pith. machine review for the scientific record. sign in

arxiv: 2604.07276 · v1 · submitted 2026-04-08 · 💻 cs.DC · cs.AI· cs.LG

Recognition: unknown

Making Room for AI: Multi-GPU Molecular Dynamics with Deep Potentials in GROMACS

Luca Pennati , Andong Hu , Ivy Peng , Lukas M\"ullender , Stefano Markidis
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:02 UTC · model grok-4.3

classification 💻 cs.DC cs.AIcs.LG
keywords molecular dynamicsdeep potentialsGROMACSDeePMDmulti-GPUscalingneural network potentialsprotein simulations
0
0 comments X

The pith

GROMACS now supports production-scale molecular dynamics with deep neural network potentials on multi-GPU systems.

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

The paper integrates the DeePMD-kit framework into GROMACS by extending the NNPot interface with a DeePMD backend and adding a decoupled domain decomposition layer. Inference runs concurrently across processes using two MPI collectives per step to broadcast coordinates and aggregate forces. Benchmarks on a 15,668-atom protein system with a trained DPA-1 model show strong-scaling efficiency of 66 percent at 16 devices and weak-scaling efficiency of 80 percent at 16 devices, with over 90 percent of time spent in inference. A reader would care because this makes simulations with near-quantum accuracy feasible at the speed and scale of classical molecular dynamics software.

Core claim

The authors add a DeePMD backend to the GROMACS NNPot interface and introduce a domain decomposition layer that decouples inference from the main simulation loop. Two MPI collectives handle coordinate broadcast and force redistribution each step, allowing concurrent GPU-accelerated inference on all ranks. They train a 1.6-million-parameter DPA-1 model on solvated protein fragments, validate it on small systems, and benchmark scaling on up to 32 A100 and MI250x GPUs, concluding that production MD with near ab initio fidelity is feasible at scale in GROMACS.

What carries the argument

DeePMD backend for the GROMACS NNPot interface with a decoupled domain decomposition layer that uses two MPI collectives per step to exchange coordinates and forces while running inference concurrently on all processes.

If this is right

  • Production molecular dynamics of solvated proteins with near ab initio accuracy becomes practical on existing GROMACS installations using 16 to 32 GPUs.
  • Strong scaling reaches 66 percent efficiency at 16 devices and weak scaling reaches 80 percent at 16 devices for 15,000-atom systems.
  • More than 90 percent of wall time is spent in DeePMD inference while MPI collectives contribute less than 10 percent.
  • Irreducible ghost-atom costs set by the cutoff radius and load imbalance across ranks are the main scaling limits.

Where Pith is reading between the lines

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

  • The same decoupled inference layer could be reused to plug in other neural-network potentials without re-engineering the GROMACS core.
  • For systems much larger than 15,000 atoms the ghost-atom overhead may require shorter cutoffs or hybrid classical-AI potential schemes to maintain efficiency.
  • Low MPI overhead suggests the approach would transfer to other distributed MD codes facing similar inference integration needs.

Load-bearing premise

The trained DPA-1 model provides forces accurate enough for the target protein systems and the combined inference plus MPI overhead stays low enough for long production simulations.

What would settle it

A direct comparison of long protein trajectories generated by the integrated GROMACS-DeePMD code against reference ab initio or experimental data that shows systematic deviations in structure or dynamics beyond acceptable error thresholds for the application.

Figures

Figures reproduced from arXiv: 2604.07276 by Andong Hu, Ivy Peng, Luca Pennati, Lukas M\"ullender, Stefano Markidis.

Figure 1
Figure 1. Figure 1: Structure of the 1HCI protein (15,668 atoms) obtained after 500 time [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: General deep model architecture. Zi denotes the atom type, Ri atom positions, Di the atom descriptor, and ei the atom energy. There exist four major DP model classes, characterized by different descriptor architectures. Deep Potential - Smooth Edition (DP-SE) [15] is the first DP model developed. The descriptor is built by combining a local environment matrix Ri , describing neighbor geometry in invariant … view at source ↗
Figure 3
Figure 3. Figure 3: DP-SE (a), DPA-1 (b), DPA-2 (c), and DPA-3 (d) descriptor architectures. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Neighbor list in case of domain decomposition. Atoms ’A’ and ’B’ [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: GROMACS MD simulation main loop. The conceptual step order is: [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: DeePMD-kit integration in the GROMACS MD engine in the case [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of the force RMSE during training of the DPA-1 model. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison between the protein gyration radii about the three [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Memory footprint and performance overhead of a GROMACS [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The PyTorch inference task requires a high amount [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Strong scaling test on NVIDIA A100 and AMD MI250x GPUs for [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Weak scaling test on NVIDIA A100 and AMD MI250x GPUs for [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Trace of one MD simulation step obtained with the [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

GROMACS is a de-facto standard for classical Molecular Dynamics (MD). The rise of AI-driven interatomic potentials that pursue near-quantum accuracy at MD throughput now poses a significant challenge: embedding neural-network inference into multi-GPU simulations retaining high-performance. In this work, we integrate the MLIP framework DeePMD-kit into GROMACS, enabling domain-decomposed, GPU-accelerated inference across multi-node systems. We extend the GROMACS NNPot interface with a DeePMD backend, and we introduce a domain decomposition layer decoupled from the main simulation. The inference is executed concurrently on all processes, with two MPI collectives used each step to broadcast coordinates and to aggregate and redistribute forces. We train an in-house DPA-1 model (1.6 M parameters) on a dataset of solvated protein fragments. We validate the implementation on a small protein system, then we benchmark the GROMACS-DeePMD integration with a 15,668 atom protein on NVIDIA A100 and AMD MI250x GPUs up to 32 devices. Strong-scaling efficiency reaches 66% at 16 devices and 40% at 32; weak-scaling efficiency is 80% to 16 devices and reaches 48% (MI250x) and 40% (A100) at 32 devices. Profiling with the ROCm System profiler shows that >90% of the wall time is spent in DeePMD inference, while MPI collectives contribute <10%, primarily since they act as a global synchronization point. The principal bottlenecks are the irreducible ghost-atom cost set by the cutoff radius, confirmed by a simple throughput model, and load imbalance across ranks. These results demonstrate that production MD with near ab initio fidelity is feasible at scale in GROMACS.

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

1 major / 2 minor

Summary. The paper describes the integration of DeePMD-kit into GROMACS via an extended NNPot interface and a decoupled domain-decomposition layer, using two MPI collectives per step for coordinate broadcast and force aggregation/redistribution. An in-house 1.6 M parameter DPA-1 model is trained on solvated protein fragments; the implementation is validated on a small protein and then benchmarked on a 15,668-atom protein system using up to 32 NVIDIA A100 and AMD MI250x GPUs. Reported results include strong-scaling efficiencies of 66 % at 16 devices and 40 % at 32 devices, weak-scaling efficiencies of 40–48 % at 32 devices, profiling showing >90 % of wall time in DeePMD inference, and a simple throughput model identifying ghost-atom cutoff costs and load imbalance as principal bottlenecks. The work concludes that production MD with near ab initio fidelity is feasible at scale in GROMACS.

Significance. If the accuracy of the DPA-1 model is demonstrated, the integration would enable large-scale, high-fidelity molecular dynamics in a widely used production MD package, lowering the barrier for near-quantum-accuracy simulations of biomolecular systems. The concrete scaling numbers, ROCm profiling data, and simple throughput model provide practical, reproducible guidance for similar NN-potential integrations; these elements ground the performance claims and constitute a clear strength of the manuscript.

major comments (1)
  1. [Abstract and validation description] Abstract and validation description: the central claim that the integration 'demonstrates that production MD with near ab initio fidelity is feasible at scale' is not supported by any quantitative accuracy metrics for the trained DPA-1 model. No energy or force RMSE values, test-set statistics, or direct comparisons to DFT references are reported for the solvated protein fragments, the small validation protein, or the 15,668-atom benchmark system. This omission is load-bearing because the 'near ab initio fidelity' half of the feasibility claim rests entirely on unshown model accuracy rather than on the reported throughput results.
minor comments (2)
  1. [Profiling and throughput model] The simple throughput model is referenced but not shown or derived in sufficient detail to allow independent verification of the ghost-atom cost analysis.
  2. [Benchmarking results] Scaling efficiencies are reported as single-point values without error bars, number of repeated runs, or discussion of run-to-run variability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough review and for highlighting this important issue with the support for our central claim. We respond point-by-point below.

read point-by-point responses

  1. Referee: [Abstract and validation description] Abstract and validation description: the central claim that the integration 'demonstrates that production MD with near ab initio fidelity is feasible at scale' is not supported by any quantitative accuracy metrics for the trained DPA-1 model. No energy or force RMSE values, test-set statistics, or direct comparisons to DFT references are reported for the solvated protein fragments, the small validation protein, or the 15,668-atom benchmark system. This omission is load-bearing because the 'near ab initio fidelity' half of the feasibility claim rests entirely on unshown model accuracy rather than on the reported throughput results.

    Authors: We agree with the referee that the manuscript contains no quantitative accuracy metrics (energy/force RMSE, test-set statistics, or DFT comparisons) for the in-house DPA-1 model on any of the systems mentioned. The paper's core contribution is the software integration (extended NNPot interface, decoupled domain-decomposition layer, and two-MPI-collective communication pattern) together with the multi-GPU scaling results and throughput model. Model training details are provided only to describe the benchmark workload; the 'near ab initio fidelity' phrasing in the abstract and conclusions is therefore not backed by data shown here. We will revise the abstract, introduction, and conclusions to qualify the claim, stating that the integration enables production-scale MD at the fidelity of the trained deep potential (whose accuracy rests on its DFT training data). We will also add a brief clarifying sentence in the methods section. These changes will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: implementation paper with direct hardware benchmarks and no derivation chain

full rationale

The manuscript is an engineering report on integrating DeePMD-kit into GROMACS via an extended NNPot interface, introducing a decoupled domain-decomposition layer, and executing concurrent inference with two MPI collectives. It reports training a 1.6 M parameter DPA-1 model on solvated protein fragments, validation on a small system, and measured scaling efficiencies (strong scaling 66 % at 16 devices, 40 % at 32; weak scaling 40–48 % at 32 devices) plus profiling (>90 % time in inference) on A100 and MI250x hardware. No equations, fitted parameters renamed as predictions, self-citations forming load-bearing uniqueness arguments, or ansatzes smuggled via prior work appear. All performance numbers are direct wall-time measurements from GPU runs, not reductions to the paper’s own inputs. The assertion of “near ab initio fidelity” rests on the external deep-potential framework rather than any internal derivation that collapses by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the DeePMD model can substitute for classical potentials in GROMACS without invalidating the MD trajectory and that the reported scaling holds for production workloads. The 1.6M-parameter model is trained on external data.

free parameters (1)
  • DPA-1 model parameters
    1.6 million parameters in the in-house trained DPA-1 model fitted to a dataset of solvated protein fragments.
axioms (1)
  • domain assumption Standard molecular dynamics assumptions remain valid when replacing classical potentials with the DeePMD neural network potential.
    Invoked when stating that near ab initio fidelity MD is now feasible in GROMACS.

pith-pipeline@v0.9.0 · 5647 in / 1399 out tokens · 43349 ms · 2026-05-10T17:02:55.668131+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

49 extracted references · 21 canonical work pages

  1. [1]

    GROMACS 4.5: a high-throughput and highly parallel open source molecular simulation toolkit,

    S. Pronk, S. P ´all, R. Schulz, P. Larsson, P. Bjelkmar, R. Apostolov, M. R. Shirts, J. C. Smith, P. M. Kasson, D. van der Spoel, B. Hess, and E. Lindahl, “GROMACS 4.5: a high-throughput and highly parallel open source molecular simulation toolkit,” Bioinformatics, vol. 29, no. 7, pp. 845–854, 02 2013. [Online]. Available: https://doi.org/10.1093/bioinfor...

    work page doi:10.1093/bioinformatics/btt055 2013

  2. [2]

    GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers,

    M. J. Abraham, T. Murtola, R. Schulz, S. P ´all, J. C. Smith, B. Hess, and E. Lindahl, “GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers,”Soft- wareX, vol. 1-2, pp. 19–25, 2015

    2015

  3. [3]

    Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS,

    S. P ´all, A. Zhmurov, P. Bauer, M. Abraham, M. Lundborg, A. Gray, B. Hess, and E. Lindahl, “Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS,”The Journal of Chemical Physics, vol. 153, no. 13, p. 134110, 10 2020. [Online]. Available: https://doi.org/10.1063/5.0018516

    work page doi:10.1063/5.0018516 2020

  4. [4]

    A Second Generation Force Field for the Simulation of Proteins, Nucleic Acids, and Organic Molecules,

    W. D. Cornell, P. Cieplak, C. I. Bayly, I. R. Gould, K. M. Merz, D. M. Ferguson, D. C. Spellmeyer, T. Fox, J. W. Caldwell, and P. A. Kollman, “A Second Generation Force Field for the Simulation of Proteins, Nucleic Acids, and Organic Molecules,”Journal of the American Chemical Society, vol. 117, no. 19, pp. 5179–5197, 1995. [Online]. Available: https://do...

    work page doi:10.1021/ja00124a002 1995

  5. [5]

    All-Atom Empirical Potential for Molecular Modeling and Dynamics Studies of Proteins,

    A. D. M. Jr., D. Bashford, M. Bellott, R. L. D. Jr., J. D. Evanseck, M. J. Field, S. Fischer, J. Gao, H. Guo, S. Ha, D. Joseph-McCarthy, L. Kuchnir, K. Kuczera, F. T. K. Lau, C. Mattos, S. Michnick, T. Ngo, D. T. Nguyen, B. Prodhom, W. E. Reiher, B. Roux, M. Schlenkrich, J. C. Smith, R. Stote, J. Straub, M. Watanabe, J. Wi ´orkiewicz-Kuczera, D. Yin, and ...

    work page doi:10.1021/jp973084f 1998

  6. [6]

    Deep dive into machine learning density functional theory for materials science and chemistry,

    L. Fiedler, K. Shah, M. Bussmann, and A. Cangi, “Deep dive into machine learning density functional theory for materials science and chemistry,”Phys. Rev. Mater., vol. 6, p. 040301, Apr 2022

    2022

  7. [7]

    A practical guide to machine learning interatomic potentials – Status and future,

    R. Jacobs, D. Morgan, S. Attarian, J. Meng, C. Shen, Z. Wu, C. Y . Xie, J. H. Yang, N. Artrith, B. Blaiszik, G. Ceder, K. Choudhary, G. Csanyi, E. D. Cubuk, B. Deng, R. Drautz, X. Fu, J. Godwin, V . Honavar, O. Isayev, A. Johansson, B. Kozinsky, S. Martiniani, S. P. Ong, I. Poltavsky, K. Schmidt, S. Takamoto, A. P. Thompson, J. Westermayr, and B. M. Wood,...

    2025

  8. [8]

    DeePMD-kit v3: A Multiple-Backend Framework for Machine Learning Potentials,

    J. Zeng, D. Zhang, A. Peng, X. Zhang, S. He, Y . Wang, X. Liu, H. Bi, Y . Li, C. Cai, C. Zhang, Y . Du, J.-X. Zhu, P. Mo, Z. Huang, Q. Zeng, S. Shi, X. Qin, Z. Yu, C. Luo, Y . Ding, Y .-P. Liu, R. Shi, Z. Wang, S. L. Bore, J. Chang, Z. Deng, Z. Ding, S. Han, W. Jiang, G. Ke, Z. Liu, D. Lu, K. Muraoka, H. Oliaei, A. K. Singh, H. Que, W. Xu, Z. Xu, Y .-B. Z...

    work page doi:10.1021/acs.jctc.5c00340 2025

  9. [9]

    M. P. Allen and D. J. Tildesley,Computer Simulation of Liquids. Oxford University Press, 06 2017. [Online]. Available: https: //doi.org/10.1093/oso/9780198803195.001.0001

    work page doi:10.1093/oso/9780198803195.001.0001 2017

  10. [10]

    Zhou and B

    K. Zhou and B. Liu,Molecular Dynamics Simulation: Fundamentals and Applications, 1st ed. Academic Press, 2022

    2022

  11. [11]

    Ab initio molecular dy- namics: Concepts, recent developments, and future trends,

    R. Iftimie, P. Minary, and M. E. Tuckerman, “Ab initio molecular dy- namics: Concepts, recent developments, and future trends,”Proceedings of the National Academy of Sciences, vol. 102, no. 19, pp. 6654–6659, 2005

    2005

  12. [12]

    Towards electronic structure-based ab-initio molecular dynamics simulations with hundreds of millions of atoms,

    R. Schade, T. Kenter, H. Elgabarty, M. Lass, O. Sch ¨utt, A. Lazzaro, H. Pabst, S. Mohr, J. Hutter, T. D. K ¨uhne, and C. Plessl, “Towards electronic structure-based ab-initio molecular dynamics simulations with hundreds of millions of atoms,”Parallel Computing, vol. 111, p. 102920, 2022

    2022

  13. [13]

    Machine Learn- ing for Molecular Simulation,

    F. No ´e, A. Tkatchenko, K.-R. M¨uller, and C. Clementi, “Machine Learn- ing for Molecular Simulation,”Annual Review of Physical Chemistry, vol. 71, no. V olume 71, 2020, pp. 361–390, 2020

    2020

  14. [14]

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

    H. Wang, L. Zhang, J. Han, and W. E, “DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics,”Computer Physics Communications, vol. 228, pp. 178–184, 2018. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0010465518300882

    2018

  15. [15]

    End-to- end symmetry preserving inter-atomic potential energy model for finite and extended systems,

    L. Zhang, J. Han, H. Wang, W. A. Saidi, R. Car, and E. Weinan, “End-to- end symmetry preserving inter-atomic potential energy model for finite and extended systems,” p. 4441–4451, 2018

    2018

  16. [16]

    Pretraining of attention-based deep learning potential model for molecular simulation,

    D. Zhang, H. Bi, F.-Z. Dai, W. Jiang, X. Liu, L. Zhang, and H. Wang, “Pretraining of attention-based deep learning potential model for molecular simulation,”npj Computational Materials, vol. 10, no. 1, p. 94, 2024. [Online]. Available: https://doi.org/10.1038/ s41524-024-01278-7

    2024

  17. [17]

    DPA-2: a large atomic model as a multi-task learner,

    D. Zhang, X. Liu, X. Zhang, C. Zhang, C. Cai, H. Bi, Y . Du, X. Qin, A. Peng, J. Huang, B. Li, Y . Shan, J. Zeng, Y . Zhang, S. Liu, Y . Li, J. Chang, X. Wang, S. Zhou, J. Liu, X. Luo, Z. Wang, W. Jiang, J. Wu, Y . Yang, J. Yang, M. Yang, F.-Q. Gong, L. Zhang, M. Shi, F.-Z. Dai, D. M. York, S. Liu, T. Zhu, Z. Zhong, J. Lv, J. Cheng, W. Jia, M. Chen, G. Ke...

    work page doi:10.1038/s41524-024-01493-2 2024

  18. [18]

    Zhang et al., A Graph Neural Network for the Era of Large Atomistic Models

    D. Zhang, A. Peng, C. Cai, W. Li, Y . Zhou, J. Zeng, M. Guo, C. Zhang, B. Li, H. Jiang, T. Zhu, W. Jia, L. Zhang, and H. Wang, “A Graph Neural Network for the Era of Large Atomistic Models,” 2025. [Online]. Available: https://arxiv.org/abs/2506.01686

    work page arXiv 2025

  19. [19]

    A flexible algorithm for calculating pair interac- tions on SIMD architectures,

    S. P ´all and B. Hess, “A flexible algorithm for calculating pair interac- tions on SIMD architectures,”Computer Physics Communications, vol. 184, no. 12, pp. 2641–2650, 2013

    2013

  20. [20]

    General Hartree-Fock program.Computer Physics Com- munications43, 355–365 (1987)

    S. Liem, D. Brown, and J. Clarke, “Molecular dynamics simulations on distributed memory machines,”Computer Physics Communications, vol. 67, no. 2, pp. 261–267, 1991. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/001046559190021C

    work page arXiv 1991

  21. [21]

    The midpoint method for parallelization of particle simulations,

    K. J. Bowers, R. O. Dror, and D. E. Shaw, “The midpoint method for parallelization of particle simulations,”The Journal of Chemical Physics, vol. 124, no. 18, p. 184109, 05 2006. [Online]. Available: https://doi.org/10.1063/1.2191489

    work page doi:10.1063/1.2191489 2006

  22. [22]

    GROMACS Reference Manual,

    GROMACS Development Team, “GROMACS Reference Manual,” https://manual.gromacs.org/current/reference-manual/index.html, 2025

    2025

  23. [23]

    Redesign- ing GROMACS Halo Exchange: Improving Strong Scaling with GPU- initiated NVSHMEM,

    M. Doijade, A. Alekseenko, A. Brown, A. Gray, and S. P ´all, “Redesign- ing GROMACS Halo Exchange: Improving Strong Scaling with GPU- initiated NVSHMEM,”arXiv preprint arXiv:2509.21527, 2025

    work page arXiv 2025

  24. [24]

    Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials,

    A. Thompson, L. Swiler, C. Trott, S. Foiles, and G. Tucker, “Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials,”Journal of Computational Physics, vol. 285, pp. 316–330, 2015. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0021999114008353

    2015

  25. [25]

    SIMPLE-NN: An efficient package for training and executing neural-network interatomic potentials,

    K. Lee, D. Yoo, W. Jeong, and S. Han, “SIMPLE-NN: An efficient package for training and executing neural-network interatomic potentials,”Computer Physics Communications, vol. 242, pp. 95– 103, 2019. [Online]. Available: https://www.sciencedirect.com/science/ article/pii/S0010465519301298

    2019

  26. [26]

    An implementation of artificial neural-network potentials for atomistic materials simulations: Performance for TiO2,

    N. Artrith and A. Urban, “An implementation of artificial neural-network potentials for atomistic materials simulations: Performance for TiO2,” Computational Materials Science, vol. 114, pp. 135–150, 2016

    2016

  27. [27]

    TorchANI: A Free and Open Source PyTorch-Based Deep Learning Implementation of the ANI Neural Network Potentials,

    X. Gao, F. Ramezanghorbani, O. Isayev, J. S. Smith, and A. E. Roitberg, “TorchANI: A Free and Open Source PyTorch-Based Deep Learning Implementation of the ANI Neural Network Potentials,”Journal of Chemical Information and Modeling, vol. 60, no. 7, pp. 3408–3415, Jul 2020

    2020

  28. [28]

    TorchMD: A deep learning framework for molecular simulations,

    S. Doerr, M. Majewski, A. P ´erez, A. Kramer, C. Clementi, F. Noe, T. Giorgino, and G. De Fabritiis, “TorchMD: A deep learning framework for molecular simulations,”Journal of chemical theory and computation, vol. 17, no. 4, pp. 2355–2363, 2021

    2021

  29. [29]

    TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations,

    R. P. Pel ´aez, G. Simeon, R. Galvelis, A. Mirarchi, P. Eastman, S. Doerr, P. Th ¨olke, T. E. Markland, and G. D. Fabritiis, “TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations,”Journal of Chemical Theory and Computation, vol. 20, no. 10, pp. 4076–4087,

  30. [30]

    Available: https://doi.org/10.1021/acs.jctc.4c00253

    [Online]. Available: https://doi.org/10.1021/acs.jctc.4c00253

    work page doi:10.1021/acs.jctc.4c00253

  31. [31]

    Andreas Bender, Nadine Schneider, Marwin Segler, W

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, “E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials,” Nature Communications, vol. 13, no. 1, p. 2453, May 2022. [Online]. Available: https://doi.org/10.1038/s41467-022-29939-5

    work page doi:10.1038/s41467-022-29939-5 2022

  32. [32]

    Learning local equivariant representations for large-scale atomistic dynamics,

    A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, and B. Kozinsky, “Learning local equivariant representations for large-scale atomistic dynamics,”Nature Communications, vol. 14, no. 1, p. 579, Feb 2023. [Online]. Available: https://doi.org/10.1038/s41467-023-36329-y

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

  33. [33]

    Materials Learning Algorithms (MALA): Scalable machine learning for electronic structure calculations in large-scale atomistic simulations,

    A. Cangi, L. Fiedler, B. Brzoza, K. Shah, T. J. Callow, D. Kotik, S. Schmerler, M. C. Barry, J. M. Goff, A. Rohskopf, D. J. V ogel, N. Modine, A. P. Thompson, and S. Rajamanickam, “Materials Learning Algorithms (MALA): Scalable machine learning for electronic structure calculations in large-scale atomistic simulations,”Computer Physics Communications, vol...

    2025

  34. [34]

    CHARMM at 45: Enhancements in Accessibility, Functionality, and Speed,

    W. Hwang, S. L. Austin, A. Blondel, E. D. Boittier, S. Boresch, M. Buck, J. Buckner, A. Caflisch, H. Chang, X. Cheng, Y . K. Choi, J. Chu, M. F. Crowley, Q. Cui, A. Damjanovic, Y . Deng, M. Devereux, X. Ding, M. F. Feig, J. Gao, D. R. Glowacki, J. E. G. II, M. B. Hamaneh, E. D. Harder, R. L. Hayes, J. Huang, Y . Huang, P. S. Hudson, W. Im, S. M. Islam, W....

    work page doi:10.1021/acs.jpcb.4c04100 2024

  35. [35]

    Development of Range-Corrected Deep Learning Potentials for Fast, Accurate Quantum Mechanical/Molecular Mechanical Simulations of Chemical Reactions in Solution,

    J. Zeng, T. J. Giese, S ¸. Ekesan, and D. M. York, “Development of Range-Corrected Deep Learning Potentials for Fast, Accurate Quantum Mechanical/Molecular Mechanical Simulations of Chemical Reactions in Solution,”Journal of Chemical Theory and Computation, vol. 17, no. 11, pp. 6993–7009, Nov 2021

    2021

  36. [36]

    AmberTools,

    D. A. Case, H. M. Aktulga, K. Belfon, D. S. Cerutti, G. A. Cisneros, V . W. D. Cruzeiro, N. Forouzesh, T. J. Giese, A. W. G ¨otz, H. Gohlke, S. Izadi, K. Kasavajhala, M. C. Kaymak, E. King, T. Kurtzman, T. Lee, P. Li, J. Liu, T. Luchko, R. Luo, M. Manathunga, M. R. Machado, H. M. Nguyen, K. A. O’Hearn, A. V . Onufriev, F. Pan, S. Pantano, R. Qi, A. Rahnam...

    work page doi:10.1021/acs.jcim.3c01153 2023

  37. [37]

    Implementation and Validation of an OpenMM Plugin for the Deep Potential Representation of Potential Energy,

    Y . Ding and J. Huang, “Implementation and Validation of an OpenMM Plugin for the Deep Potential Representation of Potential Energy,” International Journal of Molecular Sciences, vol. 25, no. 3, 2024

    2024

  38. [38]

    LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales,

    A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M. Brown, P. S. Crozier, P. J. in ’t Veld, A. Kohlmeyer, S. G. Moore, T. D. Nguyen, R. Shan, M. J. Stevens, J. Tranchida, C. Trott, and S. J. Plimpton, “LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales,”Computer Physics Commu...

    2022

  39. [39]

    AENET–LAMMPS and AENET–TINKER: interfaces for accurate and efficient molecular dynamics simulations with machine learning poten- tials,

    M. S. Chen, T. Morawietz, H. Mori, T. E. Markland, and N. Artrith, “AENET–LAMMPS and AENET–TINKER: interfaces for accurate and efficient molecular dynamics simulations with machine learning poten- tials,”The Journal of Chemical Physics, vol. 155, no. 7, 2021

    2021

  40. [40]

    MLMOD: Machine Learning Methods for Data-Driven Modeling in LAMMPS,

    P. Atzberger, “MLMOD: Machine Learning Methods for Data-Driven Modeling in LAMMPS,”Journal of Open Source Software, vol. 8, p. 5620, 09 2023

    2023

  41. [41]

    chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations,

    P. Fuchs, W. Chen, S. Thaler, and J. Zavadlav, “chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations,”arXiv preprint arXiv:2506.04055, 2025

    work page arXiv 2025

  42. [42]

    Billion atom molecular dynamics simulations of carbon at extreme conditions and experimental time and length scales,

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

    work page doi:10.1145/3458817.3487400 2021

  43. [43]

    Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning,

    W. Jia, H. Wang, M. Chen, D. Lu, L. Lin, R. Car, W. E, and L. Zhang, “Pushing the limit of molecular dynamics with ab initio accuracy to 100 million atoms with machine learning,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, ser. SC ’20. IEEE Press, 2020

    2020

  44. [44]

    Extending the limit of molecular dynamics with ab initio accuracy to 10 billion atoms,

    Z. Guo, D. Lu, Y . Yan, S. Hu, R. Liu, G. Tan, N. Sun, W. Jiang, L. Liu, Y . Chen, L. Zhang, M. Chen, H. Wang, and W. Jia, “Extending the limit of molecular dynamics with ab initio accuracy to 10 billion atoms,” inProceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, ser. PPoPP ’22. New York, NY , USA: Associati...

    work page doi:10.1145/3503221.3508425 2022

  45. [45]

    InProceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis(Atlanta, GA, USA)(SC ’24)

    J. Li, B. Li, Z. Guo, M. Li, E. Li, L. Liu, G. Yuan, Z. Wang, G. Tan, and W. Jia, “Scaling Molecular Dynamics with ab initio Accuracy to 149 Nanoseconds per Day,” inProceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis, ser. SC ’24. IEEE Press, 2024. [Online]. Available: https://doi.org/10.1109/SC414...

    work page doi:10.1109/sc41406.2024.00036 2024

  46. [46]

    MACE: higher order equivariant message passing neural networks for fast and accurate force fields,

    I. Batatia, D. P. Kov ´acs, G. N. C. Simm, C. Ortner, and G. Cs ´anyi, “MACE: higher order equivariant message passing neural networks for fast and accurate force fields,” inProceedings of the 36th International Conference on Neural Information Processing Systems, ser. NIPS ’22. Red Hook, NY , USA: Curran Associates Inc., 2022

    2022

  47. [47]

    PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges,

    O. T. Unke and M. Meuwly, “PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments, and Partial Charges,”Journal of Chemical Theory and Computation, vol. 15, no. 6, pp. 3678–3693, Jun

  48. [48]

    Available: https://doi.org/10.1021/acs.jctc.9b00181

    [Online]. Available: https://doi.org/10.1021/acs.jctc.9b00181

    work page doi:10.1021/acs.jctc.9b00181

  49. [49]

    Enabling AI Deep Potentials for Ab Initio-quality Molecular Dynamics Simulations in GROMACS,

    A. Hu, L. Pennati, S. Markidis, and I. Peng, “Enabling AI Deep Potentials for Ab Initio-quality Molecular Dynamics Simulations in GROMACS,” in2026 34th Euromicro International Conference on Parallel, Distributed, and Network-Based Processing (PDP). Los Alamitos, CA, USA: IEEE Computer Society, Mar. 2026

    2026