Solid-state electrolytes are critical for advancing lithium-ion battery technologies due to their potential to provide higher safety, improved energy density, and better long-term stability, compared to conventional liquid electrolytes. Among solid-state electrolytes, sulfide-based materials such as Li₆PS₅Cl (LPSC) are especially promising because of their high ionic conductivity, low interface impedance, and mechanical flexibility. LPSC has an argyrodite structure that inherently offers abundant lithium-ion migration pathways, ensuring efficient ionic transport. However, practical applications of LPSC are often hindered in the presence of intrinsic defects, such as lithium vacancies and LiCl substitutions, thus altering the local atomic environments and macroscopic ionic transport properties. Conventional computational methods like ab initio molecular dynamics (AIMD) offer a quantum-level precision, but are computationally prohibitive in large scales, whereas classical empirical potentials lack the accuracy to capture complex defect-induced local rearrangements. Therefore, developing a computational method that combines quantum accuracy with large-scale simulations remains an important challenge.

This study used a deep learning-based large-atom model (DPA-SSE), employing a "pretraining-fine-tuning-distillation" strategy specifically designed for sulfide solid electrolytes. Initially, the model was pretrained based on extensive density functional theory (DFT) data covering diverse sulfide materials and defect structures. Subsequently, it was fine-tuned using a carefully selected subset of 551 high-value DFT configurations specifically related to LPSC and its defect environments, chosen from a larger dataset of 11 631 configurations. These configurations included comprehensive DFT-calculated energies, atomic forces, and virial stress tensors. The fine-tuned model could achieve a quantum-level precision with a significantly reduced mean absolute error in predicting energies and forces. Finally, a distilled model was proposed via leveraging the fine-tuned model to label extensive datasets efficiently, enabling simulations of large-scale systems with thousands to millions of atoms in nanosecond time scales.

The molecular dynamics simulations are conducted to systematically investigate the effect of LiCl defect concentration on the ionic conductivity in LPSC at different temperatures by the distilled DPA-SSE model. At a lower temperature (i.e., 333 K), the simulations show that introducing LiCl defects moderately improves the ionic conductivity, compared to pristine LPSC, indicating that defects at a low concentration can create a beneficial local disorder facilitating ion transport. However, at an elevated temperature (i.e., 500 K), the simulations clearly demonstrate a negative correlation between ionic conductivity and increasing defect concentrations. Specifically, the measured ionic conductivities at different defect concentrations (i.e., 0%, 3.8%, and 7.7%) are approximately 3.23 × 10^-2, 2.06 × 10^-2 S/cm, and 1.36 × 10^-2 S/cm, respectively. The reduced ionic conductivity at higher defect concentrations is attributed to an increased structural instability, leading to a partial decomposition or a significant rearrangement of PS4 tetrahedral units, which form critical frameworks for efficient ion migration. The radial distribution function (RDF) analysis further corroborates these findings, showing intensified local coordination and disrupted ion diffusion pathways at higher defect levels.

This work demonstrated the feasibility and effectiveness of using a deep learning-based large-atom model to accurately and efficiently simulate ionic transport in defect-rich sulfide solid electrolytes. The "pretraining-fine-tuning-distillation" approach significantly reduced the computational cost associated with conventional quantum mechanical simulations, while maintaining a high predictive accuracy. This study provided valuable theoretical insights that could inform the design and optimization of high-performance solid-state electrolytes via revealing the intricate relationship among defect concentrations, local structural dynamics and ionic conductivity. Moreover, the methodological framework presented could have a broad applicability, offering a powerful and scalable computational tool for studying a variety of complex materials in energy storage applications.