Discovery of Lithium Solid State Electrolytes through Machine Learned Probability Density Analysis

Solid state Li-ion conducting electrolytes are a class of materials with appreciable ionic conductivities, but which are still often several of orders of magnitude lower than liquids impeding their commercial adoption in commercial solid-state Li and Li-ion batteries. The major bottleneck to the design of solid electrolytes with high ionic conductivity is the lack of an efficient predictive model of Li-ion diffusivity from either the composition or the crystal structure. Traditional ab initio molecular dynamics (AIMD) simulations have very high accuracy, but are limited by both the number of atoms and timescales they can simulate due to the O(n3) computational cost. Here, we propose using the Li-ion potential energy surface calculated from AIMD to predict ionic conductivity. We show direct correlations (R2= 0.72) between several topological descriptors and both the magnitude of the ionic conductivity and diffusion mechanism across many categories of ionic conductors including oxides, halides, and sulfides. Generating the Li-ion probability density typically requires a full AIMD simulation, however, to accelerate the calculation for new materials, we train a three-dimensional U-net convolutional neural network to predict the Li-ion probability density directly from the pseudo-electron density of the crystal (Figure 1). Using AIMD simulations of fifty known Li-ion conductors as training data, we train a model with 91% test accuracy. Our model is able to efficiently calculate the probability density for new materials including arbitrary compositions, space groups, and site occupations (i.e., disordered structures) which greatly expands the design space over which we can optimize and discover ionic conductors. By screening the ICSD we identify several for new candidate solid state electrolytes and re-identified Li6NBr3. This composition could not be directly simulated using AIMD due to the large unit cell volume, but its ionic conductivity has been experimentally measured as 0.0035 mS/cm at 150°C demonstrating both the accuracy and utility of our model for solid state electrolyte discovery. Figure 1