Subtle Energy Corrections Reshape Condensed-Phase Property Predictions from a Machine-Learning Interatomic Potential
physics.chem-ph
Submitted: 2026-11-04
Updated: 2026-11-04
Comments: 17 pages, 7 figures, 1 table
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
The gist: Accurate reproduction of electronic-structure labels by machine-learning interatomic potentials does not necessarily ensure reliable condensed-phase properties.
Abstract
Accurate reproduction of electronic-structure labels by machine-learning interatomic potentials does not necessarily ensure reliable condensed-phase properties. We investigate the effects of basis-set superposition error (BSSE) and an explicit Axilrod-Teller-Muto (ATM) three-body dispersion contribution on molecular-liquid predictions. Starting from MACELES-OFF, we fine-tune a baseline potential at the ω B97M-V/def2-TZVPD level. A separate branch learns counterpoise (CP) energy and force corrections with the baseline frozen, while a D4-parametrized ATM term is added during molecular-dynamics inference. Mean atom-normalized CP and ATM contributions are 0.222 and 0.205 meV/atom, respectively, for the analyzed SPICE2 configurations from OMol25. We compare baseline, CP-only, ATM-only, and CP + ATM variants for 39 organic liquids. Relative to baseline, the mean system-wise density changes are-1.17%, -10.38%, and-11.93% for CP-only, ATM-only, and CP + ATM, respectively; the corresponding enthalpy-of-vaporization changes are-2.79%, -14.74%, and-18.32%. For the 11 liquids with experimental self-diffusion data, the mean self-diffusion changes are +17.66%, +203.02%, and +230.36%. Both corrections thus lower density and enthalpy of vaporization while increasing molecular mobility on average. ATM dominates these changes, while CP provides a smaller contribution; their combination yields the largest average shifts. These results show that corrections with sub-meV/atom mean values on the analyzed configurations can substantially reshape the predicted structural, thermodynamic, and transport properties of molecular liquids.
Sources
- MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry
- AceFF: A State-of-the-Art Machine Learning Potential for Small Molecules
- The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models
- COACH: Toward the performance limit of traditional hybrid density functionals for molecular chemistry
- Egret-1: Pretrained Neural Network Potentials for Efficient and Accurate Bioorganic Simulation
- Orb-v3: atomistic simulation at scale
- UMA: A Family of Universal Models for Atoms
Related papers
- A System-Independent Metadynamics Strategy for Reactive Training Data: Application to Gas-Phase Organic Reactions
- A Machine Learning Model for the Chemistry of a Solvated Electron
- Compression of virtual spaces in transcorrelated methods via singular value decomposition: application to the G2 set
- Charge Transfer with a Spin. II: A Framework for Diabatization which Localizes Charge and Spin
- Diabatic Seam Space Sampling for Hydrogen Tunneling Systems with Nuclear-Electronic Orbital Theory
- Augmented Roothaan-Hall Hessian Applied to Spin-Restricted Open-Shell Density-Functional Theory