Beyond Static Accuracy: Ensemble-Dependent Reliability of Universal Machine-Learned Interatomic Potentials in Hybrid Amorphous Membranes

Universal machine-learned interatomic potentials (uMLIPs) are increasingly used as ready-to-run atomistic simulators and as teacher models for constructing task-specific potentials, yet their reliability is commonly judged from static energy and force errors. We show that this criterion can miss substantial errors in variable-cell molecular dynamics (MD). MatterSim and MACE were evaluated for three DFT-relaxed amorphous TiO₂–SiO₂ membranes containing different amounts of acetylacetone (ACA) under NVT and NPT conditions. Both uMLIPs exhibit large NPT cell excursions, while Ti–ACA coordination loss is strongly amplified at high ACA loading. In contrast, 1-ps PBEsol DFT-NPT trajectories show less than 0.3% volume change and no ACA detachment. Static volume scans reveal systematic pressure offsets of about 1.9–2.4 GPa, but these offsets are similar across compositions and imply only a roughly 2% zero-pressure volume shift. A DFT audit of 105 configurations sampled from 18 uMLIP NPT trajectories further shows that global pressure and force errors neither amplify along the trajectory nor spike at coordination-loss onset. The results motivate ensemble-aware, trajectory-level validation before uMLIPs are deployed for variable-cell simulation or teacher-driven data generation.