Reproduce “Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hongwei Du, Dingyang Lv, Baole Wei, Yu Ren, Feng Yu, Xin He, Bonan Zhu, Jiahui Liu and 6 more
A benchmark of 11 pretrained interatomic force fields on 7,599 multicomponent material configurations finds that accuracy depends on the target property: the model with the lowest formation-energy error does not lead on every derived calculation. The error analysis links larger force errors to unfamiliar environments, short and compressed contacts, local chemical and geometric context, and element-specific electronic responses.
Key insights
- Training-environment distance is a useful warning signal for force and stress errors, but not a complete explanation; local chemistry and geometry add predictive information.
- Compression is more error-prone than extension: at matched element pairs and absolute distance deviations, compression-side force errors are 1.81–1.95 times larger.
- Element-specific electronic responses contribute to recurring error patterns, while angular representations reduce test force errors for most elements in the separate Metal-43 comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.