Machine learning interatomic potentials for NaPSO glasses: the critical role of training data

The performance of five machine learning interatomic potentials (MLIPs), based on MACE, DeePMD, and GRACE-FS architectures, is assessed in reproducing the structural and mechanical properties of Na 4 P 2 S 7-x O x ​ mixed oxy-thiophosphate glasses, promising candidates for next-generation all-solid-state sodium batteries. The glass series (0 ≤ x ≤3.5) was chosen to explore the effect of oxygen incorporation on short- and medium-range structural order (SRO and MRO), a particularly challenging task as experimental data show non-linear trends in density, conductivity, and structural units with composition. Universal MLIPs, trained on generic databases (MP0, MATPES-r2SCAN, METPES-PBE) or on datasets comprising only the elements relevant to glassy solid electrolytes (GRACE-GSE), provide stable molecular dynamics but often predict artifacts such as edge-sharing tetrahedra or P–P chains, and fail to reproduce the SRO evolution. A DeePMD MLIP, trained via concurrent learning on crystalline Na–P–S–O structures without stoichiometry-specific data, exhibits similar limitations. Among the universal MLIPs, MATPES-r2SCAN reproduces density, mechanical properties, and anion (S and O) speciation with reasonable accuracy. Fine-tuning MP0 on the specific glass compositions (FT-MP0) significantly improves structural reproduction, accurately capturing density trends, mechanical properties, and the preference of oxygen for bridging positions. Nonetheless, even FT-MP0 cannot accurately reproduce the disproportionation reaction of dimeric P 1 units converting to isolated P 0 units and P 2 chains associated with oxygen incorporation. These results highlight that universal MLIPs are valuable starting points for approximate simulations or database generation, but fine-tuning on both composition and relevant structural features is essential to accurately reproduce the short- and medium-range order of NaPSO glasses.