pith. machine review for the scientific record. sign in

arxiv: 2604.24816 · v1 · submitted 2026-04-27 · ❄️ cond-mat.mtrl-sci

Recognition: unknown

Trillion-atom molecular dynamics simulations with ab initio accuracy

Pengfei Suo , Wudi Cao , Xingxing Wu , Wenjie Zhang , Zheyong Fan , Shuanghan Xian , Rui Wang , Cheng Qian
show 9 more authors
Chao Liang Qinghong Yuan Xiaoshuang Chen Pengfei Guan Jingde Bu Hongzhen Tian Yanjing Su Feng Ding Lin-Wang Wang
Authors on Pith no claims yet

Pith reviewed 2026-05-08 02:52 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords molecular dynamicsneuroevolution potentialtrillion atomsab initio accuracymesoscale simulationmachine learning force fieldsparallel scalingmaterial modeling
0
0 comments X

The pith

A molecular dynamics simulation of 1.62 trillion atoms achieves ab initio accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that molecular dynamics can be performed on 1.62 trillion atoms while maintaining the accuracy of quantum calculations through the neuroevolution potential framework. This scale reaches the mesoscale, bridging the gap between atomic-level details and observable material properties under microscopes. The simulation achieves substantial speed improvements and efficient scaling across many processors. If successful, it allows for direct quantum-precise modeling of complex, real-world material phenomena at previously inaccessible sizes.

Core claim

The authors report an unprecedented molecular dynamics simulation of 1.62 trillion atoms using the neuroevolution potential framework to attain ab initio accuracy. Their implementation delivers a time-to-solution 100 times faster than previous machine learning force field simulations and demonstrates 86.9 percent weak scaling efficiency from a single processor to 45,000 processors. This redefines the limits of atomistic simulations by enabling mesoscopic modeling with quantum-level precision.

What carries the argument

The neuroevolution potential (NEP) framework, which serves as an efficient machine learning surrogate for quantum forces in molecular dynamics.

If this is right

  • Mesoscale material properties can be modeled directly from atomic interactions without intermediate approximations.
  • Multiscale phenomena in materials become simulatable at full atomistic resolution.
  • Computational resources for scientific simulations can handle systems three orders of magnitude larger than before.
  • High parallel efficiency supports utilization of large computing facilities for single simulations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Such simulations could be used to study how defects evolve over mesoscale distances in real materials.
  • The method might integrate with imaging techniques to provide atomic explanations for observed mesoscale structures.
  • Transferability of the trained model to untested material systems remains to be verified in larger contexts.

Load-bearing premise

The neuroevolution potential model trained on smaller ab initio datasets retains full accuracy when applied to a trillion-atom mesoscale system.

What would settle it

A direct comparison of the large-scale simulation results against known experimental mesoscale properties or smaller ab initio benchmarks for the same material would disprove the accuracy claim if significant discrepancies appear.

Figures

Figures reproduced from arXiv: 2604.24816 by Chao Liang, Cheng Qian, Feng Ding, Hongzhen Tian, Jingde Bu, Lin-Wang Wang, Pengfei Guan, Pengfei Suo, Qinghong Yuan, Rui Wang, Shuanghan Xian, Wenjie Zhang, Wudi Cao, Xiaoshuang Chen, Xingxing Wu, Yanjing Su, Zheyong Fan.

Figure 1
Figure 1. Figure 1: Schematic illustration of the NEP architecture. (a) The dashed circles indicate the different cutoff radii for the radial and angular descriptors, denoted Rc and rc, respectively. (b) The input layer comprising radial (qn) and angular (qnl) descriptors is mapped to the atomic energy Ui through a hidden layer. As shown in view at source ↗
Figure 3
Figure 3. Figure 3: Training results of NEP model for H2O. (a)-(c) The comparison between the MatPL/NEP predictions and DFT values of energy, force and virial. We performed scaling measurements across all three target systems—water, copper, and hafnium dioxide—to assess the parallel efficiency of our NEP-based MD simulations. For water, strong scaling was evaluated on a system containing 110,578,630,656 atoms. In weak‑scaling… view at source ↗
Figure 5
Figure 5. Figure 5: Strong scaling: (a) the water system of 110,578,630,656 atoms. (b) the HfO view at source ↗
read the original abstract

Material properties are fundamentally dictated by multiscale phenomena, which often reach mesoscale in size. The {\mu}m mesoscale is also the size which can be observed directly under an optical microscope, bridging the atomistic microscopic description with the continuous model macroscopic world. In this work, we report an unprecedented molecular dynamics (MD) simulation comprising 1.62 trillion atoms. Utilizing the neuroevolution potential (NEP) framework, we attained ab initio accuracy on China's New-generation Intelligent Supercomputer. Our implementation achieves a time-to-solution (s/step/atom) 100 times faster than previous state-of-the-art machine learning force field simulations, and 1,000 times faster than the Gordon Bell Prize-winning application from six years ago. Furthermore, we demonstrate an 86.9% weak scaling efficiency from a single GPGPU to 45,000 GPGPUs. These results redefine atomistic simulation boundaries, enabling direct mesoscopic modeling with quantum-level precision.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript reports an unprecedented molecular dynamics simulation of 1.62 trillion atoms using the neuroevolution potential (NEP) framework to achieve ab initio accuracy. It demonstrates a 100-fold speedup in time-to-solution over prior machine-learning force-field simulations and 86.9% weak-scaling efficiency from one to 45,000 GPGPUs on China's New-generation Intelligent Supercomputer, enabling direct mesoscale modeling with quantum-level precision.

Significance. If the accuracy claim is substantiated, the work would substantially advance computational materials science by bridging atomistic and mesoscopic scales, allowing quantum-accurate simulations of phenomena observable under optical microscopes. The scaling performance itself is a notable engineering achievement that expands the practical reach of atomistic methods.

major comments (2)
  1. [§4] §4 (Accuracy and Validation): The claim of 'ab initio accuracy' for the 1.62-trillion-atom system rests on transferability of an NEP model trained on smaller ab initio datasets, yet no quantitative error metrics (force RMSE, energy errors, or mesoscale property comparisons such as phonon dispersion or defect formation energies) are shown for system sizes approaching the trillion-atom regime. Direct ab initio references are impossible at this scale, so the central assertion requires explicit tests demonstrating absence of scale-dependent errors in collective modes or long-range correlations.
  2. [§5.3] §5.3 (Weak Scaling): The reported 86.9% efficiency is load-bearing for the feasibility claim, but the manuscript does not specify atoms per GPGPU, communication volume, or load-balancing details that would allow independent assessment of whether the parallel implementation introduces numerical artifacts at 45,000 GPGPUs.
minor comments (2)
  1. [Abstract] The abstract and introduction should include a brief statement of the training dataset size and composition (number of ab initio structures, elements covered) to contextualize the transferability assumption.
  2. [Figure 3] Figure captions for scaling plots should explicitly state the per-GPGPU atom count and the observable used to measure time-to-solution.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and positive assessment of the work's significance. We address each major comment below with clarifications and proposed revisions.

read point-by-point responses

  1. Referee: [§4] §4 (Accuracy and Validation): The claim of 'ab initio accuracy' for the 1.62-trillion-atom system rests on transferability of an NEP model trained on smaller ab initio datasets, yet no quantitative error metrics (force RMSE, energy errors, or mesoscale property comparisons such as phonon dispersion or defect formation energies) are shown for system sizes approaching the trillion-atom regime. Direct ab initio references are impossible at this scale, so the central assertion requires explicit tests demonstrating absence of scale-dependent errors in collective modes or long-range correlations.

    Authors: We agree that direct ab initio calculations at the trillion-atom scale are impossible and that the accuracy claim relies on model transferability. The manuscript validates the NEP model using force and energy errors on held-out ab initio datasets from smaller systems, along with mesoscale properties including phonon dispersions and defect formation energies. To address scale-dependent errors, we will add an explicit discussion in the revised §4 of how validated properties (e.g., long-range correlations in phonon modes) remain consistent across system sizes up to the largest computationally accessible regimes, thereby supporting transferability to the trillion-atom simulation. revision: partial

  2. Referee: [§5.3] §5.3 (Weak Scaling): The reported 86.9% efficiency is load-bearing for the feasibility claim, but the manuscript does not specify atoms per GPGPU, communication volume, or load-balancing details that would allow independent assessment of whether the parallel implementation introduces numerical artifacts at 45,000 GPGPUs.

    Authors: We will revise §5.3 to include the requested implementation details: each GPGPU handles approximately 36 million atoms in the largest run, with communication volumes dominated by halo exchanges of neighbor lists and forces (quantified per step), and load balancing achieved via spatial domain decomposition with periodic rebalancing. These additions will allow assessment of numerical stability, which is further supported by reported energy conservation in the production runs. revision: yes

Circularity Check

0 steps flagged

No circularity in the reported simulation scaling demonstration

full rationale

The paper reports an engineering and computational achievement: a 1.62-trillion-atom MD run using a pre-trained neuroevolution potential (NEP) model on a large supercomputer, together with measured wall-clock performance and weak-scaling efficiency. No derivation chain is presented in which a claimed first-principles result or prediction is shown by the paper's own equations to be identical to its inputs by construction. The NEP accuracy claim rests on prior training against ab initio data for smaller systems; the trillion-atom run itself is an application, not a re-derivation that forces the accuracy result. No self-definitional steps, fitted-input-as-prediction maneuvers, or load-bearing self-citation chains appear in the abstract or described methodology. The work is therefore self-contained as a scaling demonstration.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract provides no explicit free parameters, axioms, or invented entities; the NEP accuracy claim implicitly rests on prior training of the potential on ab initio data whose details are not shown here.

pith-pipeline@v0.9.0 · 5522 in / 1179 out tokens · 41506 ms · 2026-05-08T02:52:48.736356+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

  1. [1]

    and Bacon, D.J

    Hull, D. and Bacon, D.J. eds. 2011. Introduction to dislocations. Butterworth-Heinemann

    2011

  2. [2]

    and Kang, K

    Jung, S., Gwon, H., Hong, J., Park, K., Seo, D., Kim, H., Hyun, J., Yang, W. and Kang, K. 2014. Understanding the Degradation Mechanisms of LiNi0.5 Co0.2 Mn0.3 O2 Cathode Material in Lithium Ion Batteries. Advanced Energy Materials. 4, 1 (Jan. 2014), 1300787. https://doi.org/10.1002/aenm.201300787

    work page doi:10.1002/aenm.201300787 2014

  3. [3]

    and Huang, W

    Li, J. and Huang, W. 2018. From Multiscale to Mesoscience: Addressing Mesoscales in Mesoregimes of Different Levels. Annual Review of Chemical and Biomolecular Engineering. 9, 1 (June 2018), 41–60. https://doi.org/10.1146/annurev-chembioeng-060817-084249

    work page doi:10.1146/annurev-chembioeng-060817-084249 2018

  4. [4]

    Behler, J. 2021. Four Generations of High-Dimensional Neural Network Potentials. Chemical Reviews. 121, 16 (Aug. 2021), 10037– 10072. https://doi.org/10.1021/acs.chemrev.0c00868

    work page doi:10.1021/acs.chemrev.0c00868 2021

  5. [5]

    and Chmiela, Stefan and Sauceda, Huziel E

    Unke, O.T., Chmiela, S., Sauceda, H.E., Gastegger, M., Poltavsky, I., Schü tt, K.T., Tkatchenko, A. and Mü ller, K.-R. 2021. Machine Learning Force Fields. Chemical Reviews. 121, 16 (Aug. 2021), 10142–10186. https://doi.org/10.1021/acs.chemrev.0c01111

    work page doi:10.1021/acs.chemrev.0c01111 2021

  6. [6]

    and Behler, J

    Kocer, E., Ko, T.W. and Behler, J. 2022. Neural Network Potentials: A Concise Overview of Methods. Annual Review of Physical Chemistry. 73, 1 (Apr. 2022), 163–186. https://doi.org/10.1146/annurev-physchem-082720-034254

    work page doi:10.1146/annurev-physchem-082720-034254 2022

  7. [7]

    and Jiang, B

    Xia, J., Zhang, Y. and Jiang, B. 2025. The evolution of machine learning potentials for molecules, reactions and materials. Chemical Society Reviews. 54, 10 (2025), 4790–4821. https://doi.org/10.1039/D5CS00104H

    work page doi:10.1039/d5cs00104h 2025

  8. [8]

    and Ala-Nissila, T

    Fan, Z., Zeng, Z., Zhang, C., Wang, Y., Song, K., Dong, H., Chen, Y. and Ala-Nissila, T. 2021. Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport. Physical Review B. 104, 10 (Sept. 2021), 104309. https://doi.org/10.1103/PhysRevB.104.104309

    work page doi:10.1103/physrevb.104.104309 2021

  9. [9]

    and Wang, L

    Suo, P., Wu, X., Tian, H. and Wang, L. 2026. Towards Scalable and Efficient Machine-Learning Force Fields: The MatPL package and Its Advancements on Neuroevolution Potentials. ChemRxiv. (2026). https://doi.org/https://doi.org/10.26434/chemrxiv.15001665

    work page doi:10.26434/chemrxiv.15001665 2026

  10. [10]

    Thompson, A.P. et al. 2022. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications. 271, (Feb. 2022), 108171. https://doi.org/10.1016/j.cpc.2021.108171

    work page doi:10.1016/j.cpc.2021.108171 2022

  11. [11]

    Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces.Physical Review Letters, 98(14):146401, April 2007

    Behler, J. and Parrinello, M. 2007. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces. Physical Review Letters. 98, 14 (Apr. 2007), 146401. https://doi.org/10.1103/PhysRevLett.98.146401

    work page doi:10.1103/physrevlett.98.146401 2007

  12. [12]

    Gaussian

    Bartó k, A.P., Payne, M.C., Kondor, R. and Csá nyi, G. 2010. Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons. Physical Review Letters. 104, 13 (Apr. 2010), 136403. https://doi.org/10.1103/PhysRevLett.104.136403

    work page doi:10.1103/physrevlett.104.136403 2010

  13. [13]

    and Tucker, G.J

    Thompson, A.P., Swiler, L.P., Trott, C.R., Foiles, S.M. and Tucker, G.J. 2015. Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials. Journal of Computational Physics. 285, (Mar. 2015), 316–330. https://doi.org/10.1016/j.jcp.2014.12.018

    work page doi:10.1016/j.jcp.2014.12.018 2015

  14. [14]

    DeePMD-kit : A deep learning package for many-body potential energy representation and molecular dynamics

    Wang, H., Zhang, L., Han, J. and E, W. 2018. DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics. Computer Physics Communications. 228, (July 2018), 178–184. https://doi.org/10.1016/j.cpc.2018.03.016

    work page doi:10.1016/j.cpc.2018.03.016 2018

  15. [15]

    Learning local equivariant representations for large-scale atomistic dynamics,

    Musaelian, A., Batzner, S., Johansson, A., Sun, L., Owen, C.J., Kornbluth, M. and Kozinsky, B. 2023. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications. 14, 1 (Feb. 2023), 579. https://doi.org/10.1038/s41467-023-36329-y

    work page doi:10.1038/s41467-023-36329-y 2023

  16. [16]

    and Zhang, L

    Jia, W., Wang, H., Chen, M., Lu, D., Lin, L., Car, R., Weinan, E. and Zhang, L. 2020. Pushing the Limit of Molecular Dynamics with Ab Initio Accuracy to 100 Million Atoms with Machine Learning. SC20: International Conference for High Performance Computing, Networking, Storage and Analysis (Atlanta, GA, USA, Nov. 2020), 1– 14

    2020

  17. [17]

    and Batzner, S

    Kozinsky, B., Musaelian, A., Johansson, A. and Batzner, S. 2023. Scaling the Leading Accuracy of Deep Equivariant Models to Biomolecular Simulations of Realistic Size. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (Denver CO USA, Nov. 2023), 1– 12

    2023

  18. [18]

    and Oleynik, I.I

    Nguyen-Cong, K., Willman, J.T., Moore, S.G., Belonoshko, A.B., Gayatri, R., Weinberg, E., Wood, M.A., Thompson, A.P. and Oleynik, I.I. 2021. Billion atom molecular dynamics simulations of carbon at extreme conditions and experimental time and length scales. Proceedings of the International Conference for High Performance Computing, Networking, Storage and...

    2021

  19. [19]

    Liang, T. et al. 2025. NEP89: Universal neuroevolution potential for inorganic and organic materials across 89 elements. (Apr. 2025). https://doi.org/10.48550/arXiv.2504.21286

    work page doi:10.48550/arxiv.2504.21286 2025

  20. [20]

    and Fan, Z

    Dong, H., Shi, Y., Ying, P., Xu, K., Liang, T., Wang, Y., Zeng, Z., Wu, X., Zhou, W., Xiong, S., Chen, S. and Fan, Z. 2024. Molecular dynamics simulations of heat transport using machine-learned potentials: A mini-review and tutorial on GPUMD with neuroevolution potentials. Journal of Applied Physics. 135, 16 (Apr. 2024), 161101. https://doi.org/10.1063/5.0200833

    work page doi:10.1063/5.0200833 2024

  21. [21]

    and Fan, Z

    Ying, P., Qian, C., Zhao, R., Wang, Y., Xu, K., Ding, F., Chen, S. and Fan, Z. 2025. Advances in modeling complex materials: The rise of neuroevolution potentials. Chemical Physics Reviews. 6, 1 (Mar. 2025), 011310. https://doi.org/10.1063/5.0259061

    work page doi:10.1063/5.0259061 2025

  22. [22]

    and Schmidhuber, J

    Schaul, T., Glasmachers, T. and Schmidhuber, J. 2011. High dimensions and heavy tails for natural evolution strategies. Proceedings of the 13th annual conference on Genetic and evolutionary computation (Dublin Ireland, July 2011), 845–852

    2011

  23. [23]

    Xu, K. et al. 2025. GPUMD 4.0: A high‐performance molecular dynamics package for versatile materials simulations with machine‐ learned potentials. Materials Genome Engineering Advances. (Aug. 2025), e70028. https://doi.org/10.1002/mgea.70028

    work page doi:10.1002/mgea.70028 2025

  24. [24]

    Song, K. et al. 2024. General-purpose machine-learned potential for 16 elemental metals and their alloys. Nature Communications. 15, 1 (Nov. 2024), 10208. https://doi.org/10.1038/s41467-024-54554-x

    work page doi:10.1038/s41467-024-54554-x 2024

  25. [25]

    and Jia, W

    Guo, Z., Lu, D., Yan, Y., Hu, S., Liu, R., Tan, G., Sun, N., Jiang, W., Liu, L., Chen, Y., Zhang, L., Chen, M., Wang, H. and Jia, W. 2022. Extending the limit of molecular dynamics with ab initio accuracy to 10 billion atoms. Proceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (Seoul Republic of Korea, Apr. 202...

    2022

  26. [26]

    and Jia, W

    Wang, X., Meng, X., Guo, Z., Li, M., Liu, L., Li, M., Xiao, Q., Zhao, T., Sun, N., Tan, G. and Jia, W. 2025. 29-Billion Atoms Molecular Dynamics Simulation With Ab Initio Accuracy on 35 Million Cores of New Sunway Supercomputer. IEEE Transactions on Computers. 74, 5 (May 2025), 1634–1648. https://doi.org/10.1109/TC.2025.3540646

    work page doi:10.1109/tc.2025.3540646 2025

  27. [27]

    and Su, Y

    Liu, J., Byggmä star, J., Fan, Z., Qian, P. and Su, Y. 2023. Large-scale machine-learning molecular dynamics simulation of primary radiation damage in tungsten. Physical Review B. 108, 5 (Aug. 2023), 054312. https://doi.org/10.1103/PhysRevB.108.054312

    work page doi:10.1103/physrevb.108.054312 2023

  28. [28]

    The interface is still the device

    2012. The interface is still the device. Nature Materials. 11, 2 (Feb. 2012), 91–91. https://doi.org/10.1038/nmat3244

    work page doi:10.1038/nmat3244 2012

  29. [29]

    and Wallace, R.M

    Kim, S.Y., Sun, Z., Roy, J., Wang, X., Chen, Z., Appenzeller, J. and Wallace, R.M. 2024. Fundamental Understanding of Interface Chemistry and Electrical Contact Properties of Bi and MoS2. ACS Applied Materials & Interfaces. 16, 40 (Oct. 2024), 54790–54798. https://doi.org/10.1021/acsami.4c10082

    work page doi:10.1021/acsami.4c10082 2024

  30. [30]

    and Vanderzande, D.J.M

    Oosterbaan, W.D., Bolsé e, J., Wang, L., Vrindts, V., Lutsen, L.J., Lemaur, V., Beljonne, D., McNeill, C.R., Thomsen, L., Manca, J.V. and Vanderzande, D.J.M. 2014. On the Relation between Morphology and FET Mobility of Poly(3‐alkylthiophene)s at the Polymer/SiO2 and Polymer/Air Interface. Advanced Functional Materials. 24, 14 (Apr. 2014), 1994–2004. https...

    work page doi:10.1002/adfm.201303298 2014

  31. [31]

    Xie, J. et al. 2024. Low Resistance Contact to P-Type Monolayer WSe2. Nano Letters. 24, 20 (May 2024), 5937–5943. https://doi.org/10.1021/acs.nanolett.3c04195

    work page doi:10.1021/acs.nanolett.3c04195 2024

  32. [32]

    and Lee, W.B

    Kim, D.H., Kwak, S.J., Jeong, J.H., Yoo, S., Nam, S.K., Kim, Y. and Lee, W.B. 2021. Molecular Dynamics Simulation of Silicon Dioxide Etching by Hydrogen Fluoride Using the Reactive Force Field. ACS Omega. 6, 24 (June 2021), 16009–16015. https://doi.org/10.1021/acsomega.1c01824

    work page doi:10.1021/acsomega.1c01824 2021

  33. [33]

    and Tersoff, J

    Perebeinos, V. and Tersoff, J. 2014. Carbon Nanotube Deformation and Collapse under Metal Contacts. Nano Letters. 14, 8 (Aug. 2014), 4376–4380. https://doi.org/10.1021/nl5012646

    work page doi:10.1021/nl5012646 2014