Skip to content
DraconiticResearch
Search Sources API
190,851 studies31 sourcesUpdated 11 min ago
Peer-reviewedCC BY-NC-NDnpj Computational Materials

A graph neural network for the era of large atomistic models

  • Duo Zhang1,2,3
  • Anyang Peng1
  • Chun Cai1,3
  • Wentao Li4
  • Yuanchang Zhou5,6
  • Jinzhe Zeng7,8,9
  • Mingyu Guo1,3,10
  • Chengqian Zhang1,2,3
  • Bowen Li11
  • Hong Yi Jiang12
  • Tong Zhu13,14,11
  • Weile Jia5,6
2 more authors
  • Linfeng Zhang1,3
  • Han Wang15,16
16 affiliations
  1. 1AI for Science Institute, Beijing, PR China
  2. 2Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, PR China
  3. 3DP Technology, Beijing, PR China
  4. 4Department of Chemical Engineering, Tsinghua University, Beijing, PR China
  5. 5State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, PR China
  6. 6University of Chinese Academy of Sciences, Beijing, PR China
  7. 7School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, PR China
  8. 8Suzhou Big Data & AI Research and Engineering Center, Suzhou, PR China
  9. 9Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, PR China
  10. 10School of Chemistry, Sun Yat-sen University, Guangzhou, PR China
  11. 11Shanghai Engineering Research Center of Molecular Therapeutics & New Drug Development, School of Chemistry and Molecular Engineering, East China Normal University, Shanghai, PR China
  12. 12College of Chemistry and Molecular Engineering, Peking University, Beijing, PR China
  13. 13Institute for Advanced algorithms research, Shanghai, PR China
  14. 14NYU-ECNU Center for Computational Chemistry at NYU Shanghai, Shanghai, PR China
  15. 15HEDPS, CAPT, College of Engineering, Peking University, Beijing, PR China
  16. 16National Key Laboratory of Computational Physics, Institute of Applied Physics and Computational Mathematics, Beijing, PR China
Open at sourcePDFDOICite

Abstract

Foundation models, or large atomistic models (LAMs), aim to universally represent the ground-state potential energy surface (PES) of atomistic systems as defined by density functional theory (DFT). The scaling law is pivotal in the development of large models, suggesting that their generalizability in downstream tasks consistently improves with increased model size, expanded training datasets, and larger computational budgets. In this study, we present DPA3, a multi-layer graph neural network founded on line graph series (LiGS), designed explicitly for the era of LAMs. We demonstrate that the generalization error of the DPA3 model adheres to the scaling law. The scalability in the number of model parameters is attained by stacking additional layers within DPA3. Additionally, the model employs a dataset encoding mechanism that decouples the scaling of training data size from the model size within its multi-task training framework. When trained as problem-oriented potential energy models, the DPA3 model exhibits superior accuracy in the majority of benchmark cases, encompassing systems with diverse features, including molecules, bulk materials, surface and cluster catalysis, two-dimensional materials, and battery materials. When trained as a LAM on the OpenLAM-v1 dataset, the DPA-3.1-3M model exhibits state-of-the-art performance in the LAMBench benchmark suite for LAMs, demonstrating lowest overall zero-shot generalization error across 17 downstream tasks from a broad spectrum of research domains. This performance suggests superior accuracy as an out-of-the-box potential model, requiring minimal fine-tuning data for downstream scientific applications.

Related

  1. SolviScope: An Interaction-Aware Graph Neural Network for Solvent Ranking of Small Organic Molecules in Pure and Multicomponent Solvent Systems
    Maansi Jain, Ayan Kashyap, Chanukya Nanduru +1
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  2. Why Multi-Layer Message Passing Works: Completeness Theory for Graph Neural Network Interatomic Potentials
    Pingbing Ming, Han Wang
    arXivPreprintCC BY-NC-SA 4.0Full textPhysical chemistryMathematical physicsarXiv
  3. Out-of-Distribution Inverse Design of Elastic Networks with Differentiable Graph Neural Network Molecular Dynamics
    Sergey A. Shteingolts, Salman N. Salman, Ron Levie +1
    arXivPreprintCC BY-NC-ND 4.0Full textPhysical chemistryarXiv
  4. Prediction of MOF Adsorption Isotherms and High-Throughput Screening Using a Physics-Informed Membranes Graph Neural Network
    Li Linfeng, Xingshuo Fan, Ning Zhang
    ChemRxivPreprintCC BYFull textMaterials & polymersChemRxiv
  5. Multi-Task Graph Neural Network Predictions of Auger-Electron and X-ray Photoelectron Spectroscopy
    Adam E. A. Fouda, Patrick Phillips, Phay J. Ho
    arXivPreprintCC BY 4.0Full textPhysical chemistryarXiv

Cite

BibTeXRIS
BibTeX
@article{zhang2026graph,
  title = {{A graph neural network for the era of large atomistic models}},
  author = {Duo Zhang and Anyang Peng and Chun Cai and Wentao Li and Yuanchang Zhou and Jinzhe Zeng and Mingyu Guo and Chengqian Zhang and Bowen Li and Hong Yi Jiang and Tong Zhu and Weile Jia and Linfeng Zhang and Han Wang},
  journal = {npj Computational Materials},
  year = {2026},
  month = may,
  doi = {10.1038/s41524-026-02146-2},
  url = {https://doi.org/10.1038/s41524-026-02146-2},
  note = {CC BY-NC-ND},
}
RIS
TY  - JOUR
TI  - A graph neural network for the era of large atomistic models
AU  - Duo Zhang
AU  - Anyang Peng
AU  - Chun Cai
AU  - Wentao Li
AU  - Yuanchang Zhou
AU  - Jinzhe Zeng
AU  - Mingyu Guo
AU  - Chengqian Zhang
AU  - Bowen Li
AU  - Hong Yi Jiang
AU  - Tong Zhu
AU  - Weile Jia
AU  - Linfeng Zhang
AU  - Han Wang
PY  - 2026
DA  - 2026/05/25
JO  - npj Computational Materials
DO  - 10.1038/s41524-026-02146-2
UR  - https://doi.org/10.1038/s41524-026-02146-2
AB  - Foundation models, or large atomistic models (LAMs), aim to universally represent the ground-state potential energy surface (PES) of atomistic systems as defined by density functional theory (DFT). The scaling law is pivotal in the development of large models, suggesting that their generalizability in downstream tasks consistently improves with increased model size, expanded training datasets, and larger computational budgets. In this study, we present DPA3, a multi-layer graph neural network founded on line graph series (LiGS), designed explicitly for the era of LAMs. We demonstrate that the generalization error of the DPA3 model adheres to the scaling law. The scalability in the number of model parameters is attained by stacking additional layers within DPA3. Additionally, the model employs a dataset encoding mechanism that decouples the scaling of training data size from the model size within its multi-task training framework. When trained as problem-oriented potential energy models, the DPA3 model exhibits superior accuracy in the majority of benchmark cases, encompassing systems with diverse features, including molecules, bulk materials, surface and cluster catalysis, two-dimensional materials, and battery materials. When trained as a LAM on the OpenLAM-v1 dataset, the DPA-3.1-3M model exhibits state-of-the-art performance in the LAMBench benchmark suite for LAMs, demonstrating lowest overall zero-shot generalization error across 17 downstream tasks from a broad spectrum of research domains. This performance suggests superior accuracy as an out-of-the-box potential model, requiring minimal fine-tuning data for downstream scientific applications.
SN  - 2057-3960
ER  -