Skip to content
DraconiticResearch
Search Sources API
226,742 studies31 sourcesUpdated 43 min ago
Peer-reviewedCC BY 4.0Journal of Chemical Theory and Computation

Convolutional Kernel-Embedded E(3)-Equivariant Networks (KEN): Overcoming Architectural Rigidity in Machine Learning Interatomic Potentials

  1. 1Hylleraas Centre for Quantum Molecular Sciences, Department of Chemistry, University of Oslo, Postboks 1033 Blindern, 0315 Oslo, Norway

Abstract

Machine learning interatomic potentials (MLIPs) have emerged as scalable alternatives to first-principles methods such as density functional theory (DFT). Among them, E(3)-equivariant graph neural networks (GNNs) like PACE and MACE are highly accurate but structurally rigid, limiting the incorporation of rich, multichannel electronic descriptors. We introduce convolutional kernel-embedded E(3)-equivariant networks (KEN), a hybrid architecture that combines the local environmental extraction of 2D convolutional kernels with the many-body expressivity of an E(3)-equivariant GNN backbone. This design preserves symmetry while enabling flexible, multimodal input features without increasing training data requirements. Trained and tested on heavy element 26 (HE26) dataset, KEN was compared with MACE-osaka26 model. Further validation on lattice parameter prediction through Birch–Murnaghan equation of state, magnetic moment prediction, and comparison with CHGNet framework, and microcanonical molecular dynamics confirms robust energy conservation and accurate reproduction of the forces, and thus, demonstrates a viable pathway to overcoming architectural rigidity in state-of-the-art MLIPs.

Related

  1. Integrating Charge Equilibration with Equivariant Machine-Learning Interatomic Potentials
    Martin Vondrák, William J. Baldwin, Gábor Cśanyi +2
    Journal of Chemical Theory and ComputationPeer-reviewedLicence not statedFull text6 citedMaterials & polymersOpenAlex
  2. BranchIP: Learning Adaptive Equivariant Computation for Interatomic Potentials
    Laura Zichi, Gil Harari, Chuin Wei Tan +6
    arXivPreprintarXiv licenceFull textPhysical chemistryApplied physics & instrumentationarXiv
  3. Correcting the Liquid Densities of Universal Machine-Learning Interatomic Potentials Using Local-Number-Density Repulsion
    Shunsuke Mieda
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  4. Foundation interatomic potentials learn the practical chemistry they were never taught
    Alexander A. Kolganov, Mas P. Klein, Evgeny A. Pidko
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  5. pdb2reaction: End-to-End Reaction-Path Elucidation from PDB Structures Using Machine-Learning Interatomic Potentials
    Takuto Ohmura, Hajime Sato, Tohru Terada
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
BibTeX
@article{gautam2026convolutional,
  title = {{Convolutional Kernel-Embedded E(3)-Equivariant Networks (KEN): Overcoming Architectural Rigidity in Machine Learning Interatomic Potentials}},
  author = {Jha Gautam},
  journal = {Journal of Chemical Theory and Computation},
  year = {2026},
  month = oct,
  doi = {10.1021/acs.jctc.6c01109},
  url = {https://doi.org/10.1021/acs.jctc.6c01109},
  note = {CC BY 4.0},
}
RIS
TY  - JOUR
TI  - Convolutional Kernel-Embedded E(3)-Equivariant Networks (KEN): Overcoming Architectural Rigidity in Machine Learning Interatomic Potentials
AU  - Jha Gautam
PY  - 2026
DA  - 2026/10/02
JO  - Journal of Chemical Theory and Computation
DO  - 10.1021/acs.jctc.6c01109
UR  - https://doi.org/10.1021/acs.jctc.6c01109
AB  - Machine learning interatomic potentials (MLIPs) have emerged as scalable alternatives to first-principles methods such as density functional theory (DFT). Among them, E(3)-equivariant graph neural networks (GNNs) like PACE and MACE are highly accurate but structurally rigid, limiting the incorporation of rich, multichannel electronic descriptors. We introduce convolutional kernel-embedded E(3)-equivariant networks (KEN), a hybrid architecture that combines the local environmental extraction of 2D convolutional kernels with the many-body expressivity of an E(3)-equivariant GNN backbone. This design preserves symmetry while enabling flexible, multimodal input features without increasing training data requirements. Trained and tested on heavy element 26 (HE26) dataset, KEN was compared with MACE-osaka26 model. Further validation on lattice parameter prediction through Birch–Murnaghan equation of state, magnetic moment prediction, and comparison with CHGNet framework, and microcanonical molecular dynamics confirms robust energy conservation and accurate reproduction of the forces, and thus, demonstrates a viable pathway to overcoming architectural rigidity in state-of-the-art MLIPs.
SN  - 1549-9618
SN  - 1549-9626
ER  -