Paddle-Wheel Effect Mediates Polymorph Transitions but not Li Transport in the Ionic Conductor Li₃PS₄.

Ionic transport is a key determinant of performance in energy technology devices, and simultaneously a convenient testing ground for machine-learned neural-network interatomic potentials. Of particular interest is the influence of slow degrees of freedom on ion conduction. Here we investigate the influence of large-amplitude polyhedral-anion flips on Li transport in the modern solid-state ion conductor Li₃PS₄, where a paddle-wheel effect has been proposed to benefit transport, using ensembles of local equivariant graph neural network potentials.<br/>We find that at elevated temperature a large variety of models yield consistent predictions for the rate of polyhedral flips, approximately one per nanosecond at 800 K, and variance within a factor of 2 between models trained with distinct hyperparameters and across different sizes of training data (32-128 atom cells) despite the large increase in out-of-dataset testing errors versus larger-scale ab initio frames.<br/>Importantly, we find that the large-amplitude paddlewheel flips are orders of magnitude too slow to facilitate Li transport. Further, using well-tempered metadynamics enhanced sampling, we estimate the free-energy kinetic barriers for the polyhedral flips to be ≥0.7 eV, which is approximately double the activation of ion conduction. Instead, the lowest-activation polyhedral-anion flip is collective and mediates the phase transition between room-temperature and elevated-temperature polymorphs in a one-dimensional mechanism.<br/>Overall, our simulations are consistent with the orientational disorder observed experimentally, and suggest that such disorder rather than a purely dynamic paddle-wheel effect is beneficial to ion conduction. Owing to its rich dynamical landscape and polymorphism coupled to it, Li₃PS₄ can serve as a benchmarking material for testing machine-learning interatomic potentials and enhanced sampling methods.