Data mining of molecular dynamics data reveals Li diffusion characteristics in garnet Li7La3Zr2O12

Understanding Li diffusion in solid conductors is essential for the next generation Li batteries. Here we show that density-based clustering of the trajectories computed using molecular dynamics simulations helps elucidate the Li diffusion mechanism within the Li7La3Zr2O12 (LLZO) crystal lattice. This unsupervised learning method recognizes lattice sites, is able to give the site type, and can identify Li hopping events. Results show that, while the cubic LLZO has a much higher hopping rate compared to its tetragonal counterpart, most of the Li hops in the cubic LLZO do not contribute to the diffusivity due to the dominance of back-and-forth type jumps. The hopping analysis and local Li configuration statistics give evidence that Li diffusivity in cubic LLZO is limited by the low vacancy concentration. The hopping statistics also shows uncorrelated Poisson-like diffusion for Li in the cubic LLZO, and correlated diffusion for Li in the tetragonal LLZO in the temporal scale. Further analysis of the spatio-temporal correlation using site-to-site mutual information confirms the weak site dependence of Li diffusion in the cubic LLZO as the origin for the uncorrelated diffusion. This work puts forward a perspective on combining machine learning and information theory to interpret results of molecular dynamics simulations.