Jingxuan Ding1,Boris Kozinsky1
Harvard University1
Jingxuan Ding1,Boris Kozinsky1
Harvard University1
Understanding the interfacial reactions between the Li-metal anode and solid-state electrolyte (SSE) and the Li ion diffusion mechanism are keys for developing stable and efficient solid-state batteries. Yet, theoretic studies are hindered by the high computational costs of ab initio molecular dynamics in both spatial and temporal dimensions. We combine on-the-fly active learning based on Gaussian Process regression (FLARE) with local equivariant neural network interatomic potentials (Allegro) to construct a symmetric battery of Li-metal anode and SSE and perform machine-learned molecular dynamics (MLMD) of tens of thousands of atoms over several nanoseconds with ab initio accuracy. Prominent reactions are observed at the interface with product phases forming a transition layer spanning a few tens of angstroms. Li ions near the interface are observed to migrate from the anode into the electrolyte and eventually diffuse within the SSE. Finally, we examine how off-stoichiometry affects the reaction rate of the interface and enhances Li ion diffusion.