Applicability Domain of Universal Machine Learning Interatomic Potentials at Solid-State Battery Interfaces

Universal machine learning interatomic potentials (uMLIPs) are trained predominantly on bulk systems, yet are increasingly applied to solid-solid interfaces. Evaluating six uMLIPs against DFT across 24 Li-metal/solid-electrolyte interfaces, we map their applicability domain (AD) and where they break down. Total energies are predicted reasonably well (MAE < 15 meV/atom), whereas interatomic forces at the interfaces are not. Most models underestimate the interfacial formation energy by 0.32 J/m2 and overestimate the work of adhesion by 0.26 J/m2, and none reproduces the stability ordering of polyanion-containing interfaces. Despite large force errors for the interfaces, Li-ion kinetics agree well across models. The errors concentrate on the interface atoms and are larger for the polyanion chemistries. Descriptor-based analysis traces the higher force errors to interfacial environments that lie outside the pretraining data. Fine-tuning can close the accuracy gap, but only with data that samples the bonding across the interface, whereas slab data alone degrades performance. All models struggle on the same chemistries irrespective of the architecture, indicating the coverage of the pretraining data as the primary limitation. Overall, pretrained uMLIPs are suitable for qualitative screening of interfaces with simple single anion electrolytes, while those containing polyanions require interface-specific data.