Pith. sign in

REVIEW 3 major objections 7 minor 74 references

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

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper presents chemtrain-deploy, a model-agnostic pipeline that deploys JAX-defined machine-learning potentials into LAMMPS and runs molecular dynamics on millions of atoms across multiple GPUs.

desk verdict Real engineering value in the scaling data, but the 'any JAX model' and 'state-of-the-art' claims outrun the evidence; deserves review after tightening. read the letter →

arxiv 2506.04055 v1 pith:MJJ5WQZS submitted 2025-06-04 physics.comp-ph cs.LGphysics.chem-ph

Paul Fuchs , Weilong Chen , Stephan Thaler , Julija Zavadlav This is my paper
classification physics.comp-phcs.LGphysics.chem-ph
keywords machinelearningpotentialsmoleculardynamicsLAMMPSJAXdomaindecompositionmulti-GPUscalinggraphneuralnetworksXLA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to remove the architecture-specific barriers that keep machine-learning potentials (MLPs) from being used in large-scale molecular dynamics. It introduces chemtrain-deploy, a pipeline that takes any JAX-defined semi-local potential — one whose total energy is a sum of per-atom energies depending on a local graph environment — and deploys it inside LAMMPS, a widely used molecular dynamics engine. The framework combines an MLIR-based model export, an XLA/PJRT runtime that compiles the model for the GPU backend, and a domain-decomposition scheme that guarantees each processor's subgraph contains all information needed to compute exact forces. The authors validate the pipeline with three neural-network architectures, MACE, Allegro, and PaiNN, on water–vapor, aluminum, and solvated Chignolin systems, reporting close-to-ideal strong and weak scaling up to 64 GPUs and throughput high enough for million-atom simulations. If the claim holds, practitioners can run current and future graph-neural-network potentials on parallel hardware without rewriting molecular dynamics code.

What carries the argument

The load-bearing machinery is the export-and-runtime pipeline plus the cutoff inequality. A semi-local potential is defined by a per-atom energy $U_i(\mathbf{r}_i, \mathcal{A}_i)$ where $\mathcal{A}_i$ contains all atoms within $T$ message-passing hops of atom $i$; the model is exported with its graph specification and units into a self-contained format via MLIR, then compiled at runtime by XLA through the PJRT device interface. The plugin takes LAMMPS's spatial decomposition and neighbor lists, prunes edges longer than the model cutoff, and pads all tensors to fixed shapes so the compiled program does not need recompilation until buffer capacities are exceeded. The identity $R_{\text{comm}} \ge T R$ is what makes the decomposition exact: it ensures the copied halo atoms contain every $T$-th-order neighbor needed to compute the energy and force of each local atom.

What would settle it

Take a JAX-defined semi-local potential that uses dynamic control flow or data-dependent shapes not present in Allegro, MACE, or PaiNN, export it with chemtrain-deploy, and compare its forces against the untouched JAX implementation on identical coordinates: any deviation beyond floating-point tolerance would refute the claim that the pipeline supports any JAX-defined semi-local potential.

Watch

Extended reading notes

Core claim

The central discovery is that model-agnostic, multi-GPU deployment of semi-local machine-learning potentials is achievable by combining three existing pieces: a self-contained model export, a just-in-time compiler stack, and the spatial decomposition machinery of a mature molecular dynamics code. The paper shows that if the communication cutoff satisfies $R_{\text{comm}} \ge T R$, where $R$ is the model cutoff and $T$ the number of message-passing steps, each processor's local subgraph contains the complete $T$-th-order neighbor environment of every local atom. Per-atom energies can then be summed locally, and forces can be reconciled by a single reduction of partial forces, so no intermediate communication is needed during model evaluation. With graph pruning and fixed-shape buffering to avoid costly XLA recompilations, the resulting plugin runs Allegro, MACE, and PaiNN on systems of roughly one million atoms with near-ideal strong scaling and close-to-ideal weak scaling on up to 64 GPUs.

Load-bearing premise

The pipeline is only demonstrated on three JAX architectures, so the claim that it handles any JAX-defined semi-local potential rests on the assumption that MLIR export and XLA/PJRT compilation preserve both correctness and efficiency for arbitrary JAX models, including those with dynamic control flow or unusual graph operations.

Editorial extensions

If this is right

  • Any JAX-defined semi-local potential — including Behler–Parrinello-style networks, NequIP, and DimeNet++ — can in principle be deployed in LAMMPS without writing new pair-style code, provided it exports cleanly.
  • For message-passing models with $T$ layers and cutoff $R$, the communication halo must extend to $T R$; this directly sets the parallel overhead and predicts strong-scaling efficiency via the ratio of copied atoms to local atoms.
  • Multi-GPU execution removes the single-GPU memory ceiling that currently caps graph-neural-network potential simulations below roughly half a million atoms, opening one-million-atom systems and beyond to machine-learning-potential molecular dynamics.
  • Different machine-learning potential architectures can be benchmarked in the same LAMMPS simulation environment with identical integration settings and reference datasets, making architecture comparisons more direct.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If semantic preservation of MLIR export holds beyond the three tested architectures, chemtrain-deploy could serve as a neutral benchmark harness that settles architecture-comparison debates under identical molecular dynamics conditions.
  • The scaling formula suggests a design principle the paper states only implicitly: for fixed per-GPU domain sizes, increasing the model cutoff or the number of message-passing layers shrinks parallel efficiency, so strictly local models will win at large scale unless the extra body order buys enough accuracy.
  • A stress-test extension would deploy a JAX potential with dynamic control flow or irregular graph updates, such as a model that prunes edges inside the network, and check whether the fixed-shape buffering and collective recompilation logic preserves both correctness and throughput.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper presents chemtrain-deploy, a framework that exports JAX-defined semi-local machine learning potentials to a self-contained representation and evaluates them inside LAMMPS through MLIR export and XLA/PJRT compilation. The framework uses LAMMPS domain decomposition with a communication cutoff R_comm >= T R so that each domain contains the T-th neighbor environments of local atoms; forces are assembled by summing partial forces from atom copies, avoiding intermediate communication. Strong and weak scaling up to 64 GPUs are reported for MACE, Allegro, and PaiNN on water, aluminum, and solvated Chignolin systems of up to about 1.2 million atoms, with comparisons to JAX, M.D. and a classical MEAM potential. A parameter-free analytic scaling model (Eq. 3) is introduced and found to be consistent with the measured data. The paper claims that chemtrain-deploy supports any JAX-defined semi-local potential and achieves state-of-the-art efficiency.

Significance. If the claims hold, chemtrain-deploy is a useful engineering contribution: it provides an open-source path from JAX-trained models to multi-GPU LAMMPS simulations and offers a parameter-free analytic estimate of parallel efficiency (Eq. 3) that is not fitted to the measurements. The three-architecture validation is a reasonable first demonstration, and the reported scaling curves are internally consistent with the analytic model. However, the two headline claims are not established by the current evidence. The model-agnostic claim needs either a scoped statement of the supported JAX language subset plus numerical equivalence checks, or explicit limitations; the state-of-the-art efficiency claim needs head-to-head comparisons with other multi-GPU MLP deployment frameworks. The paper will be of interest to the MLP and molecular dynamics community once these points are addressed.

major comments (3)
  1. [Section 2 (ModelExport, Fig. 1b/e); Discussion] The abstract and Discussion state that chemtrain-deploy supports 'any JAX-defined semi-local potential,' but this is not established. The export path relies on MLIR lowering and PJRT/XLA compilation, yet the paper gives no specification of the supported JAX language subset (e.g., whether Python control flow, host callbacks, jax.custom_vjp, or dynamic shapes are permitted) and no test of models beyond the three demonstrated JIT-compatible architectures. Moreover, no end-to-end numerical equivalence check between the trained JAX model and the deployed LAMMPS pair style is reported: Table 1 compares model predictions with reference DFT data, not with the JAX evaluator. Because semantic preservation of the exported program is load-bearing for the model-agnostic claim, this needs either a scoped claim or explicit validation on a broader set of models.
  2. [Abstract; Section 2 (Memory requirements) and Fig. 4] The abstract's 'state-of-the-art efficiency' claim is not supported by the reported comparisons. In Section 2 ('Memory requirements'), the only software baseline is JAX, M.D. on a single GPU, and in Fig. 4 the only multi-GPU baseline is classical MEAM; no head-to-head timing against contemporary multi-GPU MLP deployment frameworks such as DeePMD-kit v3, SevenNet, or Orb is included. The supplementary comparison with Allegro-LAMMPS covers two systems and is not quantified in the main text. I recommend either adding such benchmarks or rephrasing the claim to 'competitive efficiency with existing single-GPU JAX approaches and good scaling behavior.'
  3. [Section 2 (Scaling to million-atom systems); Supplementary Figure 2] The text reports that Allegro strong scaling 'often exceed[s] the anticipated ideal parallel efficiency,' i.e., parallel efficiency greater than 1 in Eq. 4. This is surprising because Eq. 3 predicts efficiency below 1 for any finite T R, and no measurement protocol is given (number of repeats, warm-up treatment, whether the 250 production steps are averaged). Please document the timing methodology and either explain the superlinear regime or present the data as close-to-ideal rather than 'exceeding ideal.'
minor comments (7)
  1. [Throughout] There are numerous typographical errors ('exsisting', 'mutliple', 'Traning', 'alumnium', 'molecular dyamics', 'avnailable'); please proofread the manuscript.
  2. [Section 4.3] The Chignolin simulation is described with a 0.5 ps time step; this is almost certainly a typo for 0.5 fs, and should be corrected because it affects reproducibility.
  3. [References] Several references are incomplete, with missing year, volume, or page information (e.g., refs. 1, 2, 6, 7, 13, 14); please standardize the bibliography.
  4. [Figure 2] The legend states that ideal and approximate (Eq. 3) strong scaling are both shown as dashed lines in black and model colors, respectively; using different line styles would make the comparison easier to read.
  5. [Section 2] The main text cites 'Supplementary Note 1' for the MACE accuracy/scalability tradeoff; please ensure this supplementary note is included and its content is summarized in the main text.
  6. [Figure 1] The 'T=2' annotation appears before the message-passing order T is defined in the text; consider defining it in the figure caption for standalone readability.
  7. [Table 2] The caption says measurements are for a single A100 GPU, but the text states only that JAX, M.D. simulations were performed on A100; please clarify that the chemtrain-deploy maximum sizes in Table 2 were also measured on the same hardware.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Eq. 3 is an a priori, parameter-free scaling model compared against (not fit to) measurements; accuracy is benchmarked against external literature; the unverified model-agnostic claim is an evidence gap, not a by-construction reduction.

full rationale

The paper's central deliverables are (i) a software path exporting JAX-defined semi-local potentials into LAMMPS via MLIR/XLA/PJRT, (ii) a distributed force/energy evaluation scheme with R_comm >= T R, (iii) measured strong/weak scaling on three architectures, and (iv) accuracy benchmarks. None of these reduces to its own inputs by construction. The scaling model, Eq. 3, is derived a priori from the geometric assumption that per-domain cost is proportional to (L + 2TR)^d; it contains no fitted parameters (T, R, L, P, d are all known from the model and setup), and the paper uses it as a reference curve against which measurements are compared, even noting where data 'slightly outperform' it (Results, Scaling to million-atom systems) — a genuine, falsifiable comparison rather than a fit relabeled as prediction. Equations (1), (2), and (5) are the standard semi-local energy decomposition and the exact force sum rule; Eq. 2 follows by differentiating Eq. 1, and the R_comm >= T R condition follows from the bound that any T-th neighbor lies within distance T R — no quantity is defined in terms of the claimed result. Accuracy validation is anchored outside the paper: trained models are compared to external datasets (H2O-PBE0TS, ANI-AL, SPICE) and to literature benchmarks (NequIP, DeepMD, MEAM, TorchMD-NET, Table 1). The only self-referential element is the use of the authors' own chemtrain platform (ref. [51]) for training, which is normal platform extension rather than load-bearing circularity, because the deployment, accuracy, and scaling results are all measured directly and compared to external references. The broad claim that 'any JAX-defined semi-local potential' is supported rests on demonstrations of only three architectures; if MLIR/XLA export silently fails for some JAX models, the claim would be over-broad — an evidence gap that belongs under correctness risk, not circularity. No passage defines model-agnostic support in terms of the three demonstrated models, no unique theorem is imported from the authors' prior work, and no fitted parameter is presented as a prediction. Verdict: no significant circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central scaling claim relies on standard force-decomposition identities and domain-decomposition assumptions, but no numbers are fitted to data. The main untested premise is generic model export; the scalability formula is a stated approximation rather than a fit.

free parameters (1)
  • Buffer capacity multiplier = 1.1
    Set in the example LAMMPS input (Fig. 1) to control recompilation frequency; hand-tuned, affects warm-up cost but not the central scaling model.
assumptions (5)
  • domain assumption The total energy decomposes into a sum of per-atom energies U_i that depend only on T-th order neighbors within radius T R (Eqs. 1 and 5).
    This is the definition of semi-local potentials that the framework claims to support; if a model has longer-range or global interactions, the domain decomposition argument does not apply.
  • domain assumption LAMMPS spatial decomposition with ghost cutoff R_comm >= T R yields subgraphs that contain the complete environment of every local atom (Distributed potential computation).
    Correctness of energies and forces depends on this bound; for neighbor lists pruned to model cutoff it can be violated if model cutoffs vary per species or are not monotone in distance.
  • domain assumption Runtime scales linearly with the number of atoms in each domain including ghost atoms (Parallelization cost, Eq. 3).
    The predicted speedup S in Eq. 3 is based on this linear-cost assumption and on uniform rectangular domains; real systems with density inhomogeneity may deviate, though measured scaling is consistent.
  • domain assumption MLIR export preserves the model graph and XLA/PJRT can compile it for the target backend (ModelExport and library sections).
    Model-agnostic support is only as strong as this pipeline; only three GNN architectures are demonstrated.
  • standard math Force on atom i is the sum of partial forces from all atoms in its T-th order environment (Eq. 2).
    This is the chain-rule decomposition of the energy expression; standard calculus.

how reviews work

0 comments
Cite this review

Pith. "Pith review of chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations." pith.science (2026). https://pith.science/paper/MJJ5WQZS

@misc{pith2026250604055,
  author       = {Pith},
  title        = {Pith review of: chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJJ5WQZS}},
  note         = {Machine review of arXiv:2506.04055}
}
read the original abstract

Machine learning potentials (MLPs) have advanced rapidly and show great promise to transform molecular dynamics (MD) simulations. However, most existing software tools are tied to specific MLP architectures, lack integration with standard MD packages, or are not parallelizable across GPUs. To address these challenges, we present chemtrain-deploy, a framework that enables model-agnostic deployment of MLPs in LAMMPS. chemtrain-deploy supports any JAX-defined semi-local potential, allowing users to exploit the functionality of LAMMPS and perform large-scale MLP-based MD simulations on multiple GPUs. It achieves state-of-the-art efficiency and scales to systems containing millions of atoms. We validate its performance and scalability using graph neural network architectures, including MACE, Allegro, and PaiNN, applied to a variety of systems, such as liquid-vapor interfaces, crystalline materials, and solvated peptides. Our results highlight the practical utility of chemtrain-deploy for real-world, high-performance simulations and provide guidance for MLP architecture selection and future design.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 26 canonical work pages

  1. [1]

    Unke, O.T., Chmiela, S., Sauceda, H.E., Gastegger, M., Poltavsky, I., Sch¨ utt, K.T., Tkatchenko, A., M¨ uller, K.-R.: Machine Learning Force Fields121(16), 10142–10186 https://doi.org/10.1021/acs.chemrev.0c01111

    doi:10.1021/acs.chemrev.0c01111

  2. [2]

    No´ e, F., Tkatchenko, A., M¨ uller, K.-R., Clementi, C.: Machine Learn- ing for Molecular Simulation71, 361–390 https://doi.org/10.1146/ annurev-physchem-042018-052331

    work page

  3. [3]

    The Journal of chemical physics145(17) (2016)

    Behler, J.: Perspective: Machine learning potentials for atomistic simulations. The Journal of chemical physics145(17) (2016)

    work page 2016

  4. [4]

    Merchant, A., Batzner, S., Schoenholz, S.S., Aykol, M., Cheon, G., Cubuk, E.D.: Scaling deep learning for materials discovery624(7990), 80–85 https://doi.org/ 10.1038/s41586-023-06735-9

    doi:10.1038/s41586-023-06735-9

  5. [5]

    org/10.1073/pnas.0500193102

    Iftimie, R., Minary, P., Tuckerman, M.E.: Ab initio molecular dynamics: Con- cepts, recent developments, and future trends102(19), 6654–6659 https://doi. org/10.1073/pnas.0500193102

    doi:10.1073/pnas.0500193102

  6. [6]

    Cornell, W.D., Cieplak, P., Bayly, C.I., Gould, I.R., Merz, K.M., Ferguson, D.M., Spellmeyer, D.C., Fox, T., Caldwell, J.W., Kollman, P.A.: A Second Generation 18 Force Field for the Simulation of Proteins, Nucleic Acids, and Organic Molecules 117(19), 5179–5197 https://doi.org/10.1021/ja00124a002

    doi:10.1021/ja00124a002

  7. [7]

    Nikitin, A.M., Milchevskiy, Y.V., Lyubartsev, A.P.: A new AMBER-compatible force field parameter set for alkanes20(3), 2143 https://doi.org/10.1007/ s00894-014-2143-6

    work page

  8. [8]

    Marrink, S.J., Vries, A.H., Mark, A.E.: Coarse Grained Model for Semiquantita- tive Lipid Simulations108(2), 750–760 https://doi.org/10.1021/jp036508g

    doi:10.1021/jp036508g

Show all 74 references
  1. [9]

    Behler, J.: Atom-centered symmetry functions for constructing high-dimensional neural network potentials134(7), 074106 https://doi.org/10.1063/1.3553717

    doi

  2. [10]

    Chemical Science 8(4), 3192–3203 (2017) https://doi.org/10.1039/C6SC05720A

    Smith, J.S., Isayev, O., Roitberg, A.E.: ANI-1: An extensible neural network potential with DFT accuracy at force field computational cost. Chemical Science 8(4), 3192–3203 (2017) https://doi.org/10.1039/C6SC05720A

    2017 doi

  3. [11]

    In: Precup, D., Teh, Y.W

    Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1263–...

    2017

  4. [12]

    Physical Review B99(1), 014104 (2019) https://doi.org/10.1103/ PhysRevB.99.014104

    Drautz, R.: Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B99(1), 014104 (2019) https://doi.org/10.1103/ PhysRevB.99.014104

    2019

  5. [13]

    Zhang, L., Han, J., Wang, H., Car, R., E, W.: Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics120(14), 143001 https://doi.org/10.1103/PhysRevLett.120.143001

    doi

  6. [14]

    https://doi.org/10.48550/ arXiv.2504.06231

    Rhodes, B., Vandenhaute, S., ˇSimkus, V., Gin, J., Godwin, J., Duignan, T., Neumann, M.: Orb-v3: Atomistic Simulation at Scale. https://doi.org/10.48550/ arXiv.2504.06231 . http://arxiv.org/abs/2504.06231

    doi arXiv

  7. [15]

    https://doi.org/10.48550/arXiv

    Zhang, D., Peng, A., Cai, C., Li, W., Zhou, Y., Zeng, J., Guo, M., Zhang, C., Li, B., Jiang, H., Zhu, T., Jia, W., Zhang, L., Wang, H.: Graph Neural Network Model for the Era of Large Atomistic Models. https://doi.org/10.48550/arXiv. 2506.01686 . http://arxiv.org/abs/2506.01686

    doi

  8. [16]

    Kov´ acs, D.P., Moore, J.H., Browning, N.J., Batatia, I., Horton, J.T., Pu, Y., Kapil, V., Witt, W.C., Magd˘ au, I.-B., Cole, D.J., Cs´ anyi, G.: MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules 147(21), 17598–17611 https://doi.org/10.1021...

    doi

  9. [17]

    Nature Machine Intelligence7(1), 56–67 (2025) 19 https://doi.org/10.1038/s42256-024-00956-x

    Batatia, I., Batzner, S., Kov´ acs, D.P., Musaelian, A., Simm, G.N.C., Drautz, R., Ortner, C., Kozinsky, B., Cs´ anyi, G.: The design space of E(3)-equivariant atom- centred interatomic potentials. Nature Machine Intelligence7(1), 56–67 (2025) 19 https://doi.org/10.1038/s42256...

    2025 doi

  10. [18]

    Nature Communications14(1), 579 (2023) https://doi.org/10.1038/ s41467-023-36329-y

    Musaelian, A., Batzner, S., Johansson, A., Sun, L., Owen, C.J., Kornbluth, M., Kozinsky, B.: Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications14(1), 579 (2023) https://doi.org/10.1038/ s41467-023-36329-y

    2023

  11. [19]

    Nature Communications 13(1), 1–11 (2022) https://doi.org/10.1038/s41467-022-29939-5

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt, T.E., Kozinsky, B.: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications 13(1), 1–11 (2022) https://doi.org/10.103...

    2022 doi

  12. [20]

    The Journal of Chemical Physics148(24), 241722 (2018) https://doi.org/10.1063/1.5019779 arXiv:1712.06113 [cond-mat, physics:physics]

    Sch¨ utt, K.T., Sauceda, H.E., Kindermans, P.-J., Tkatchenko, A., M¨ uller, K.-R.: SchNet - a deep learning architecture for molecules and materials. The Journal of Chemical Physics148(24), 241722 (2018) https://doi.org/10.1063/1.5019779 arXiv:1712.06113 [cond-mat, physics:physics]

    2018 arXiv

  13. [21]

    https://doi.org/10

    Sch¨ utt, K.T., Unke, O.T., Gastegger, M.: Equivariant Message Passing for the Prediction of Tensorial Properties and Molecular Spectra. https://doi.org/10. 48550/arXiv.2102.03150 . http://arxiv.org/abs/2102.03150

    doi arXiv

  14. [22]

    https://doi.org/10.48550/arXiv.2206.07697

    Batatia, I., Kov´ acs, D.P., Simm, G.N.C., Ortner, C., Cs´ anyi, G.: MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. https://doi.org/10.48550/arXiv.2206.07697 . http://arxiv.org/abs/2206. 07697

    doi arXiv

  15. [23]

    Sch¨ utt, K.T., Kessel, P., Gastegger, M., Nicoli, K.A., Tkatchenko, A., M¨ uller, K.- R.: SchNetPack: A Deep Learning Toolbox For Atomistic Systems15(1), 448–455 https://doi.org/10.1021/acs.jctc.8b00908

    doi

  16. [24]

    jcim.0c00451

    Gao, X., Ramezanghorbani, F., Isayev, O., Smith, J.S., Roitberg, A.E.: TorchANI: A Free and Open Source PyTorch-Based Deep Learning Implementation of the ANI Neural Network Potentials60(7), 3408–3415 https://doi.org/10.1021/acs. jcim.0c00451

    doi

  17. [25]

    Doerr, S., Majewski, M., P´ erez, A., Kr¨ amer, A., Clementi, C., Noe, F., Giorgino, T., De Fabritiis, G.: TorchMD: A Deep Learning Framework for Molecular Simulations17(4), 2355–2363 https://doi.org/10.1021/acs.jctc.0c01343

    doi

  18. [26]

    https://doi.org/10.48550/ arXiv.2506.02023

    Han, K., Deng, B., Farimani, A.B., Ceder, G.: DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials. https://doi.org/10.48550/ arXiv.2506.02023 . http://arxiv.org/abs/2506.02023

    doi

  19. [27]

    Anderson, J.A., Lorenz, C.D., Travesset, A.: General purpose molecular dynamics simulations fully implemented on graphics processing units227(10), 5342–5359 https://doi.org/10.1016/j.jcp.2008.01.047 20

    2008 doi

  20. [28]

    Pronk, S., P´ all, S., Schulz, R., Larsson, P., Bjelkmar, P., Apostolov, R., Shirts, M.R., Smith, J.C., Kasson, P.M., Van Der Spoel, D.: GROMACS 4.5: A high- throughput and highly parallel open source molecular simulation toolkit29(7), 845–854

  21. [29]

    Computer Physics Communications271, 108171 (2022) https: //doi.org/10.1016/j.cpc.2021.108171

    Thompson, A.P., Aktulga, H.M., Berger, R., Bolintineanu, D.S., Brown, W.M., Crozier, P.S., in ’t Veld, P.J., Kohlmeyer, A., Moore, S.G., Nguyen, T.D., Shan, R., Stevens, M.J., Tranchida, J., Trott, C., Plimpton, S.J.: LAMMPS - a flexible simulation tool for particle-based mate...

    2022

  22. [30]

    Zhang, Y., Wang, H., Chen, W., Zeng, J., Zhang, L., Wang, H., E, W.: DP-GEN: A concurrent learning platform for the generation of reliable deep learning based potential energy models253, 107206 https://doi.org/10.1016/j.cpc.2020.107206

  23. [31]

    1038/s41467-021-21376-0 2003.04934

    Smith, J.S., Nebgen, B., Mathew, N., Chen, J., Lubbers, N., Burakovsky, L., Tretiak, S., Nam, H.A., Germann, T., Fensin, S., Barros, K.: Automated discovery of a robust interatomic potential for aluminum12(1), 1257 https://doi.org/10. 1038/s41467-021-21376-0 2003.04934

    arXiv

  24. [32]

    https://doi

    Levine, D.S., Shuaibi, M., Spotte-Smith, E.W.C., Taylor, M.G., Hasyim, M.R., Michel, K., Batatia, I., Cs´ anyi, G., Dzamba, M., Eastman, P., Frey, N.C., Fu, X., Gharakhanyan, V., Krishnapriyan, A.S., Rackers, J.A., Raja, S., Rizvi, A., Rosen, A.S., Ulissi, Z., Vargas, S., Zitn...

    2025 doi

  25. [33]

    1038/s41467-021-27241-4

    Thaler, S., Zavadlav, J.: Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting12(1), 6884 https://doi.org/10. 1038/s41467-021-27241-4

  26. [34]

    Thaler, S., Stupp, M., Zavadlav, J.: Deep coarse-grained potentials via relative entropy minimization157(24), 244103 https://doi.org/10.1063/5.0124538

    doi

  27. [35]

    1063/5.0235189

    R¨ ocken, S., Burnet, A.F., Zavadlav, J.: Predicting solvation free energies with an implicit solvent machine learning potential161(23), 234101 https://doi.org/10. 1063/5.0235189

  28. [36]

    Cheng, B.: Latent Ewald summation for machine learning of long-range interac- tions11(1), 1–8 https://doi.org/10.1038/s41524-025-01577-7

    doi

  29. [37]

    https://doi

    Fuchs, P., Sanocki, M., Zavadlav, J.: Learning Non-Local Molecular Interactions Via Equivariant Local Representations and Charge Equilibration. https://doi. org/10.48550/arXiv.2501.19179 . http://arxiv.org/abs/2501.19179

    doi arXiv

  30. [38]

    https://doi.org/10.48550/arXiv.2303

    Kosmala, A., Gasteiger, J., Gao, N., G¨ unnemann, S.: Ewald-Based Long-Range 21 Message Passing for Molecular Graphs. https://doi.org/10.48550/arXiv.2303. 04791 . http://arxiv.org/abs/2303.04791

    doi arXiv

  31. [39]

    https://doi.org/10.48550/arXiv.2502.13797

    Caruso, A., Venturin, J., Giambagli, L., Rolando, E., No´ e, F., Clementi, C.: Extending the RANGE of Graph Neural Networks: Relaying Attention Nodes for Global Encoding. https://doi.org/10.48550/arXiv.2502.13797 . http://arxiv.org/ abs/2502.13797

    doi arXiv

  32. [40]

    https://doi

    Frank, J.T., Chmiela, S., M¨ uller, K.-R., Unke, O.T.: Euclidean Fast Attention: Machine Learning Global Atomic Representations at Linear Cost. https://doi. org/10.48550/arXiv.2412.08541 . http://arxiv.org/abs/2412.08541

    doi

  33. [41]

    arXiv preprint arXiv:2210.07237 (2022)

    Fu, X., Wu, Z., Wang, W., Xie, T., Keten, S., Gomez-Bombarelli, R., Jaakkola, T.: Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations. arXiv preprint arXiv:2210.07237 (2022)

    2022 arXiv

  34. [42]

    https://doi.org/10.48550/arXiv.2502.12147

    Fu, X., Wood, B.M., Barroso-Luque, L., Levine, D.S., Gao, M., Dzamba, M., Zit- nick, C.L.: Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction. https://doi.org/10.48550/arXiv.2502.12147 . http://arxiv. org/abs/2502.12147

    doi arXiv

  35. [43]

    https://doi.org/10.48550/arXiv.2408

    P´ ota, B., Ahlawat, P., Cs´ anyi, G., Simoncelli, M.: Thermal Conductivity Predic- tions with Foundation Atomistic Models. https://doi.org/10.48550/arXiv.2408. 00755 . http://arxiv.org/abs/2408.00755

    doi arXiv

  36. [44]

    https://doi.org/10

    Loew, A., Sun, D., Wang, H.-C., Botti, S., Marques, M.A.L.: Universal Machine Learning Interatomic Potentials Are Ready For Phonons. https://doi.org/10. 48550/arXiv.2412.16551 . http://arxiv.org/abs/2412.16551

    doi arXiv

  37. [45]

    https://doi.org/10.48550/arXiv.2402.13984

    Raja, S., Amin, I., Pedregosa, F., Krishnapriyan, A.S.: Stability-Aware Train- ing of Machine Learning Force Fields with Differentiable Boltzmann Estimators. https://doi.org/10.48550/arXiv.2402.13984 . http://arxiv.org/abs/2402.13984

    doi arXiv

  38. [46]

    Park, Y., Kim, J., Hwang, S., Han, S.: Scalable Parallel Algorithm for Graph Neu- ral Network Interatomic Potentials in Molecular Dynamics Simulations20(11), 4857–4868 https://doi.org/10.1021/acs.jctc.4c00190

    doi

  39. [47]

    Rohskopf, A., Sievers, C., Lubbers, N., Cusentino, M., Goff, J., Janssen, J., McCarthy, M., Zapiain, D.M.O., Nikolov, S., Sargsyan, K., Sema, D., Sikorski, E., Williams, L., Thompson, A., Wood, M.: FitSNAP: Atomistic machine learning with LAMMPS8(84), 5118 https://doi.org/10.2...

    doi

  40. [48]

    https://doi.org/10.48550/arXiv.2310.03121

    Eastman, P., Galvelis, R., Pel´ aez, R.P., Abreu, C.R.A., Farr, S.E., Gallicchio, E., Gorenko, A., Henry, M.M., Hu, F., Huang, J., Kr¨ amer, A., Michel, J., Mitchell, J.A., Pande, V.S., Rodrigues, J.P., Rodriguez-Guerra, J., Simmonett, A.C., Singh, S., Swails, J., Turner, P., ...

    doi arXiv

  41. [49]

    Zeng, J., Zhang, D., Peng, A., Zhang, X., He, S., Wang, Y., Liu, X., Bi, H., Li, Y., Cai, C., Zhang, C., Du, Y., Zhu, J.-X., Mo, P., Huang, Z., Zeng, Q., Shi, S., Qin, X., Yu, Z., Luo, C., Ding, Y., Liu, Y.-P., Shi, R., Wang, Z., Bore, S.L., Chang, J., Deng, Z., Ding, Z., Han,...

  42. [50]

    Zeng, J., Giese, T.J., Zhang, D., Wang, H., York, D.M.: DeePMD-GNN: A DeePMD-kit Plugin for External Graph Neural Network Potentials65(7), 3154–3160 https://doi.org/10.1021/acs.jcim.4c02441

    doi

  43. [51]

    Computer Physics Communications310, 109512 (2025) https://doi.org/10.1016/j.cpc.2025.109512

    Fuchs, P., Thaler, S., R¨ ocken, S., Zavadlav, J.: Chemtrain: Learning deep potential models via automatic differentiation and statistical physics. Computer Physics Communications310, 109512 (2025) https://doi.org/10.1016/j.cpc.2025.109512

    2025

  44. [52]

    Schoenholz, S., Cubuk, E.D.: JAX MD: A Framework for Differentiable Physics 33, 11428–11441

  45. [53]

    In: 2021 IEEE/ACM Interna- tional Symposium on Code Generation and Optimization (CGO), pp

    Lattner, C., Amini, M., Bondhugula, U., Cohen, A., Davis, A., Pienaar, J., Riddle, R., Shpeisman, T., Vasilache, N., Zinenko, O.: MLIR: Scaling compiler infrastructure for domain specific computation. In: 2021 IEEE/ACM Interna- tional Symposium on Code Generation and Optimizat...

    2021

  46. [54]

    https://github.com/openxla/xla

    OpenXLA. https://github.com/openxla/xla

  47. [55]

    https://openxla.org/xla/pjrt

    PJRT - Uniform Device API. https://openxla.org/xla/pjrt

  48. [56]

    https://doi.org/10.5281/zenodo.15009305

    Developers, T.: TensorFlow. https://doi.org/10.5281/zenodo.15009305 . https:// doi.org/10.5281/zenodo.15009305

    doi

  49. [57]

    Eastman, P., Behara, P.K., Dotson, D.L., Galvelis, R., Herr, J.E., Horton, J.T., Mao, Y., Chodera, J.D., Pritchard, B.P., Wang, Y., De Fabritiis, G., Markland, T.E.: SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials10(1), 11 https://...

    doi

  50. [58]

    Lee, B.-J., Shim, J.-H., Baskes, M.I.: Semiempirical atomic potentials for the fcc metals cu, ag, au, ni, pd, pt, al, and pb based on first and second nearest- neighbor modified embedded atom method. Phys. Rev. B68, 144112 (2003) https: //doi.org/10.1103/PhysRevB.68.144112

    2003 doi

  51. [59]

    Modelling and Simulation in Materials Science and Engineering27(8), 085015 (2019) https://doi.org/10.1088/ 1361-651X/ab4b36

    Mahata, A., Asle Zaeem, M.: Size effect in molecular dynamics simulation of 23 nucleation process during solidification of pure metals: Investigating modified embedded atom method interatomic potentials. Modelling and Simulation in Materials Science and Engineering27(8), 08501...

    2019

  52. [60]

    MODELLING AND SIMULATION IN MATERIALS SCIENCE AND ENGINEERING18(015012) (2010) https://doi

    Stukowski, A.: Visualization and analysis of atomistic simulation data with OVITO-the open visualization tool. MODELLING AND SIMULATION IN MATERIALS SCIENCE AND ENGINEERING18(015012) (2010) https://doi. org/10.1088/0965-0393/18/1/015012

    2010 doi

  53. [61]

    In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis

    Kozinsky, B., Musaelian, A., Johansson, A., Batzner, S.: Scaling the leading accu- racy of deep equivariant models to biomolecular simulations of realistic size. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. S...

    2023

  54. [62]

    Behler, J., Parrinello, M.: Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces98(14), 146401 https://doi.org/10.1103/ PhysRevLett.98.146401

  55. [63]

    https://doi

    Gasteiger, J., Giri, S., Margraf, J.T., G¨ unnemann, S.: Fast and Uncertainty- Aware Directional Message Passing for Non-Equilibrium Molecules. https://doi. org/10.48550/arXiv.2011.14115 . http://arxiv.org/abs/2011.14115

    2011 doi arXiv

  56. [64]

    Wang, J., Olsson, S., Wehmeyer, C., P´ erez, A., Charron, N.E., De Fabritiis, G., No´ e, F., Clementi, C.: Machine Learning of Coarse-Grained Molecular Dynamics Force Fields5(5), 755–767 https://doi.org/10.1021/acscentsci.8b00913

    doi

  57. [65]

    https://doi.org/10.26434/chemrxiv-2024-bdfr0

    Kabylda, A., Frank, J.T., Dou, S.S., Khabibrakhmanov, A., Sandonas, L.M., Unke, O.T., Chmiela, S., Muller, K.-R., Tkatchenko, A.: Molecular Simula- tions with a Pretrained Neural Network and Universal Pairwise Force Fields. https://doi.org/10.26434/chemrxiv-2024-bdfr0 . https:...

  58. [66]

    Fuchs, P., Sanocki, M., Zavadlav, J.: Learning Non-Local Molecular Interactions via Equivariant Local Representations and Charge Equilibration (2025)

    2025

  59. [67]

    The Journal of chemical physics153(4) (2020)

    Phillips, J.C., Hardy, D.J., Maia, J.D., Stone, J.E., Ribeiro, J.V., Bernardi, R.C., Buch, R., Fiorin, G., H´ enin, J., Jiang, W., et al.: Scalable molecular dynamics on cpu and gpu architectures with namd. The Journal of chemical physics153(4) (2020)

    2020

  60. [68]

    The Journal of Chemical Physics97(3), 1990–2001 (1992) https: //doi.org/10.1063/1.463137 24

    Tuckerman, M., Berne, B.J., Martyna, G.J.: Reversible multiple time scale molec- ular dynamics. The Journal of Chemical Physics97(3), 1990–2001 (1992) https: //doi.org/10.1063/1.463137 24

    1992 doi

  61. [69]

    Physical Review Letters98(14), 146401 (2007) https://doi.org/10.1103/PhysRevLett.98.146401

    Behler, J., Parrinello, M.: Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces. Physical Review Letters98(14), 146401 (2007) https://doi.org/10.1103/PhysRevLett.98.146401

    2007 doi

  62. [70]

    Chem- ical Reviews121(16), 9759–9815 (2021) https://doi.org/10.1021/acs.chemrev

    Musil, F., Grisafi, A., Bart´ ok, A.P., Ortner, C., Cs´ anyi, G., Ceriotti, M.: Physics-Inspired Structural Representations for Molecules and Materials. Chem- ical Reviews121(16), 9759–9815 (2021) https://doi.org/10.1021/acs.chemrev. 1c00021

    2021 doi

  63. [71]

    Physical Review B87(18), 184115 (2013) https://doi.org/10.1103/PhysRevB.87

    Bart´ ok, A.P., Kondor, R., Cs´ anyi, G.: On representing chemical environments. Physical Review B87(18), 184115 (2013) https://doi.org/10.1103/PhysRevB.87. 184115

    2013 doi

  64. [72]

    https://doi.org/10.48550/arXiv.2003.03123

    Gasteiger, J., Groß, J., G¨ unnemann, S.: Directional Message Passing for Molec- ular Graphs. https://doi.org/10.48550/arXiv.2003.03123 . http://arxiv.org/abs/ 2003.03123

    2003 doi arXiv

  65. [73]

    Wollschl¨ ager, T., Gao, N., Charpentier, B., Ketata, M.A., G¨ unnemann, S.: Uncertainty Estimation for Molecules: Desiderata and Methods (2023)

    2023

  66. [74]

    Sanchez-Burgos, I., Muniz, M.C., Espinosa, J.R., Panagiotopoulos, A.Z.: A Deep Potential model for liquid–vapor equilibrium and cavitation rates of water 158(18), 184504 https://doi.org/10.1063/5.0144500 25

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.