a machine-learned interatomic potential, by benchmarking its predictions against density functional theory (DFT) and two widely used embedded-atom method (EAM) potentials for PdHx systems. All approaches were in close agreement for predicting equilibrium structural and thermodynamic properties, such as lattice constants and cohesive energies. In contrast, simulations of nonequilibrium hydrogen diffusion kinetics revealed substantial differences among the methods. Both EAM potentials predicted qualitatively incorrect hydrogen diffusion kinetics, exhibiting nearly instantaneous diffusion dynamics regardless of temperature and failing to capture the formation of molecular H2. UMA correctly reproduced the temperature dependence of hydrogen diffusion and qualitatively matched the DFT-predicted kinetic behavior, although diffusion timescales remained faster than DFT, particularly at lower temperatures. Overall, UMA provides the best balance between predictive accuracy and computational efficiency for large-scale palladium-hydrogen simulations, with DFT serving as a supporting role to obtain more accurate estimates of kinetic timescales, if needed. Our results show that existing EAM potentials (which were historically the most widely used atomistic methods for large-scale studies of these systems) should be used with caution, particularly for palladium-hydride diffusion kinetics.