Certifying the Kinetics of Machine-Learned Interatomic Potentials from Single-Point Energies
- 1School of Chemistry, Chemical Engineering and Biotechnology, Nanyang Technological University, 21 Nanyang Link, Singapore, 637371
Abstract
Machine-learned interatomic potentials are validated by force errors on equilibrium configurations, yet activated rates depend on transition regions that equilibrium sampling rarely visits. In a onedimensional metastable double well, the worst-case rate sensitivity to force error grows as eβB/2 with barrier height B; a path-space bound valid at any friction shows that force error certifies rare-event frequencies only to within a factor that grows at the same rate. The leverage density weights energy errors by their influence on reactive transport: it is the normalised dissipation density of the reactive current. For reversible overdamped dynamics, the Dirichlet and Thomson principles give two-sided capacity bounds, converted to metastable rate bounds by basin population corrections. A path representation evaluates their averages from reactive trajectories, committor values and single-point energies. Sampling uses reference paths to assess sampled transitions and model paths to expose additional or overpopulated pathways. In alanine dipeptide, closely matched force errors produce rate changes from 0.49 to 3.5, consistent with estimated intervals. In Li3OCl, mechanism-resolved upper endpoints track six models’ vacancy-hop rates with a systematic underestimation of about 7%. Labels selected from both ensembles improve kinetics even as force errors rise. We propose reporting the energy-error contrast M between the leverage density and the basins, and the spread σν under the leverage density, alongside force error; both are independent of the energy zero.
Related
- Correcting the Liquid Densities of Universal Machine-Learning Interatomic Potentials Using Local-Number-Density Repulsion
- pdb2reaction: End-to-End Reaction-Path Elucidation from PDB Structures Using Machine-Learning Interatomic Potentials
- Applicability Domain of Universal Machine Learning Interatomic Potentials at Solid-State Battery Interfaces
- Accurate and Rapid Vibrational Spectra of Molecular Crystals from Machine-Learned Interatomic Potentials
- How accurate are foundational machine learning interatomic potentials for heterogeneous catalysis?
BibTeX
@misc{yu2026certifying,
title = {{Certifying the Kinetics of Machine-Learned Interatomic Potentials from Single-Point Energies}},
author = {Xiyao Yu},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010329/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010329/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Certifying the Kinetics of Machine-Learned Interatomic Potentials from Single-Point Energies
AU - Xiyao Yu
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010329/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010329/v1
AB - Machine-learned interatomic potentials are validated by force errors on equilibrium configurations, yet activated rates depend on transition regions that equilibrium sampling rarely visits. In a onedimensional metastable double well, the worst-case rate sensitivity to force error grows as eβB/2 with barrier height B; a path-space bound valid at any friction shows that force error certifies rare-event frequencies only to within a factor that grows at the same rate. The leverage density weights energy errors by their influence on reactive transport: it is the normalised dissipation density of the reactive current. For reversible overdamped dynamics, the Dirichlet and Thomson principles give two-sided capacity bounds, converted to metastable rate bounds by basin population corrections. A path representation evaluates their averages from reactive trajectories, committor values and single-point energies. Sampling uses reference paths to assess sampled transitions and model paths to expose additional or overpopulated pathways. In alanine dipeptide, closely matched force errors produce rate changes from 0.49 to 3.5, consistent with estimated intervals. In Li3OCl, mechanism-resolved upper endpoints track six models’ vacancy-hop rates with a systematic underestimation of about 7%. Labels selected from both ensembles improve kinetics even as force errors rise. We propose reporting the energy-error contrast M between the leverage density and the basins, and the spread σν under the leverage density, alongside force error; both are independent of the energy zero.
ER -