Machine-Learned Potentials for Accelerated CCSD(T)-level Evaluations: Applications to OH(H2O)n
- 1Department of Chemistry
- 2University of Washington
- 3Hong Kong University of Science and Technology
Abstract
An end-to-end machine learning workflow for the development of potential energy surfaces at the DLPNO-CCSD(T)/aug-cc-pVTZ level is described and applied to hydroxide-water complexes with two to five water molecules. This workflow utilizes a two-part machine learning process, achieving both accuracy and efficiency by combining a molecular orbital-based machine learning (MOB-ML) model with a GPU-accelerated equivariant graph neural network (EGNN) model. For both machine learning models trained in this workflow, the training data is obtained through the combination of small diffusion Monte Carlo (DMC) simulations initiated at each of the optimized structures of interest for a given system size. A farthest point sampling algorithm based on a principal component analysis is employed to generate a diverse set of geometries that spans configuration space, including transition states. By transforming Cartesian coordinates into a graph-based representation, the EGNN model eliminates additional preprocessing of molecular configurations, provides an unbiased treatment of multiple isomers, and allows for isomerization processes among the isomers. An error analysis is performed on the final EGNN models, comparing the potential energies predicted for optimized structures, transition states, and structures that are obtained from snapshots of the ground state wave function sampled by large-scale DMC simulations to those evaluated at the DLPNO-CCSD(T)/aug-ccpVTZ level of theory and basis. Finally, production-run DMC simulations are performed using the validated potentials for both the OH(H 2 O) 2-5 and OD(D 2 O) 2-5 systems, and an analysis of the relative anharmonic zero-point energies is performed. It is found that deuteration does not change the relative energy ordering of the isomers of the OH(H 2 O) 3-5 complexes. Evidence of proton delocalization is identified in the prism isomer of OH(H 2 O) 5 in which a hydrogen atom in one of the solvating water molecules is found to be, on average, equidistant from the oxygen atoms in the hydroxide ion and the hydrogen-bonded water molecule. This delocalization is found to be retained with full deuteration of this complex.
Related
- SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials
- From Energy-Force Weighting to Primal-Dual Optimization of Machine-Learned Interatomic Potentials
- Dielectric Response of Short and Long Range Models of Nonuniform Liquids and Implications for Machine Learned Interatomic Potentials
- Linear response for machine-learned interatomic potentials: phonons at arbitrary wave vectors from the primitive cell
- Machine Learning Local Potentials for Accelerated Electron-Phonon Interactions Calculations within the Projector Augmented-Wave Framework
BibTeX
@misc{jacobson2026machinelearned,
title = {{Machine-Learned Potentials for Accelerated CCSD(T)-level Evaluations: Applications to OH(H2O)n}},
author = {Greta M. Jacobson and Chenrui Shao and Tanya I. Bakalov and Lixue Cheng and Anne B. McCoy},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008636/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008636/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Machine-Learned Potentials for Accelerated CCSD(T)-level Evaluations: Applications to OH(H2O)n
AU - Greta M. Jacobson
AU - Chenrui Shao
AU - Tanya I. Bakalov
AU - Lixue Cheng
AU - Anne B. McCoy
PY - 2026
DA - 2026/09/10
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008636/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008636/v1
AB - An end-to-end machine learning workflow for the development of potential energy surfaces at the DLPNO-CCSD(T)/aug-cc-pVTZ level is described and applied to hydroxide-water complexes with two to five water molecules. This workflow utilizes a two-part machine learning process, achieving both accuracy and efficiency by combining a molecular orbital-based machine learning (MOB-ML) model with a GPU-accelerated equivariant graph neural network (EGNN) model. For both machine learning models trained in this workflow, the training data is obtained through the combination of small diffusion Monte Carlo (DMC) simulations initiated at each of the optimized structures of interest for a given system size. A farthest point sampling algorithm based on a principal component analysis is employed to generate a diverse set of geometries that spans configuration space, including transition states. By transforming Cartesian coordinates into a graph-based representation, the EGNN model eliminates additional preprocessing of molecular configurations, provides an unbiased treatment of multiple isomers, and allows for isomerization processes among the isomers. An error analysis is performed on the final EGNN models, comparing the potential energies predicted for optimized structures, transition states, and structures that are obtained from snapshots of the ground state wave function sampled by large-scale DMC simulations to those evaluated at the DLPNO-CCSD(T)/aug-ccpVTZ level of theory and basis. Finally, production-run DMC simulations are performed using the validated potentials for both the OH(H 2 O) 2-5 and OD(D 2 O) 2-5 systems, and an analysis of the relative anharmonic zero-point energies is performed. It is found that deuteration does not change the relative energy ordering of the isomers of the OH(H 2 O) 3-5 complexes. Evidence of proton delocalization is identified in the prism isomer of OH(H 2 O) 5 in which a hydrogen atom in one of the solvating water molecules is found to be, on average, equidistant from the oxygen atoms in the hydroxide ion and the hydrogen-bonded water molecule. This delocalization is found to be retained with full deuteration of this complex.
ER -