Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems

T0 review · 3 major / 5 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 A fully JAX-based library packages MACE, NequIP, and ViSNet into one training and simulation pipeline, with molecular dynamics on a single GPU fast enough to make protein-scale runs practical.

desk verdict A genuinely useful JAX MLIP library, but the speed and accuracy claims are softer than the abstract suggests; worth a proper review with targeted fixes. read the letter →

arxiv 2505.22397 v2 pith:TTYVCQSE submitted 2025-05-28 physics.chem-ph

Christoph Brunken , Olivier Peltre , Heloise Chomet , Lucien Walewski , Manus McAuliffe , Valentin Heyraud , Solal Attias , Martin Maarand
show 6 more authors
Yessine Khanfir Edan Toledo Fabio Falcioni Marie Bluntzer Silvia Acosta-Gutiérrez Jules Tilly
This is my paper
classification physics.chem-ph
keywords machinelearninginteratomicpotentialsJAXMACENequIPViSNetJAX-MDmoleculardynamicsSPICE2
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

This white paper presents mlip, a unified, entirely JAX-based library for training, fine-tuning, and simulating with machine learning interatomic potentials. The authors claim that its tight integration with the JAX-MD simulator yields state-of-the-art molecular dynamics speeds: for example, MACE-large on a 138-atom mini-protein runs at 6.3 ms per step with JAX+JAX-MD versus 44.2 ms per step with the paper's Torch+ASE route. They also introduce a modified MACE architecture that reaches accuracy close to the standard ν=3 model at roughly twice the inference speed. If these claims hold, a practitioner without a machine learning background can deploy pre-trained organics models for long, GPU-accelerated MD simulations that were previously impractical.

What carries the argument

The central mechanism is the JAX software stack: full just-in-time compilation with XLA means the entire force-field prediction and the MD integrator run as one compiled GPU program with no CPU–GPU data transfer between steps. To avoid costly recompilation when neighbor lists change, the library pads the neighbor lists and only reallocates and reruns an episode when the edge buffer overflows. The modified MACE model's key mechanism is a species-dependent gating weight applied to node features, built from the scalar output of the symmetric contraction; this increases body order at negligible extra cost, plus an edge-wise MLP that injects sender-and-receiver species information into the interaction term.

What would settle it

Benchmark the same three models using their original, optimized PyTorch implementations (the official MACE, NequIP, and ViSNet repositories) on the same H100 GPU and the same 1UAO and 1ABT systems; if the JAX+JAX-MD times are not at least as fast as the official implementations, the central speed claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single JAX-native library can deliver the full MLIP workflow—data preprocessing, model training, fine-tuning, batched inference, and molecular dynamics—while making MLIP-based MD fast enough for near-industrial use. The load-bearing evidence is Table 2, which reports wall-clock speeds on one H100 GPU: JAX+JAX-MD outperforms the paper's own Torch+ASE ports by roughly 3–7x depending on model and system. A second, independent contribution is a modified MACE model: gating the node features with scalars produced during the symmetric contraction, plus making neighbor interactions explicitly dependent on both atomic species, raises the effective body order without increasing the correlation order, giving near-ν=3 accuracy at ν=2-like speed.

Load-bearing premise

The headline speed advantage assumes the paper's own Torch+ASE ports are a fair baseline for the Torch MLIP ecosystem.

Editorial extensions

If this is right

  • If the tabulated speeds hold, a single H100 GPU can run a 1-nanosecond simulation of a ~140-atom protein system with MACE-large in about 16 minutes, and with the modified MACE in about 8 minutes.
  • The modified MACE design suggests that body order can be raised through gating rather than through a more expensive correlation order, a trick that could be ported to other equivariant message-passing architectures.
  • The library lowers the barrier for non-specialists, since pre-trained organics models can be loaded and simulated with a few lines of code, making MLIP more accessible to industry users.
  • The ability to add new model architectures and loss functions within one framework should accelerate method development and cross-model comparisons.

Reading between the lines

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

  • If the JAX+JAX-MD speed advantage survives comparison against the original authors' optimized Torch implementations, JAX-MD-style integration may become the default backend for MLIP simulation, not just for this library.
  • The species-dependent gating modification is a general architectural idea; it could plausibly be applied to NequIP or other equivariant models to improve accuracy without sacrificing inference speed, though the paper does not test this.
  • The paper itself notes that the relative speed advantage of JAX+JAX-MD should shrink on larger systems as GPU capacity saturates; extrapolating beyond the 1,205-atom benchmark is therefore speculative.
  • The emphasis on inference speed over training speed positions the library for production simulation use cases, but the paper's own warning implies validation error alone is not enough to certify physical correctness—an inference not tested here.
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 / 5 minor

Summary. This paper introduces mlip, a JAX-based library for training, evaluating, and simulating machine learning interatomic potentials (MLIPs). The library ships with three model architectures (MACE, NequIP, ViSNet), two molecular dynamics wrappers (ASE and JAX-MD), and a set of pre-trained models on a curated SPICE2 dataset. The authors report validation energy/force MAE for the pre-trained models, runtime benchmarks on two protein systems (1UAO and 1ABT) comparing JAX+JAX-MD, JAX+ASE, and Torch+ASE routes, and propose a modified MACE architecture with gated, species-dependent features that is claimed to reach accuracy close to ν=3 MACE at ν=2-like speed.

Significance. If the claims hold, the library is a valuable open-source contribution to the MLIP ecosystem: it provides a unified, JAX-based training/simulation workflow, pre-trained organic models, and a transparently documented benchmark setup. The internal comparison between JAX+JAX-MD and JAX+ASE is useful and shows clear benefits from JIT-compilation and GPU-resident simulation. The modified MACE variant is a promising speed-accuracy trade-off, provided its accuracy is quantitatively verified. The authors ship code, disclose important caveats (notably that their Torch+ASE baselines are self-ported and not representative of native implementations), and provide full hyperparameters, which are strengths. The main weaknesses are that the headline 'state-of-the-art MD simulation speeds' claim is not supported by the reported baselines, and the modified-MACE accuracy claim lacks numerical details.

major comments (3)
  1. [Model benchmarking, Table 2] The Introduction and the Practical examples section claim that the JAX-MD backend enables 'state-of-the-art MD simulation speeds,' and Table 2 is the only quantitative support for this claim. However, the Torch+ASE baselines in Table 2 are the authors' own implementations, and the text explicitly states that these 'should not be considered representative of the performance of the code developed by the original authors.' No benchmarks against native MACE, NequIP, or ViSNet repositories, nor against accelerated backends such as cuEquivariance (which the roadmap lists as future work), are provided. The observed 3-7x speedups therefore compare the JAX pipeline against unoptimized self-ported references, and they do not, by themselves, establish 'state-of-the-art' performance. The internal comparison between JAX+JAX-MD and JAX+ASE is sound and informative, but the 'state-of-the-art' qualifier is unsupported as written. Please either benchmark against native implementations or revise the claim to describe the observed speedups relative to the library's own Torch ports.
  2. [Appendix C, Eqs. (1)-(6), Figure 2] The central claim of Appendix C is that the Modified MACE model reaches accuracy 'close' to the vanilla ν=3 MACE at an inference speed similar to ν=2, but this claim is not quantitatively verifiable from the manuscript. No numerical MAE or RMSE values, error bars, or seed variance are reported for the modified model; Figure 2 is the only evidence, and the text itself warns that validation errors are 'not sufficient to attest to a model's ability to simulate correct physics.' Because the 2x speed advantage of Modified MACE is only meaningful when paired with a quantified accuracy anchor, please provide a table with per-subset energy and force MAE (and ideally RMSE) for vanilla MACE ν=2, vanilla MACE ν=3, and Modified MACE, including results from at least a few training seeds.
  3. [Dataset and pre-trained models; Figure 2] The validation comparisons in Figure 2 and the runtime benchmarks in Table 2 lack error bars, standard deviations, or seed information, and the models were 'selected from many training runs' on the same SPICE2 validation metrics used to report the results. This selection makes it difficult to assess whether reported differences (e.g., 'MACE-large, on average, outperforms MACE-medium in force RMSE') are robust or reflect selection noise. At minimum, the number of seeds used for the final models should be reported, and standard deviations should be provided for the headline accuracy and timing numbers; this is especially important for the Appendix C claim, where the accuracy comparison is a load-bearing part of the contribution.
minor comments (5)
  1. [Throughout] The spelling of ViSNet is inconsistent: the text and tables use 'ViSNet', 'VisNet', and 'Visnet' (e.g., Table 2, Table 4, and the model description). Please standardize.
  2. [Model benchmarking, Figure 2 and text] The text states that 'MACE-medium achieves lower energy RMSE than MACE-large in every subset, while MACE-large, on average, outperforms MACE-medium in force RMSE,' but Figure 2 presents only MAE. Please either add an RMSE panel or remove the RMSE-specific claims.
  3. [Related work] The sentence 'Energy conservation is ensured by the construction of the vector-value kernel function' contains a typo ('vector-value' should be 'vector-valued').
  4. [Appendix C, Eq. (1)] The notation in Eq. (1) is difficult to parse, particularly the simultaneous use of tildes and multiple subscripts on the mixing weights (e.g., 'W ˜ην Zik˜k,ην'). Please clearly define all indices and summation ranges, or move the detailed notation to a table.
  5. [Dataset and pre-trained models] The force-filter thresholds (total force norm exceeding 0.1 eV/Å or per-atom force greater than 15 eV/Å) are stated as adopted after improving benchmark performance, but no sensitivity analysis is given. A brief comment on how robust the model performance is to these thresholds would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's load-bearing claims are empirical benchmarks and implementation descriptions, not derived results; self-citations are background and limitations are candidly disclosed.

full rationale

The paper's central claims—library efficiency (Table 2), pre-trained model accuracy (Figure 2), and the Modified MACE trade-off (Appendix C)—are empirical measurements or code descriptions, not results derived from inputs. No equation is fitted to the quantity it then predicts; no uniqueness theorem is imported; no ansatz is smuggled via self-citation. The MACE implementation is based on Geiger and Batatia's JAX code and the original MACE papers [43-45], while reference [4] (which has overlapping authors) appears only as background and as one of several MACE citations; it is not load-bearing. References [62] and [63] are roadmap self-citations only. The paper explicitly disclaims the Torch+ASE baselines ('should not be considered representative of the performance of the code developed by the original authors') and the validation metrics ('not sufficient to attest to a model's ability to simulate correct physics'), so the potential weaknesses are benchmarking-validity and generalization concerns, not circular reductions. Appendix C proposes explicit new equations (Eqs. 1-6) whose accuracy claim is an empirical outcome reported in Figure 2; even though no numerical MAE values are tabulated there, the absence of supporting numbers is an evidence gap, not a circularity. Accordingly, the circularity score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper's central efficiency claim rests on disclosed engineering choices (JAX, padding, JAX-MD) and hand-chosen training and data-filtering parameters; the modified MACE adds learnable gating and species-pair weights. No new physical postulates are introduced.

free parameters (7)
  • Force filter thresholds (total force norm, per-atom force) = 0.1 eV/Å total; 15 eV/Å per atom
    Hand-chosen dataset exclusion cuts. The paper states the filter was adopted because it 'led to improved performance on key benchmarks', so the reported validation metrics are partly conditioned on these thresholds.
  • Loss weight schedule (energy, force, flip epoch) = 40 and 1000, flipped at epoch 115
    Hand-chosen training weights (Appendix A); they steer the energy/force accuracy balance visible in Figure 2.
  • MACE-large architecture choices = 2 layers, 128 channels, correlation 2, l_max 3
    Selected for MD stability over the MACE-OFF-aligned configuration, at the cost of higher energy MAE; a design choice that shapes the paper's headline accuracy numbers.
  • Training schedule constants = EMA decay 0.99; warmup 4000; transition 360000; clip 500
    Hand-chosen hyperparameters in Appendix A shared across all three architectures.
  • Gating mixing weights W and biases b (Appendix C) = not reported
    Species-wise learnable weights introduced in Eq. 1 of Appendix C; trained on SPICE2, no final values given.
  • Species embeddings s_i = dimension d = 8
    Learnable per-species embeddings introduced in Eq. 3 of Appendix C; trained on SPICE2.
  • Edge-feature MLP weights producing beta = not reported
    MLP parameters mapping species-pair features to interaction weights in Eq. 5 of Appendix C; trained on SPICE2.
assumptions (4)
  • domain assumption SPICE2 DFT labels (ωB97M-D3(BJ)/def2-TZVPPD) are adequate ground truth for the energies and forces used to train and evaluate all models.
    Invoked in 'Curated SPICE2 Dataset'; all Figure 2 MAE values are deviations from these DFT references, so accuracy claims are relative to this electronic-structure level.
  • domain assumption Completing a 1 ns MD run without divergence is a sufficient demonstration of simulation stability.
    Table 2 states 'All models included in the table achieved stable simulations' with no stability criterion (e.g., energy drift or temperature bound) defined.
  • domain assumption A 5 Å neighbor cutoff captures the interactions needed for accurate energy and force prediction in these organic and biomolecular systems.
    graph_cutoff_angstrom = 5 is set for every model in Appendix A; accuracy claims depend on this locality assumption.
  • domain assumption JIT compilation with padded neighbor lists and episode re-runs preserves the correctness of MD trajectories.
    The library avoids recompilation by padding neighbor lists and re-running overflowed episodes; the paper provides no validation that this mechanism leaves trajectories unaffected.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems." pith.science (2026). https://pith.science/paper/TTYVCQSE

@misc{pith2026250522397,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTYVCQSE}},
  note         = {Machine review of arXiv:2505.22397}
}
read the original abstract

Machine Learning Interatomic Potentials (MLIP) are a novel in silico approach for molecular property prediction, creating an alternative to disrupt the accuracy/speed trade-off of empirical force fields and density functional theory (DFT). In this white paper, we present our MLIP library which was created with two core aims: (1) provide to industry experts without machine learning background a user-friendly and computationally efficient set of tools to experiment with MLIP models, (2) provide machine learning developers a framework to develop novel approaches fully integrated with molecular dynamics tools. The library includes in this release three model architectures (MACE, NequIP, and ViSNet), and two molecular dynamics (MD) wrappers (ASE, and JAX-MD), along with a set of pre-trained organics models. The seamless integration with JAX-MD, in particular, facilitates highly efficient MD simulations, bringing MLIP models significantly closer to industrial application. The library is available on GitHub and on PyPI under the Apache license 2.0.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. NepTrain and NepTrainKit: Automated Active Learning and Visualization Toolkit for Neuroevolution Potentials

    cs.LG 2025-06 conditional novelty 6.0 of 10

    A new open-source toolkit automates active learning and dataset management for neuroevolution potentials, with a CsPbI3 case study showing comparable accuracy to hand-curated NEP models.

Reference graph

Works this paper leans on

72 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Stevens, Fabian Grünewald, P

    Jan A. Stevens, Fabian Grünewald, P. A. Marco van Tilburg, Melanie König, Benjamin R. Gilbert, Troy A. Brier, Zane R. Thornburg, Zaida Luthey-Schulten, and Siewert J. Marrink. Molecular dynamics simu- lation of an entire cell.Frontiers in Chemistry, 11, January 2023. doi: 10.3389/fchem.2023.1106495. URL https://doi.org/10.3389/fchem.2023.1106495. Section:...

    arXiv 2023

  2. [2]

    Shaw, Peter J

    David E. Shaw, Peter J. Adams, Asaph Azaria, Joseph A. Bank, Brannon Batson, Alistair Bell, Michael Bergdorf, Jhanvi Bhatt, J. Adam Butts, Timothy Correia, and et al. Anton 3: Twenty microseconds of molecular dynamics simulation before lunch. InProceedings of the International Conference for High Perfor- mance Computing, Networking, Storage and Analysis (...

    arXiv 2021

  3. [3]

    Johnson, Chris Leary, Dougal Maclaurin, ..., and Qiao Zhang

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew J. Johnson, Chris Leary, Dougal Maclaurin, ..., and Qiao Zhang. Jax: composable transformations of python+numpy programs. InProceedings of the 31st Conference on Neural Information Processing Systems (NeurIPS 2018), 2018. URLhttps://github. com/google/jax

    work page 2018

  4. [4]

    Elena, Dávid P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, Dávid P. Kovács, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Samuel M. Blau, Vlad Cărare, James P. Darby, Sandip De, Flaviano Della Pia, Volker L. Deringer, Rokas Elijošius, Zakariya El-Machachi, Fabio Falcio...

    arXiv 2024

  5. [5]

    Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G

    Daniel S. Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G. Taylor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia, Gábor Csányi, Misko Dzamba, Peter Eastman, Nathan C. Frey, Xiang Fu, Vahe Gharakhanyan, Aditi S. Krishnapriyan, Joshua A. Rackers, Sanjeev Raja, Ammar Rizvi, Andrew S. Rosen, Zachary Ulissi, Santiago Vargas, C. Lawrence Zi...

    arXiv 2025

  6. [6]

    Bartók, Mike C

    Albert P. Bartók, Mike C. Payne, Risi Kondor, and Gábor Csányi. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons.Physical Review Letters, 104(13), April

    work page

  7. [7]

    Gaussian Approximation Potentials: a brief tutorial introduction

    Albert P. Bartók and Gábor Csányi. Gaussian approximation potentials: a brief tutorial introduction, 2020. URL https://arxiv.org/abs/1502.01366

    work page Pith review arXiv 2020

  8. [8]

    Bartók, James Kermode, Noam Bernstein, and Gábor Csányi

    Albert P. Bartók, James Kermode, Noam Bernstein, and Gábor Csányi. Machine learning a general- purpose interatomic potential for silicon.Physical Review X, 8(4), December 2018. ISSN 2160-3308. doi: 10.1103/physrevx.8.041048. URL http://dx.doi.org/10.1103/PhysRevX.8.041048

    doi:10.1103/physrevx.8.041048 2018

Show all 72 references
  1. [9]

    Bartók, Risi Kondor, and Gábor Csányi

    Albert P. Bartók, Risi Kondor, and Gábor Csányi. On representing chemical environments.Physical Review B, 87(18), May 2013. ISSN 1550-235X. doi: 10.1103/physrevb.87.184115. URLhttp://dx.doi.org/ 10.1103/PhysRevB.87.184115. 14 Machine Learning Interatomic Potentials: library fo...

    2013 doi

  2. [10]

    Generalized neural-network representation of high-dimensional potential-energy surfaces.Physical Review Letters, 98(14), April 2007

    Jörg Behler and Michele Parrinello. Generalized neural-network representation of high-dimensional potential-energy surfaces.Physical Review Letters, 98(14), April 2007. ISSN 1079-7114. doi: 10.1103/ physrevlett.98.146401. URL http://dx.doi.org/10.1103/PhysRevLett.98.146401

    2007 doi

  3. [11]

    Thompson, L.P

    A.P. Thompson, L.P. Swiler, C.R. Trott, S.M. Foiles, and G.J. Tucker. Spectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials.Journal of Computational Physics, 285:316–330, March 2015. ISSN 0021-9991. doi: 10.1016/j.jcp.2014.12.018...

    2015 doi

  4. [12]

    Sauceda, Igor Poltavsky, Kristof T

    Stefan Chmiela, Alexandre Tkatchenko, Huziel E. Sauceda, Igor Poltavsky, Kristof T. Schütt, and Klaus- Robert Müller. Machine learning of accurate energy-conserving molecular force fields.Science Advances, 3(5):e1603015, 2017. doi: 10.1126/sciadv.1603015

    2017 doi

  5. [13]

    Sauceda, Klaus-Robert Müller, and Alexandre Tkatchenko

    Stefan Chmiela, Huziel E. Sauceda, Klaus-Robert Müller, and Alexandre Tkatchenko. Towards exact molecular dynamics simulations with machine-learned force fields.Nature Communications, 9(1):3887,

  6. [14]

    Sauceda, Alexandre Tkatchenko, and Klaus-Robert Müller.Accurate molecular dynamics enabled by efficient physically-constrained machine learning approaches, pages 129–154

    Stefan Chmiela, Huziel E. Sauceda, Alexandre Tkatchenko, and Klaus-Robert Müller.Accurate molecular dynamics enabled by efficient physically-constrained machine learning approaches, pages 129–154. Springer International Publishing, 2020. doi: 10.1007/978-3-030-40245-7\_7

    2020 doi

  7. [15]

    Unke, Adil Kabylda, Huziel E

    Stefan Chmiela, Valentin Vassilev-Galindo, Oliver T. Unke, Adil Kabylda, Huziel E. Sauceda, Alexandre Tkatchenko, and Klaus-Robert Müller. Accurate global machine learning force fields for molecules with hundreds of atoms.Science Advances, 9(2):eadf0873, 2023. doi: 10.1126/sci...

    2023 doi

  8. [16]

    Atomic cluster expansion for accurate and transferable interatomic potentials.Physical Review B, 99(1), January 2019

    Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials.Physical Review B, 99(1), January 2019. ISSN 2469-9969. doi: 10.1103/physrevb.99.014104. URL http: //dx.doi.org/10.1103/PhysRevB.99.014104

    2019 doi

  9. [17]

    Alexander V. Shapeev. Moment tensor potentials: A class of systematically improvable interatomic potentials. Multiscale Modeling & Simulation, 14(3):1153–1173, January 2016. ISSN 1540-3467. doi: 10.1137/15m1054183. URL http://dx.doi.org/10.1137/15M1054183

    2016 doi

  10. [18]

    J. S. Smith, O. Isayev, and A. E. Roitberg. Ani-1: an extensible neural network potential with dft accuracy at force field computational cost.Chemical Science, 8(4):3192–3203, 2017. ISSN 2041-6539. doi: 10.1039/c6sc05720a. URL http://dx.doi.org/10.1039/C6SC05720A

    2017 doi

  11. [19]

    Smith, Roman Zubatyuk, Benjamin Nebgen, Nicholas Lubbers, Kipton Barros, Adrian E

    Justin S. Smith, Roman Zubatyuk, Benjamin Nebgen, Nicholas Lubbers, Kipton Barros, Adrian E. Roitberg, Olexandr Isayev, and Sergei Tretiak. The ani-1ccx and ani-1x data sets, coupled-cluster and density functional theory properties for molecules.Scientific Data, 7(1), May 2020...

    2020 doi

  12. [20]

    Smith, Benjamin T

    Justin S. Smith, Benjamin T. Nebgen, Roman Zubatyuk, Nicholas Lubbers, Christian Devereux, Kipton Barros, Sergei Tretiak, Olexandr Isayev, and Adrian E. Roitberg. Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning.Nat...

    2019 doi

  13. [21]

    Smith, Kate K

    Christian Devereux, Justin S. Smith, Kate K. Huddleston, Kipton Barros, Roman Zubatyuk, Olexandr Isayev, and Adrian E. Roitberg. Extending the applicability of the ani deep learning molecular potential to sulfur and halogens.Journal of Chemical Theory and Computation, 16(7):41...

    2020

  14. [22]

    Comparing ani-2x, ani-1ccx neural networks, force field, and dft methods for predicting conformational potential energy of organic molecules

    Mozafar Rezaee, Saeid Ekrami, and Seyed Majid Hashemianzadeh. Comparing ani-2x, ani-1ccx neural networks, force field, and dft methods for predicting conformational potential energy of organic molecules. Scientific Reports, 14(1), May 2024. ISSN 2045-2322. doi: 10.1038/s41598-...

    2024 doi

  15. [23]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data- efficient and accurate interatomic potentials.Nature Communications, 13(1), M...

    2022 doi

  16. [24]

    Does equivariance matter at scale?,

    Johann Brehmer, Sönke Behrends, Pim de Haan, and Taco Cohen. Does equivariance matter at scale?,

  17. [25]

    K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. Müller. Schnet – a deep learning architecture for molecules and materials.The Journal of Chemical Physics, 148(24), March 2018. ISSN 1089-7690. doi: 10.1063/1.5019779. URLhttp://dx.doi.org/10.1063/1.5019779

    2018 doi

  18. [26]

    Schütt, Pan Kessel, Michael Gastegger, Kim A

    Kristof T. Schütt, Pan Kessel, Michael Gastegger, Kim A. Nicoli, Alexandre Tkatchenko, and Klaus-Robert Müller. SchNetPack: A Deep Learning Toolbox For Atomistic Systems.Journal of Chemical Theory and Computation, 15(1):448–455, 2019. doi: 10.1021/acs.jctc.8b00908. URLhttps://...

    2019 doi

  19. [27]

    Schütt, Stefaan S

    Kristof T. Schütt, Stefaan S. P. Hessmann, Niklas W. A. Gebauer, Jonas Lederer, and Michael Gastegger. SchNetPack 2.0: A neural network toolbox for atomistic machine learning.The Journal of Chemical Physics, 158(14):144801, 04 2023. ISSN 0021-9606. doi: 10.1063/5.0138367. URLh...

    2023 doi

  20. [28]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical Review Letters, 120(14), April 2018. ISSN 1079-

    2018

  21. [29]

    Anstine, Roman Zubatyuk, and Olexandr Isayev

    Dylan M. Anstine, Roman Zubatyuk, and Olexandr Isayev. Aimnet2: a neural network potential to meet your neutral, charged, organic, and elemental-organic needs.Chemical Science, 2025. ISSN 2041-6539. doi: 10.1039/d4sc08572h. URLhttp://dx.doi.org/10.1039/D4SC08572H

    2025 doi

  22. [30]

    E(n) equivariant graph neural networks,

    Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks,

  23. [31]

    Directional message passing for molecular graphs

    Johannes Gasteiger, Janek Groß, and Stephan Günnemann. Directional message passing for molecular graphs. InInternational Conference on Learning Representations (ICLR), 2020

    2020

  24. [32]

    Margraf, and Stephan Günnemann

    Johannes Gasteiger, Shankari Giri, Johannes T. Margraf, and Stephan Günnemann. Fast and uncertainty- aware directional message passing for non-equilibrium molecules. InMachine Learning for Molecules Workshop, NeurIPS, 2020

    2020

  25. [33]

    Gemnet: Universal directional graph neural networks for molecules

    Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pa...

    2021

  26. [34]

    Gemnet: Universal directional graph neural networks for molecules, 2024

    Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules, 2024. URLhttps://arxiv.org/abs/2106.08903

    2024 arXiv

  27. [35]

    Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing

    Yusong Wang, Tong Wang, Shaoning Li, Xinheng He, Mingyu Li, Zun Wang, Nanning Zheng, Bin Shao, and Tie-Yan Liu. Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing. Nature Communications, 15(1), January 2024. ISSN 2041-1...

    2024 doi

  28. [36]

    Ab initio characterization of protein molecular dynamics with ai2bmd.Nature, 635(8040):1019–1027, November

    Tong Wang, Xinheng He, Mingyu Li, Yatao Li, Ran Bi, Yusong Wang, Chaoran Cheng, Xiangzhen Shen, Jiawei Meng, He Zhang, Haiguang Liu, Zun Wang, Shaoning Li, Bin Shao, and Tie-Yan Liu. Ab initio characterization of protein molecular dynamics with ai2bmd.Nature, 635(8040):1019–10...

  29. [37]

    3d steerable cnns: Learning rotationally equivariant features in volumetric data, 2018

    Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns: Learning rotationally equivariant features in volumetric data, 2018. URLhttps://arxiv.org/abs/ 1807.02547. 16 Machine Learning Interatomic Potentials: library for efficient training, ...

    2018 arXiv

  30. [38]

    Clebsch-gordan nets: a fully fourier space spherical convolutional neural network, 2018

    Risi Kondor, Zhen Lin, and Shubhendu Trivedi. Clebsch-gordan nets: a fully fourier space spherical convolutional neural network, 2018. URLhttps://arxiv.org/abs/1806.09231

    2018 arXiv

  31. [39]

    Ilyes Batatia, Simon Batzner, Dávid Péter Kovács, Albert Musaelian, Gregor N. C. Simm, Ralf Drautz, Christoph Ortner, Boris Kozinsky, and Gábor Csányi. The design space of e(3)-equivariant atom-centered interatomic potentials, 2022. URLhttps://arxiv.org/abs/2205.06643

    2022 arXiv

  32. [40]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, 2018

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, 2018. URL https://arxiv.org/abs/1802.08219

    2018 arXiv

  33. [41]

    Owen, Mordechai Kornbluth, and Boris Kozinsky

    Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications, 14(1), February 2023. ISSN 2041-1723. doi: 10.1038/s4146...

    2023 doi

  34. [42]

    Scaling the leading accuracy of deep equivariant models to biomolecular simulations of realistic size, 2023

    Albert Musaelian, Anders Johansson, Simon Batzner, and Boris Kozinsky. Scaling the leading accuracy of deep equivariant models to biomolecular simulations of realistic size, 2023. URLhttps://arxiv.org/ abs/2304.10061

    2023 arXiv

  35. [43]

    doi: 10.1038/s41586-024-08127-z

    ISSN 1476-4687. doi: 10.1038/s41586-024-08127-z. URL http://dx.doi.org/10.1038/ s41586-024-08127-z

    doi

  36. [44]

    Harry Moore, Nicholas J

    Dávid Péter Kovács, J. Harry Moore, Nicholas J. Browning, Ilyes Batatia, Joshua T. Horton, Yixuan Pu, Venkat Kapil, William C. Witt, Ioan-Bogdan Magdău, Daniel J. Cole, and Gábor Csányi. Mace-off: Short-range transferable machine learning force fields for organic molecules.Jou...

    2025 doi

  37. [45]

    Evaluation of the mace force field architecture: From medicinal chemistry to materials science.The Journal of Chemical Physics, 159(4), July

    Dávid Péter Kovács, Ilyes Batatia, Eszter Sára Arany, and Gábor Csányi. Evaluation of the mace force field architecture: From medicinal chemistry to materials science.The Journal of Chemical Physics, 159(4), July

  38. [46]

    Schütt, Oliver T

    Kristof T. Schütt, Oliver T. Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra, 2021. URLhttps://arxiv.org/abs/2102.03150

    2021 arXiv

  39. [47]

    Euclidean neural networks: e3nn, April 2022

    Mario Geiger, Tess Smidt, Alby M., Benjamin Kurt Miller, Wouter Boomsma, Bradley Dice, Kostiantyn Lapchevskyi, Maurice Weiler, Michał Tyszkiewicz, Simon Batzner, Dylan Madisetti, Martin Uhrin, Jes Frellsen, Nuri Jung, Sophia Sanborn, Mingjian Wen, Josh Rackers, Marcel Rød, and...

    2022 doi

  40. [48]

    Unke and Hartmut Maennel.E3x: E(3)-equivariant deep learning made easy.arXiv preprint arXiv:2401.07595, 2024

    Oliver T. Unke and Hartmut Maennel.E3x: E(3)-equivariant deep learning made easy.arXiv preprint arXiv:2401.07595, 2024

    2024 arXiv

  41. [49]

    Lawrence Zitnick

    Saro Passaro and C. Lawrence Zitnick. Reducing so(3) convolutions to so(2) for efficient equivariant gnns. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023

    2023

  42. [50]

    Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields, 2023. URL https://arxiv.org/abs/2206.07697

    2023 arXiv

  43. [51]

    Wood, Luis Barroso-Luque, Daniel S

    Xiang Fu, Brandon M. Wood, Luis Barroso-Luque, Daniel S. Levine, Meng Gao, Misko Dzamba, and C. Lawrence Zitnick. Learning smooth and expressive interatomic potentials for physical property prediction, 2025. URLhttps://arxiv.org/abs/2502.12147

    2025 arXiv

  44. [52]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

    Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery.Nature, 2023. doi: 10.1038/s41586-023-06735-9

    2023 doi

  45. [53]

    Flax: A neural network library and ecosystem for JAX, 2024

    Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2024. URLhttp://github. com/google/flax. 17 Machine Learning Interatomic Potentials: library for efficie...

    2024

  46. [54]

    Schoenholz and Ekin D

    Samuel S. Schoenholz and Ekin D. Cubuk. Jax, m.d.: A framework for differentiable physics, 2020. URL https://arxiv.org/abs/1912.04232

    2020 arXiv

  47. [55]

    The atomic simulation environment—a python library for working with atoms

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E Castelli, Rune Christensen, Marcin Dułak, Jesper Friis, Michael N Groves, Bjørk Hammer, Cory Hargus, Eric D Hermes, Paul C Jennings, Peter Bjerre Jensen, James Kermode, John R Kitchin, Esben Leonhard Kolsbjerg,...

    2017

  48. [56]

    Dotson, Raimondas Galvelis, John E

    Peter Eastman, Pavan Kumar Behara, David L. Dotson, Raimondas Galvelis, John E. Herr, Josh T. Horton, Yuezhi Mao, John D. Chodera, Benjamin P. Pritchard, Yuanqing Wang, Gianni De Fabritiis, and Thomas E. Markland. Spice, a dataset of drug-like molecules and peptides for traini...

    2023 doi

  49. [57]

    Chodera, and Thomas E

    Peter Eastman, Benjamin P, Pritchard, John D. Chodera, and Thomas E. Markland. Nutmeg and spice: Models and data for biomolecular machine learning, 2024. URLhttps://arxiv.org/abs/2406. 13112

    2024

  50. [58]

    Krishnapriyan

    Shengjie Luo, Tianlang Chen, and Aditi S. Krishnapriyan. Enabling efficient equivariant operations in the fourier basis via gaunt tensor products, 2024. URLhttps://arxiv.org/abs/2401.10216

    2024 arXiv

  51. [59]

    mace-off23

    Harry Moore, David Peter Kovacs, Nicholas J Browning, Ilyes Batatia, Joshua T Horton, Venkat Kapil, William Witt, Ioan Magdau, Daniel Cole, and Gabor Csanyi. Research data supporting "mace-off23",

  52. [60]

    Lawrence Zitnick, and Abhishek Das

    Johannes Gasteiger, Muhammed Shuaibi, Anuroop Sriram, Stephan Günnemann, Zachary Ulissi, C. Lawrence Zitnick, and Abhishek Das. Gemnet-oc: Developing graph neural networks for large and diverse molecular simulation datasets, 2022. URLhttps://arxiv.org/abs/2204.02782

    2022 arXiv

  53. [61]

    An efficient sparse kernel generator for o(3)-equivariant deep networks, 2025

    Vivek Bharadwaj, Austin Glover, Aydin Buluc, and James Demmel. An efficient sparse kernel generator for o(3)-equivariant deep networks, 2025. URLhttps://arxiv.org/abs/2501.13986

    2025 arXiv

  54. [62]

    Universally applicable and tunable graph-based coarse-graining for machine learning force fields, 2025

    Christoph Brunken, Sebastien Boyer, Mustafa Omar, Martin Maarand, Olivier Peltre, Solal Attias, Bakary N’tji Diallo, Anastasia Markina, Olaf Othersen, and Oliver Bent. Universally applicable and tunable graph-based coarse-graining for machine learning force fields, 2025. URLht...

    2025 arXiv

  55. [63]

    16x0e","0e

    Lars L. Schaaf, Ilyes Batatia, Christoph Brunken, Thomas D. Barrett, and Jules Tilly. Boostmd: Accelerating molecular sampling by leveraging ml force field features from previous time-steps, 2024. URLhttps: //arxiv.org/abs/2412.18633. 18 Machine Learning Interatomic Potentials...

    2024 arXiv

  56. [66]

    Huber.Robust Estimation of a Location Parameter, pages 492–518

    Peter J. Huber.Robust Estimation of a Location Parameter, pages 492–518. Springer New York, New York, NY, 1992. ISBN 978-1-4612-4380-9. doi: 10.1007/978-1-4612-4380-9_35. URLhttps://doi.org/ 10.1007/978-1-4612-4380-9_35

    1992 doi

  57. [68]

    URL https://www.repository.cam.ac.uk/handle/1810/366661

  58. [2010]

    doi: 10.1103/physrevlett.104.136403

    ISSN 1079-7114. doi: 10.1103/physrevlett.104.136403. URLhttp://dx.doi.org/10.1103/ PhysRevLett.104.136403

    doi

  59. [2018]

    doi: 10.1038/s41467-018-06169-2

    doi

  60. [2022]

    URL https://arxiv.org/abs/2102.09844

    arXiv

  61. [2023]

    doi: 10.1063/5.0155322

    ISSN 1089-7690. doi: 10.1063/5.0155322. URLhttp://dx.doi.org/10.1063/5.0155322

    doi

  62. [2024]

    URL https://arxiv.org/abs/2410.23179

    arXiv

  63. [7114]

    URL http://dx.doi.org/10.1103/PhysRevLett

    doi: 10.1103/physrevlett.120.145301. URL http://dx.doi.org/10.1103/PhysRevLett. 120.145301

    doi

  64. [9626]

    URLhttp://dx.doi.org/10.1021/acs.jctc.0c00121

    doi: 10.1021/acs.jctc.0c00121. URLhttp://dx.doi.org/10.1021/acs.jctc.0c00121

    doi

Pith tools

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