Enhancing Li-Ion Conductivity in Argyrodite Li₆PS₅Cl through Disorder Engineering Investigated by Machine Learning Potential

Solid-state electrolytes (SSEs) have emerged as promising candidates for next-generation batteries, offering improved safety and higher energy density compared to conventional liquid electrolytes. Nevertheless, the practical utilization of SSEs has been hindered by the challenge of achieving high ionic conductivity at room temperature. Among SSEs, argyrodite systems have achieved a high conductivity of 10 mS/cm at room temperature through strategies involving disorder engineering and doping. To further optimize ionic conductivity and gain a comprehensive understanding of the underlying mechanisms, theoretical calculations based on density functional theory (DFT) have been applied. However, recent DFT studies have faced challenges in accurately quantifying transport properties in argyrodite systems compared to experiments, primarily due to computational limitations. For instance, small simulation cells introduce spurious correlations in Li-ion motions between periodic boundaries, while a scarcity of diffusion events makes direct simulation at room temperature difficult.<br/><br/>In this presentation, we present the use of Behler-Parrinello-type neural network potentials (NNPs) to overcome the aforementioned challenges and obtain Li-ion conductivities in argyrodite Li₆PS₅Cl with varying S/Cl disorders (0−100%). Utilizing cost-effective and accurate NNPs allows for direct simulations at room temperature, unraveling the effects of S/Cl disorders. We select the SCAN functional, which reproduces the lattice parameter within 0.2% of the experimental value. To train the NNPs, we utilize the SIMPLE-NN package and employ strained bulk crystals and ab initio MD data at 600 and 1200 K. Several validations are conducted to ensure the quality of our NNP in terms of both structural and dynamic properties. To obtain the ionic conductivity with a statistical convergence error range of 10%, we carry out systematic tests on the supercell and ensemble sizes. As a result, NNP MD simulations are performed using 5×5×5 supercells containing 6500 atoms for up to 25 ns. In calculating the diffusivity, we neglect the initial 5 ns, where most of the Li-ions remain within the cage at room temperature, resulting in non-linear mean squared displacements. Our predicted activation energies and ionic conductivities align well with experimental data. Interestingly, the conductivity peaks at 25% S/Cl disorder. By analyzing the diffusion mechanism, we find that the peak observed at 25% disorder arises from the synergetic effects of two contributing factors; the enhancement of rotational motion and the disorder effect which facilitates the Li diffusion between cages. Additional free energy analysis shows that these structures are thermodynamically accessible, suggesting the potential for optimizing Li-ion conductivity through disorder engineering in Li₆PS₅Cl. This work paves the way for estimating and studying the transport properties of SSEs using accurate and cost-effective NNPs.