Skip to content
DraconiticResearch
Search Sources API
186,631 studies31 sourcesUpdated 12 min ago
PreprintCC BY-NC-NDChemRxiv

Active Atom Identification for Information-Dense Training of Machine-Learning Potentials in Catalysis

  • Haobo Zhao1,2
  • Xiongwei Tian3,2
  • Jixuan Lin1,2
  • Yuqun Zhuo3,2
  • Yiming Lu4,5
  • Hai Xiao1,2
  • Jun Li1,2,6,7,8,9
  • Zhao Haobo
9 affiliations
  1. 1Department of Chemistry
  2. 2Tsinghua University
  3. 3Key Laboratory for Thermal Science and Power Engineering of Ministry of Education, Department of Energy and Power Engineering
  4. 4Advanced Institute for Materials Research (WPI-AIMR
  5. 5Tohoku University
  6. 6Ganjiang Innovation Academy
  7. 7Chinese Academy of Science
  8. 8Department of Chemistry and Guangdong Provincial Key Laboratory of Catalytic Chemistry
  9. 9Southern University of Science and Technology
Open at sourcePDFDOICite

Abstract

Machine-learning potentials (MLPs) extend atomistic simulations beyond the size and time scales accessible to density functional theory while retaining near-first-principles accuracy. However, their reliability depends on training data that cover the relevant potential-energy surface (PES). In heterogeneous catalysis, local changes at active sites may be obscured by redundant nonactive atoms when configurations are selected using global metrics. This limitation calls for an atom-resolved criterion for identifying where chemically relevant configurational information is concentrated. Here we introduce Active Atom Identification (AAI), an algorithm that measures local structural evolution and identifies active atoms that contribute most to configurational diversity. Building on this atom-resolved information, we develop IDEAL (Information-Dense Extraction and Adaptive Learning), which integrates data generation, active-learning selection and dataset refinement into a unified workflow for MLP construction. For CO2 adsorption at Ag/H2O interfaces, a model trained on an IDEAL-selected subset comprising 44% of the original dataset, reaches the accuracy of the full-dataset model. The broad applicability of IDEAL is further demonstrated in two additional applications: database refinement for solid-state electrolytes and transfer learning between zeolite-supported Rh clusters and their hydroxylated analogues. These results validate local structural evolution as an effective criterion for constructing information-dense MLP datasets in complex atomistic systems.

Related

  1. Simulating enzyme catalysis with electrostatically embedded machine learning potentials
    Valentin Gradisteanu, Elliot W Chan, Lester Hedges +7
    Chemical SciencePeer-reviewedCC BY 4.0Full text15 citedMaterials & polymersHAL
  2. Smarter Data: Rethinking Data Generation for Machine Learning Potentials in Heterogeneous Catalysis
    Wenbo Xie, Yixiao Han, Chenyu Wu +1
    JACS AuPeer-reviewedCC BY-NC-ND5 citedMaterials & polymersOpenAlex
  3. Folding and Function as Two Sides of One Coin: Statistical Mechanics, Information Theory, and Deep Learning of Binding, Catalysis, and Allostery in an Exactly Enumerable Lattice Protein
    Shubin Liu
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  4. Machine Learning Interatomic Potentials for Energy Materials: Architectures, Training Strategies, and Applications
    In Kee Park, Saeed Pourasad, Jinhong Mun +17
    Advanced Energy MaterialsPeer-reviewedCC BY-NC15 citedMaterials & polymersOpenAlex
  5. Towards Out-of-Distribution Uncertainty Quantification for Machine-Learned Interatomic Potentials via Pre-Training
    Stefan P. Schmid, Tobias Setzer, Maximilian Scheurer
    arXivPreprintCC BY 4.0Full textPhysical chemistryarXiv

Cite

BibTeXRIS
BibTeX
@misc{zhao2026active,
  title = {{Active Atom Identification for Information-Dense Training of Machine-Learning Potentials in Catalysis}},
  author = {Haobo Zhao and Xiongwei Tian and Jixuan Lin and Yuqun Zhuo and Yiming Lu and Hai Xiao and Jun Li and Zhao Haobo},
  howpublished = {ChemRxiv},
  year = {2026},
  month = sep,
  doi = {10.26434/chemrxiv.15009531/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009531/v1},
  note = {CC BY-NC-ND},
}
RIS
TY  - UNPB
TI  - Active Atom Identification for Information-Dense Training of Machine-Learning Potentials in Catalysis
AU  - Haobo Zhao
AU  - Xiongwei Tian
AU  - Jixuan Lin
AU  - Yuqun Zhuo
AU  - Yiming Lu
AU  - Hai Xiao
AU  - Jun Li
AU  - Zhao Haobo
PY  - 2026
DA  - 2026/09/28
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15009531/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009531/v1
AB  - Machine-learning potentials (MLPs) extend atomistic simulations beyond the size and time scales accessible to density functional theory while retaining near-first-principles accuracy. However, their reliability depends on training data that cover the relevant potential-energy surface (PES). In heterogeneous catalysis, local changes at active sites may be obscured by redundant nonactive atoms when configurations are selected using global metrics. This limitation calls for an atom-resolved criterion for identifying where chemically relevant configurational information is concentrated. Here we introduce Active Atom Identification (AAI), an algorithm that measures local structural evolution and identifies active atoms that contribute most to configurational diversity. Building on this atom-resolved information, we develop IDEAL (Information-Dense Extraction and Adaptive Learning), which integrates data generation, active-learning selection and dataset refinement into a unified workflow for MLP construction. For CO2 adsorption at Ag/H2O interfaces, a model trained on an IDEAL-selected subset comprising 44% of the original dataset, reaches the accuracy of the full-dataset model. The broad applicability of IDEAL is further demonstrated in two additional applications: database refinement for solid-state electrolytes and transfer learning between zeolite-supported Rh clusters and their hydroxylated analogues. These results validate local structural evolution as an effective criterion for constructing information-dense MLP datasets in complex atomistic systems.
ER  -