Investigation of Li-Ion Diffusion in Argyrodite Solid-State Electrolytes using Machine-Learned Potentials

Solid-state batteries incorporating solid Li-ion conductors have gained significant attention due to their advantages in safety and higher energy density compared to conventional liquid electrolyte batteries. However, these solid-state conductors inherently suffer from low ionic conductivities at room temperature, spurring investigation into diverse materials. Among them, argyrodite sulfide systems such as Li₆PS₅Cl have shown promising outcomes, achieving ionic conductivity higher than 10 mS/cm at room temperature. The high conductivities in argyrodites have been attributed to the presence of substitutional anion disorders. Yet, the precise mechanisms governing Li-ion diffusion remain unclear due to limitations in experimental analysis resolution. Theoretical investigations based on density functional theory (DFT) calculations are illuminating in this regard. However, recent DFT studies have encountered challenges in accurately quantifying diffusivities or ionic conductivities across various disorder levels, leading to substantial discrepancies with experimental results. These discrepancies mainly stem from two factors: i) the selection of semilocal exchange-correlation functionals based on the generalized-gradient approximation, which tends to overestimate lattice parameters and consequently increase conductivity, and ii) small simulation cells, which are susceptible to correlated Li-ion diffusions between periodic images and do not adequately account for the disorder of S/Cl ions. Furthermore, Li-ion jumps are rare at room temperatures, requiring simulations over an extended timescale, which is beyond the scope of current DFT calculations.<br/>In this presentation, to overcome the aforementioned computational issues, we employ Behler-Parrinello type neural network potentials (NNPs) to obtain Li-ion conductivities in argyrodite Li₆PS₅Cl with 0−100% Cl disorder at 4c sites. Leveraging the speed and accuracy of NNPs, we estimate the Li-ion conductivity directly at room temperatures, fully considering the S/Cl disorder and ergodicity of Li distributions. Using the SIMPLE-NN package, we train the NNP with strained bulk crystals and <i>ab initio</i> MD data at 600 and 1200 K. For the exchange-correlation energy among electrons, we select the SCAN functional, which reproduces the lattice parameter within 0.2% from the experimental value. We carry out NNP MD simulations with 4×4×4 supercells, encompassing more than 3,000 atoms, up to 15 ns simulation time at 300 K. Through parameter tests on the simulation timescale and ensemble averages of initial Li distributions, we were able to minimize the statistical fluctuations in the conductivity within 0.5 mS/cm. The estimated activation energies and ionic conductivities in the presence of disorders align well with experimental data. Interestingly, Li-ion conductivity peaks at 25% Cl at 4c sites. Free energy analysis confirms that these structures are thermodynamically accessible, suggesting a method to enhance the intrinsic Li-ion conductivity in Li₆PS₅Cl by tuning the site disorder. Furthermore, we analyze the diffusion mechanisms of ordered and disordered structures, focusing on the jumps between Li-ion cages. This work paves the way for predicting and analyzing properties of disordered complex solid-state electrolyte materials using accurate and cost-effective NNPs.