Local topology hallucinations in universal machine-learned interatomic potentials
Predicting metastable states with universal machine-learned interatomic potentials requires that they preserve the minima and saddles of the reference potential-energy surface. Here we identify local topology hallucination, in which distinct model minima relax into the same reference basin. This failure arises from mode-selective curvature-sign errors shared by MACE, CHGNet and M3GNet. Curvature signs predicted by two of the three models agreed with density-functional calculations in 55 of 57 positive but only 5 of 56 negative cases. In LiGe$_2$S, two MACE minima connected by a verified saddle relaxed into one density-functional basin. Reference relaxation likewise returned nine of ten selected model minima to their reference basins. A directional test with two displaced reference calculations rejected 9 of 12 prospectively tested model instabilities. Model agreement alone therefore cannot establish that predicted minima are distinct metastable states.
Authors
- Xiaoqing Liu
- Lei Shen
- Yangshuai Wang
Publication Details
- Journal
- arXiv (Cornell University)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.48550/arxiv.2611.21489
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint