Atomic Energy Accuracy of Neural Network Potentials: Harnessing Pretraining and Transfer Learning

Gang Seob Jung

doi:10.1021/acs.jcim.5c00079

Atomic Energy Accuracy of Neural Network Potentials: Harnessing Pretraining and Transfer Learning

Gang Seob Jung

Computational Chemistry and Nanomaterial

Research output: Contribution to journal › Article › peer-review

2 Scopus citations

Abstract

Machine learning-based interatomic potentials (MLIPs) have transformed the prediction of potential energy surfaces (PESs), achieving accuracy comparable to ab initio calculations. However, atomic energy predictions, often assumed to lack physical meaning, remain underexplored. In this study, we demonstrate that inaccuracies in atomic energy predictions reduce the robustness and transferability of Neural Network Potentials (NNPs) and atomic energy error can be masked in total energy predictions due to error cancellation. We validate this finding using challenging configurations involving deformation and failure under tensile loading. By pretraining atomic energy predictions using empirical potentials and applying transfer learning with density functional theory (DFT) data, we achieve notable improvements in the accuracy of total energy, forces, and stress predictions. Furthermore, this approach enhances the robustness and transferability of NNPs, emphasizing the importance of atomic energy predictions in developing high-quality and reliable MLIPs.

Original language	English
Pages (from-to)	4797-4807
Number of pages	11
Journal	Journal of Chemical Information and Modeling
Volume	65
Issue number	10
DOIs	https://doi.org/10.1021/acs.jcim.5c00079
State	Published - May 26 2025

Funding

The authors acknowledge support by the Laboratory Directed Research and Development Program (LDRD) of Oak Ridge National Laboratory (Enzyme Initiative), managed by UT-Battelle, LLC, for the US Department of Energy under contract DEAC05-00OR22725. This research used resources from the Compute and Data Environment for Science (CADES) at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725 and National Energy Research Scientific Computing Center (NERSC), a DOE Office of Science User Facility for access to additional supercomputing resources. This work is a part of a user project at the Center for Nanophase Materials Sciences (CNMS), a US Department of Energy, Office of Science User Facility at Oak Ridge National Laboratory. This manuscript has been authored by UT-Battelle, LLC, under Contract No. DE-AC05–00OR22725 with the U.S. Department of Energy. The United States Government retains and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. The Department of Energy will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan ( http://energy.gov/downloads/doe-public-access-plan ).

Access to Document

10.1021/acs.jcim.5c00079

Cite this

@article{dbc8a06b815145068da2614acf66d8e1,

title = "Atomic Energy Accuracy of Neural Network Potentials: Harnessing Pretraining and Transfer Learning",

abstract = "Machine learning-based interatomic potentials (MLIPs) have transformed the prediction of potential energy surfaces (PESs), achieving accuracy comparable to ab initio calculations. However, atomic energy predictions, often assumed to lack physical meaning, remain underexplored. In this study, we demonstrate that inaccuracies in atomic energy predictions reduce the robustness and transferability of Neural Network Potentials (NNPs) and atomic energy error can be masked in total energy predictions due to error cancellation. We validate this finding using challenging configurations involving deformation and failure under tensile loading. By pretraining atomic energy predictions using empirical potentials and applying transfer learning with density functional theory (DFT) data, we achieve notable improvements in the accuracy of total energy, forces, and stress predictions. Furthermore, this approach enhances the robustness and transferability of NNPs, emphasizing the importance of atomic energy predictions in developing high-quality and reliable MLIPs.",

author = "Jung, \{Gang Seob\}",

year = "2025",

month = may,

day = "26",

doi = "10.1021/acs.jcim.5c00079",

language = "English",

volume = "65",

pages = "4797--4807",

journal = "Journal of Chemical Information and Modeling",

issn = "1549-9596",

publisher = "American Chemical Society",

number = "10",

}

TY - JOUR

T1 - Atomic Energy Accuracy of Neural Network Potentials

T2 - Harnessing Pretraining and Transfer Learning

AU - Jung, Gang Seob

PY - 2025/5/26

Y1 - 2025/5/26

N2 - Machine learning-based interatomic potentials (MLIPs) have transformed the prediction of potential energy surfaces (PESs), achieving accuracy comparable to ab initio calculations. However, atomic energy predictions, often assumed to lack physical meaning, remain underexplored. In this study, we demonstrate that inaccuracies in atomic energy predictions reduce the robustness and transferability of Neural Network Potentials (NNPs) and atomic energy error can be masked in total energy predictions due to error cancellation. We validate this finding using challenging configurations involving deformation and failure under tensile loading. By pretraining atomic energy predictions using empirical potentials and applying transfer learning with density functional theory (DFT) data, we achieve notable improvements in the accuracy of total energy, forces, and stress predictions. Furthermore, this approach enhances the robustness and transferability of NNPs, emphasizing the importance of atomic energy predictions in developing high-quality and reliable MLIPs.

AB - Machine learning-based interatomic potentials (MLIPs) have transformed the prediction of potential energy surfaces (PESs), achieving accuracy comparable to ab initio calculations. However, atomic energy predictions, often assumed to lack physical meaning, remain underexplored. In this study, we demonstrate that inaccuracies in atomic energy predictions reduce the robustness and transferability of Neural Network Potentials (NNPs) and atomic energy error can be masked in total energy predictions due to error cancellation. We validate this finding using challenging configurations involving deformation and failure under tensile loading. By pretraining atomic energy predictions using empirical potentials and applying transfer learning with density functional theory (DFT) data, we achieve notable improvements in the accuracy of total energy, forces, and stress predictions. Furthermore, this approach enhances the robustness and transferability of NNPs, emphasizing the importance of atomic energy predictions in developing high-quality and reliable MLIPs.

UR - https://www.scopus.com/pages/publications/105004910841

U2 - 10.1021/acs.jcim.5c00079

DO - 10.1021/acs.jcim.5c00079

M3 - Article

C2 - 40320921

AN - SCOPUS:105004910841

SN - 1549-9596

VL - 65

SP - 4797

EP - 4807

JO - Journal of Chemical Information and Modeling

JF - Journal of Chemical Information and Modeling

IS - 10

ER -

Atomic Energy Accuracy of Neural Network Potentials: Harnessing Pretraining and Transfer Learning

Abstract

Funding

Access to Document

Other files and links

Fingerprint

Cite this