Atomic Structure of Na4P2S7 Glass Solid Electrolyte: Fine-Tuning Machine Learning Potentials for Enhanced Accuracy

High Resolution Image Download MS PowerPoint Slide Understanding the atomic structure of glassy solid electrolytes (GSEs) is essential for optimizing their performance in all-solid-state sodium batteries (ASSSBs). In this study, we investigate the structural organization of Na 4 P 2 S 7 glass, a representative composition of sulfide-based GSEs, using ab initio molecular dynamics (AIMD) and machine learning interatomic potentials (MLIPs). We assess the accuracy of the pretrained MACE foundation models (MP0-small, MP0-medium, and MP0-large) in reproducing the short- and medium-range order and demonstrate that fine-tuning with system-specific density functional theory (DFT) data significantly improves structural predictions. Our analysis of pair distribution functions (PDFs) and bond angle distributions (BADs) reveals that the pretrained models overestimate the presence of homonuclear P–P bonds and predict unrealistic edge-sharing P 2 S 6 units, artifacts that are corrected upon fine-tuning. The refined MLIPs also provide a better description of sulfur chain formation, which strongly influences sodium ion transport. Additionally, we validate the structural accuracy through 31 P NMR spectra calculations, showing improved agreement with experimental data. The vibrational density of states (VDOS) and elastic properties further highlight the advantages of fine-tuned models in capturing the structural dynamics of Na 4 P 2 S 7 glass. This work establishes a robust methodology for simulating sulfide-based GSEs and paves the way for future investigations into composition-structure–property relationships using MLIPs.