Beyond the First Coordination Shell: The Medium-Range Order Challenge for Machine-Learning Interatomic Potentials in Glasses

Machine-learning interatomic potentials are increasingly used to model complex inorganic glasses with near first-principles accuracy at molecular-dynamics length and time scales. However, reproducing energies, forces, densities, and first-neighbour coordination does not necessarily guarantee a reliable description of the medium-range topology of glass networks. In this perspective, based on results presented at the GlassLyon2026 conference and on recently published studies, we discuss this issue using two representative families: sodium phosphorus oxysulfide glasses and aluminosilicate glasses of the albite-anorthite series. In both cases, recent simulations show that machine-learning interatomic potentials can accurately reproduce local tetrahedral environments, bond distances, and short-range coordination. Nevertheless, more subtle descriptors of network connectivity, such as Pⁿ speciation in oxysulfides and Qⁿ distributions in aluminosilicates, remain difficult to capture quantitatively. Dispersion corrections improve density and local packing but do not fully resolve errors in medium-range connectivity. These examples show that the validation of machine-learning interatomic potentials for glasses must go beyond standard energy/force metrics and first-shell structural descriptors. We argue that future glass-specific potentials should be trained and validated against experimentally accessible medium-range observables, including NMR-derived speciation.