Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials
Abstract
Universal machine learning force-field generalization to multicomponent environments generated by compositional design remains insufficiently assessed. We construct a benchmark of 7,599 multicomponent configurations inspired by high-entropy design, elemental substitution and anion mixing. Eleven pretrained models are evaluated against density functional theory for energies, forces and stresses, with assessment extended to elastic, vibrational and adsorption-related properties. Force errors are analysed through training-reference coverage, local geometric heterogeneity, distance directionality and elemental response. Distances to training-reference environments reveal a qualitative association between coverage differences and increasing errors, while substantial variation remains at similar distances. Higher-error groups show greater local geometric heterogeneity, although OMat24 provides broad coverage of these environments. Relative to training-reference pair medians, errors remain low near the median, rise steeply on the compression side and increase more weakly on the extension side. After matching element pairs and absolute distance deviations, compression-side force errors are 1.81-1.95 times extension-side errors. Model-predicted pairwise interaction curves show greater curvature under compression. Fitting difficulty in independent elemental systems correlates with electronic band-energy responses to atomic displacements and Fermi-level shifts, and a similar pattern is observed in multicomponent systems. In parameter-matched comparisons, spherical-harmonic representations with maximum degrees of 2 and 4 lower test force errors for 38 and 40 of 43 elements, respectively, while differences in elemental difficulty remain. These findings inform force-field selection for experimental compositional design and identify targets for training-data sampling and model representations.
Related
- Thermodynamic origins of multimode spinodal decomposition in multicomponent alloys
- Can Universal Machine-Learning Interatomic Potentials Capture Small Polarons? A Geometry-Based Benchmark in Transition-Metal Oxides
- Distilling universal machine-learning potentials for moiré lattices across one million atoms
- Indirect prediction and screening of frustrated magnetic materials using machine learning
- Knowledge-Guided Construction of Machine Learning Potentials for Reactive Behavior in Energetic Materials
BibTeX
@misc{du2026origins,
title = {{Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials}},
author = {Hongwei Du and Dingyang Lv and Baole Wei and Yu Ren and Feng Yu and Xin He and Bonan Zhu and Jiahui Liu and Yongda Huang and Yongheng Li and Jianjun Liu and Siqi Shi and Hong Wang and Ziheng Lu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.09837},
eprint = {2610.09837},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.09837},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials
AU - Hongwei Du
AU - Dingyang Lv
AU - Baole Wei
AU - Yu Ren
AU - Feng Yu
AU - Xin He
AU - Bonan Zhu
AU - Jiahui Liu
AU - Yongda Huang
AU - Yongheng Li
AU - Jianjun Liu
AU - Siqi Shi
AU - Hong Wang
AU - Ziheng Lu
PY - 2026
DA - 2026/10/08
PB - arXiv
DO - 10.48550/arxiv.2610.09837
UR - https://arxiv.org/abs/2610.09837
AB - Universal machine learning force-field generalization to multicomponent environments generated by compositional design remains insufficiently assessed. We construct a benchmark of 7,599 multicomponent configurations inspired by high-entropy design, elemental substitution and anion mixing. Eleven pretrained models are evaluated against density functional theory for energies, forces and stresses, with assessment extended to elastic, vibrational and adsorption-related properties. Force errors are analysed through training-reference coverage, local geometric heterogeneity, distance directionality and elemental response. Distances to training-reference environments reveal a qualitative association between coverage differences and increasing errors, while substantial variation remains at similar distances. Higher-error groups show greater local geometric heterogeneity, although OMat24 provides broad coverage of these environments. Relative to training-reference pair medians, errors remain low near the median, rise steeply on the compression side and increase more weakly on the extension side. After matching element pairs and absolute distance deviations, compression-side force errors are 1.81-1.95 times extension-side errors. Model-predicted pairwise interaction curves show greater curvature under compression. Fitting difficulty in independent elemental systems correlates with electronic band-energy responses to atomic displacements and Fermi-level shifts, and a similar pattern is observed in multicomponent systems. In parameter-matched comparisons, spherical-harmonic representations with maximum degrees of 2 and 4 lower test force errors for 38 and 40 of 43 elements, respectively, while differences in elemental difficulty remain. These findings inform force-field selection for experimental compositional design and identify targets for training-data sampling and model representations.
ER -