Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

The paper claims that a multi-time-step integrator can use a cheap, non-conservative distilled force model for the inner loop and an exact outer correction, preserving canonical sampling and stability while speeding up neural-network-potent

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 22:59 UTC pith:ARUOACHC

load-bearing objection Genuinely useful extension of the authors' own DMTS framework; the speedups and sampling checks hold up on the tested systems, but the sampling guarantee for non-conservative forces is not a formal one and a few reporting gaps are easy to fix. the 4 major comments →

arxiv 2602.14975 v3 pith:ARUOACHC submitted 2026-02-16 physics.chem-ph cs.LG

Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Non-Conservative Forces

Nicola\"i Gouraud , C\^ome Cattin , Thomas Pl\'e , Olivier Adjoua , Louis Lagard\`ere , Jean-Philip Piquemal This is my paper
classification physics.chem-ph cs.LG
keywords molecular dynamicsneural network potentialsmultiple time steppingRESPAnon-conservative forcesknowledge distillationLangevin dynamicscanonical sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a multi-time-step integrator for neural-network potentials can use a small, fast, non-conservative distilled force model for the inner steps, as long as the outer step applies the exact difference between the target force and the distilled force. Done this way, the scheme is claimed to be more stable than the conservative distilled version, to require no system-specific fine-tuning, and to give 15-30% additional speed on top of earlier distilled multi-time-stepping, with total accelerations up to 5.64-fold over single-time-step simulation on the tested water boxes, solvated proteins, and small-molecule hydration free energies. The reason a sympathetic reader would care is that the main obstacle to using high-accuracy neural-network potentials in production molecular dynamics is their cost; this method attacks that cost at the integrator level, not by retraining the big potential, and it is claimed to carry over to any foundation neural-network potential, demonstrated on two different ones.

Core claim

On its own terms, the paper's central claim is that non-conservative distilled forces can be embedded safely in a RESPA-style multiple-time-step integrator: the cheap non-conservative model drives the inner loop at small steps, and the more expensive target potential is evaluated only at the outer step, where it supplies the correction F_L = −∇U − F_S. Because the total force field remains −∇U, the dynamics is conservative in aggregate even though the inner model is not. The paper reports that this DMTS-NC scheme is more stable than its conservative counterpart, allows outer time steps pushed to the physical resonance limit (about 6-6.5 fs for water and 5 fs for proteins, extended to 10 fs a

What carries the argument

The load-bearing object is the distilled non-conservative force model: a much smaller equivariant message-passing network (about one-thirtieth the parameters of the target model) that outputs atomic forces directly, avoiding the backpropagation cost of energy-derived forces. It enforces two physical priors by construction: rotation equivariance, and antisymmetric pairwise contributions F_ij − F_ji so that the net force on any connected molecular fragment cancels, matching Newton's third law. This model is inserted into a BAOAB-RESPA split: the inner loop applies the cheap non-conservative force n times per outer step, and the outer step applies F_L = −∇U − F_S, the exact correction from the

Load-bearing premise

Everything rests on the idea that running the inner loop with a force that does not come from any energy function, corrected once per outer step by the exact energy gradient, still samples the right statistical ensemble and stays stable at the large outer steps used; the paper checks this by simulation rather than proving it, and the safety threshold for the rewind fallback is never stated.

What would settle it

At the largest stable outer step claimed (e.g., 6-6.5 fs for water or 5 fs for proteins), run many independent multi-nanosecond DMTS-NC trajectories and compare the potential-energy and temperature histograms to a single-time-step reference; a systematic shift beyond the low-bias identical-variance envelope — or a rewind frequency materially higher than the reported once-per-10-ns — would show that the non-conservative inner model is biasing the sampled measure. A second direct check: monitor F_L = −∇U − F_S along the trajectory; if large sporadic disagreements coincide with jumps in kinetic e

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • DMTS-NC can be layered onto any neural-network potential, so the cost of the expensive model is amortized over n inner steps; on the heavier potential tested, total speedup over single-time-step reached 5.64x.
  • A generic rewind safeguard replaces system-specific active-learning fine-tuning, so a single distilled non-conservative model transfers across water boxes and two solvated proteins without per-system retraining.
  • Outer time steps can be pushed to the resonance limit (6-6.5 fs water, 5 fs proteins) and beyond with hydrogen mass repartitioning and high hydrogen friction (10 fs water, 6-7 fs proteins), with sampling statistics matching single-time-step runs.
  • Observed quantities — oxygen-oxygen radial distributions, temperature and potential-energy distributions, protein backbone RMSD, ligand binding distance, and hydration free energies (mean absolute error 0.154 kcal/mol vs STS) — are preserved within the expected low-bias, same-variance envelope.
  • Diffusion losses are smaller than speed gains in all reported regimes (e.g., 12.6% loss at 4.53x speedup for DMTS-NC; 32.6-39.2% loss at 5.28x speedup with high hydrogen friction), so the authors conclude the acceleration is not offset by reduced sampling rate.

Where Pith is reading between the lines

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

  • The reported 15-30% speedup over conservative distillation comes from skipping energy backpropagation in the inner model; for even larger foundation potentials where backpropagation dominates the cost, the relative gain should grow — a scaling prediction that follows from the architecture but is not tested here.
  • The rewind procedure's threshold is never specified; a practical deployment would need a principled way to set it, and systems with rarer, larger model disagreements (ions, interfaces, strongly strained conformers) may need different thresholds than the water/protein tests.
  • Because high hydrogen friction trades a 33-39% diffusion loss for a 26-33% speed gain, the effective gain for transport properties is smaller than the raw ns/day numbers suggest; the fast-forward Langevin variant recovers part of that loss and is the natural choice when dynamics, not just sampling, matters.
  • A useful stress test would be to run DMTS-NC on a slowly relaxing system (e.g., an ionic liquid or a heavily constrained peptide) and compare not just averages but the tails of the potential-energy distribution against single-time-step results, since rare force-disagreement events caught by the rewind might still leave faint marks in the tails.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DMTS-NC, a multi-time-step (MTS) integrator for neural network potentials in which a distilled, non-conservative force model is used for the inner loop and the exact difference between the reference force and the distilled force is applied at the outer step. The method is tested on bulk water, two solvated proteins, and a set of small molecules, reporting speedups up to 5.64-fold relative to single-time-step (STS) BAOAB and 15–30% over the previous conservative distilled MTS (DMTS) scheme. Additional accelerations are obtained by combining hydrogen mass repartitioning (HMR), high hydrogen friction (HHF), and a rewind safeguard. A proof-of-principle application to MACE-OFF23 is also presented.

Significance. If the method is unbiased, it offers a practical and model-agnostic way to accelerate molecular dynamics with expensive foundation-model neural network potentials, and the use of non-conservative forces to avoid backpropagation is an interesting development. The paper includes multiple validation layers: radial distribution functions, temperature and potential-energy distributions, hydration free energies, diffusion coefficients, and protein RMSD/DBC traces. The code and pretrained models are made available, which strengthens reproducibility. The speedups themselves are substantial and, if confirmed, would be useful to the community.

major comments (4)
  1. [Section 2, Multi-time-step integrator, and Algorithm 1] The statement that F_S and F_L 'are not required to be conservative forces: they only need to satisfy F_S+F_L=-∇U' is not supported. For BAOAB-RESPA, the standard invariant-measure argument relies on each substep being reversible with respect to the target Boltzmann-Gibbs measure. When F_S is non-conservative, the inner B-step is not a Hamiltonian flow, the composition is not reversible, and the stationary distribution of the numerical chain may be biased by an amount depending on the curl of F_S, Δ, and the thermostat. No formal error bound is provided, and the empirical checks (Fig. 1, Table 3) cover only 2–5 ns trajectories. This is load-bearing for the claim of unbiased acceleration. Please provide a rigorous derivation or, failing that, explicitly treat DMTS-NC as an approximate sampler and add direct numerical tests of the invariant measure (e.g., a curl diagnostic, a Δ-dependence
  2. [Section 2, Preventing rare model disagreements] The rewind safeguard is described only qualitatively: the trigger threshold is never specified, and the duration of the STS replacement is given as 'typically a few picoseconds' without a precise value. The text also states that rewinds occur 'on average every 10ns' but does not report the actual number of rewinds in the production runs or whether the wall-clock ns/day numbers in Tables 1, 2, 4, and 6 include rewind overhead. Without this information, the speedups cannot be fully reproduced, and the claim that rewinds 'almost' do not affect speed is not verifiable. Please specify the threshold, the rewind frequency for each system, and the accounting method for wall-clock time.
  3. [Section 3.1, Distillation of MACE-OFF23] The MACE-OFF23 proof-of-principle reports only simulation speeds (Table 6). No sampling or accuracy checks (e.g., temperature/energy distributions, RDF, or RMSD) are given for these systems. Since the paper claims the method is model-agnostic, the demonstration would be stronger with at least one basic dynamic/ensemble comparison for a MACE-based system, to show that the non-conservative distillation does not introduce instabilities or sampling bias when the reference model is MACE-OFF23.
  4. [Section 2, Model training] The reported MAE (1.46 kcal/mol/Å) is not clearly identified as a training-set or test-set error. The comparison with the conservative model of Ref. [52] (MAE=3.44 kcal/mol/Å) is only meaningful if both models are evaluated on the same held-out test split. Please clarify the data split and report test-set errors for both the NC and conservative distilled models.
minor comments (5)
  1. [Section 3, introduction] Typo: 'DMTC-NC' should be 'DMTS-NC' in the sentence 'the improved stability and speed of DMTC-NC'.
  2. [Table 1] The text says DMTS-NC/STS acceleration ranges from 2.94 to 4.31, but the table lists 4.53 for the 4800-atom water box. The text should say 4.53.
  3. [Table 3] For DHFR, the DMTS-NC average temperature is 302.26 K, about 2.3 K above the STS value. The text describes the deviations as 'low-bias'; the larger offset for this system deserves a comment.
  4. [Figure 4] The description of the inset (non-physical artifact peaks) is unclear; please label the inset directly or explain in the caption what frequencies it covers.
  5. [Conclusion] The phrase 'near ab initio accuracy' is stronger than what the 1.46 kcal/mol/Å force MAE alone supports; consider tempering to 'improved accuracy' or citing evidence from the validation tests.

Circularity Check

0 steps flagged

No significant circularity: DMTS-NC is validated against the same reference potential via wall-clock speedups and external sampling benchmarks; self-citations are contextual, not load-bearing.

full rationale

The paper's central claims are empirical: DMTS-NC is faster and more stable than conservative DMTS and STS for the tested systems. Speedups are wall-clock measurements (Tables 1, 2, 4, 6), not quantities derived from fitted parameters. The distilled non-conservative model is trained on reference labels from FeNNix-Bio1(M) or MACE-OFF23(S), but this is the intended approximation used inside the integrator. The validation then compares DMTS-NC to single-time-step BAOAB of the same reference potential through RDFs, temperature and energy distributions, hydration free energies, diffusion coefficients, and RMSD/DBC trajectories. This is an appropriate external benchmark for the integrator, not a circular reduction: the target potential is fixed, and the integrator's output is tested against it independently. The relation F_L = -∇U - F_S (Section 2, 'Multi-time-step integrator') is a definition of the correction force; it ensures the net force is correct at the outer time step but does not itself define the measured speedups or sampling accuracy. The self-citations ([52] for the DMTS framework, [33] for FeNNix-Bio1) describe prior work that this paper extends; they do not supply the validation data or force the numerical conclusions. Limitations are present — the invariant-measure preservation of a non-conservative inner force is not proven and rests on empirical checks, and the rewind threshold in Section 2 is unspecified — but these are correctness and reproducibility concerns, not instances of a derivation reducing to its own inputs. The conclusion even states that 'Future work will focus on ... subtle sampling properties,' acknowledging the open theoretical question. No fitted input is renamed as a prediction, no uniqueness theorem is imported, and no ansatz is smuggled via citation. Accordingly, no circular step can be exhibited with a specific equation or reduction.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The central method is algorithmic rather than physical: it postulates no new forces or particles. Its load-bearing assumptions are standard MTS splitting, distillation fidelity, and empirical claims about non-conservative force stability. Several hand-set parameters (HMR mass, HHF friction, rewind threshold, architecture size, chosen timesteps) control the speedup and stability numbers.

free parameters (5)
  • Hydrogen mass repartitioning increment = +3.0 Da on H
    Hand-set; shifts H stretch frequencies downward, allowing larger external time step; used in all DMTS/DMTS-NC simulations.
  • High Hydrogen Friction coefficient = γ_H = 10 ps^-1 (γ_O = 1 ps^-1)
    Hand-set; damps H oscillations, enables Δ up to 10 fs; costs 25-39% diffusion loss depending on thermostat.
  • Rewind trigger threshold = not reported
    The on-the-fly rewind procedure is activated when the force-model difference 'exceeds a threshold'; the value is never given, so the stability/speed trade-off is not reproducible.
  • Distilled NC model architecture hyperparameters = N_f=64, n_c=8, λ_max=1, n_l=2, n_s=8, N_Q=8; 286,736 parameters
    Hand-chosen to balance fidelity (MAE 1.46 kcal/mol) and speed; these choices drive the reported improvements.
  • External time step Δ and inner step δ = δ=1-1.25 fs; Δ=5-10 fs depending on system
    Selected by stability scanning; the headline speedups are computed at the maximum stable Δ, a hand-tuned operating point.
axioms (7)
  • domain assumption FeNNix-Bio1(M) and MACE-OFF23(S) provide reference force fields accurate enough for the target simulations.
    The paper does not validate either foundation model against QM; claimed 'near-ab initio accuracy' is inherited from prior training.
  • domain assumption BAOAB-RESPA splitting with non-conservative inner force F_S and exact outer correction F_L = -∇U - F_S converges to the canonical Langevin dynamics at used Δ.
    Algorithm 1 relies on standard MTS error theory; no proof is given for the non-conservative inner field, only empirical agreement.
  • domain assumption The SPICE2 subset labeled with FeNNix-Bio1 forces is representative of water, proteins, and small molecules simulated.
    Training set composition is described only as 'a subset of conformations'; exact selection not specified.
  • domain assumption The architectural priors (rotational equivariance, zero total force) prevent the severe artifacts reported for non-conservative force models.
    Invoked in Section 2 and abstract; supported only by empirical stability of the tested systems.
  • domain assumption HMR preserves configurational sampling and does not significantly alter diffusion.
    Standard method cited to [66]; the paper's own STS-HMR diffusion comparison supports the assumption.
  • domain assumption HHF and FFL thermostats maintain the target Boltzmann measure closely enough at 300 K.
    Used to reach Δ=10 fs; diffusion loss is measured, but configurational bias beyond potential energy is not exhaustively tested.
  • domain assumption Rewind events are rare enough and unbiased enough not to affect sampled statistics.
    Stated as occurring every ~10 ns at maximum stable Δ; no quantitative analysis of rewind-induced perturbation.

pith-pipeline@v1.3.0-alltime-deepseek · 17144 in / 19812 out tokens · 186612 ms · 2026-08-02T22:59:25.171676+00:00 · methodology

0 comments
read the original abstract

Following our previous work (J. Phys. Chem. Lett., 2026, 17, 5, 1288-1295), we propose the DMTS-NC approach, a distilled multi-time-step (DMTS) strategy using non-conservative (NC) forces to further accelerate atomistic molecular dynamics simulations using foundation neural network models such as FeNNix-Bio1. There, a dual-level reversible reference system propagator algorithm (RESPA) formalism couples a target accurate conservative potential to a simplified distilled representation optimized for the production of non-conservative forces. Despite being non-conservative, the distilled architecture is designed to enforce key physical priors, such as equivariance under rotation and cancellation of atomic force components. These choices facilitate the distillation process and therefore improve drastically the robustness of simulation, significantly limiting abnormal discrepancies between the two models, thus achieving excellent agreement with the forces data. Overall, the DMTS-NC scheme is found to be more stable and efficient than its conservative counterpart with additional speedups reaching 15-30% over DMTS. Requiring no fine-tuning steps, it is easier to implement and can be pushed to the limit of the systems physical resonances to maintain accuracy while providing maximum efficiency. We obtain additional speedup by combining hydrogen mass repartitioning (HMR), High Hydrogen Friction (HHF) to further extended the largest timestep up to 10fs of our schemes while conserving stability and accuracy. As for DMTS, DMTS-NC is applicable to any neural network potential and can be applied to approaches that are computationally heavier than FeNNix-Bio1. We show a proof of principle applying the approach to the distillation of MACE-OFF23 with consequent speedups ranging from 3.66 to 5.64 compared to single timestep.

Figures

Figures reproduced from arXiv: 2602.14975 by C\^ome Cattin, Jean-Philip Piquemal, Louis Lagard\`ere, Nicola\"i Gouraud, Olivier Adjoua, Thomas Pl\'e.

Figure 1
Figure 1. Figure 1: Sampling properties of bulk water simulations. All data was collected from 2 nanosecond simulations, [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Time evolution of the protein backbone RMSD and the ligand’s Distance to Bound Configuraiton (DBC) [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hydration free energy of small molecules. Error bars are computed from three independent lambda [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hydrogen velocity auto-correlation spectra for various external time steps, in a system of 1600 water [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MSD and diffusion coefficients for various time steps and thermostats. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Distilling first-principles accuracy into compact machine learning potentials for condensed-phase chemistry

    physics.chem-ph 2026-06 unverdicted novelty 6.0

    Distilled compact MLIPs from transfer-learned teachers reproduce observables more reliably than same-size models trained directly and enable practical PIMD umbrella sampling of water dissociation at TiO2 interface wit...

Reference graph

Works this paper leans on

73 extracted references · 14 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Leimkuhler and C

    B. Leimkuhler and C. Matthews.Molecular Dynamics. Springer, 2015

    2015

  2. [2]

    M. P. Allen and D. J. Tildesley.Computer Simulation of Liquids. Oxford University Press, 2017

    2017

  3. [3]

    Molecular simulations: past, present, and future (a Topical Issue in EPJB)

    G Ciccotti, C Dellago, M Ferrario, E R Hern´ andez, and M E Tuckerman. “Molecular simulations: past, present, and future (a Topical Issue in EPJB)”. en. In:Eur. Phys. J. B95.1 (Jan. 2022)

    2022

  4. [4]

    Jensen.Introduction to Computational Chemistry

    F. Jensen.Introduction to Computational Chemistry. Wiley, 2017.isbn: 9781118825990.url:https:// books.google.fr/books?id=UZOVDQAAQBAJ

    2017

  5. [5]

    Polarizable Atomic Multipole Water Model for Molecular Mechanics Simulation

    P. Ren and J. W. Ponder. “Polarizable Atomic Multipole Water Model for Molecular Mechanics Simulation”. In:The Journal of Physical Chemistry B107.24 (2003), pp. 5933–5947.issn: 1520-6106

    2003

  6. [6]

    New Interaction Parameters for Charged Amino Acid Side Chains in the GROMOS Force Field

    M. M. Reif, P. H. H¨ unenberger, and C. Oostenbrink. “New Interaction Parameters for Charged Amino Acid Side Chains in the GROMOS Force Field”. In:Journal of Chemical Theory and Computation8.10 (2012), pp. 3705–3723.issn: 1549-9618

    2012

  7. [7]

    ff14SB: Improving the Accuracy of Protein Side Chain and Backbone Parameters from ff99SB

    J. A. Maier, C. Martinez, K. Kasavajhala, L. Wickstrom, K. E. Hauser, and C. Simmerling. “ff14SB: Improving the Accuracy of Protein Side Chain and Backbone Parameters from ff99SB”. In:Journal of Chemical Theory and Computation11.8 (2015), pp. 3696–3713.issn: 1549-9618. 14

    2015

  8. [8]

    Improved Peptide and Protein Torsional Energetics with the OPLS-AA Force Field

    M. J. Robertson, J. Tirado-Rives, and W. L. Jorgensen. “Improved Peptide and Protein Torsional Energetics with the OPLS-AA Force Field”. In:Journal of Chemical Theory and Computation11.7 (2015), pp. 3499– 3509.issn: 1549-9618.doi:10.1021/acs.jctc.5b00356.url:https://doi.org/10.1021/acs.jctc. 5b00356

    work page doi:10.1021/acs.jctc.5b00356.url:https://doi.org/10.1021/acs.jctc 2015

  9. [9]

    CHARMM36m: an improved force field for folded and intrinsically disordered proteins

    J. Huang et al. “CHARMM36m: an improved force field for folded and intrinsically disordered proteins”. In:NATURE METHODS(2017).doi:{10.1038/NMETH.4067}

    work page doi:10.1038/nmeth.4067 2017

  10. [10]

    Chapter 7 Empirical Force Fields for Proteins: Current Status and Future Di- rections

    Alexander D. MacKerell. “Chapter 7 Empirical Force Fields for Proteins: Current Status and Future Di- rections”. In: vol. 1. Annual Reports in Computational Chemistry. Elsevier, 2005, pp. 91–102.doi:10 . 1016 / S1574 - 1400(05 ) 01007 - 8.url:https : / / www . sciencedirect . com / science / article / pii / S1574140005010078

    2005

  11. [11]

    Generalized neural-network representation of high-dimensional potential- energy surfaces

    J¨ org Behler and Michele Parrinello. “Generalized neural-network representation of high-dimensional potential- energy surfaces”. In:Physical Review Letters98.14 (2007), p. 146401

    2007

  12. [12]

    Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons

    Albert P Bart´ ok, Mike C Payne, Risi Kondor, and G´ abor Cs´ anyi. “Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons”. In:Physical Review Letters104.13 (2010), p. 136403

    2010

  13. [13]

    Accurate neural network description of surface phonons in reactive gas–surface dynamics: N2+ Ru (0001)

    Khosrow Shakouri, J¨ org Behler, J¨ org Meyer, and Geert-Jan Kroes. “Accurate neural network description of surface phonons in reactive gas–surface dynamics: N2+ Ru (0001)”. In:The Journal of Physical Chemistry Letters8.10 (2017), pp. 2131–2136

    2017

  14. [14]

    ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost

    Justin S Smith, Olexandr Isayev, and Adrian E Roitberg. “ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost”. In:Chemical Science8.4 (2017), pp. 3192–3203

    2017

  15. [15]

    SchNet: a continuous-filter convolutional neural network for modeling quantum interactions

    K. T. Sch¨ utt, P.-J. Kindermans, H. E. Sauceda, S. Chmiela, A. Tkatchenko, and K.-R. M¨ uller. “SchNet: a continuous-filter convolutional neural network for modeling quantum interactions”. In:Proceedings of the 31st International Conference on Neural Information Processing Systems. NIPS’17. Long Beach, California, USA: Curran Associates Inc., 2017, pp. 9...

    2017

  16. [16]

    Towards exact molec- ular dynamics simulations with machine-learned force fields

    Stefan Chmiela, Huziel E Sauceda, Klaus-Robert M¨ uller, and Alexandre Tkatchenko. “Towards exact molec- ular dynamics simulations with machine-learned force fields”. In:Nature Communications9.1 (2018), pp. 1– 10

    2018

  17. [17]

    Incorporating long-range physics in atomic-scale machine learning

    Andrea Grisafi and Michele Ceriotti. “Incorporating long-range physics in atomic-scale machine learning”. In:The Journal of Chemical Physics151.20 (2019), p. 204105

    2019

  18. [18]

    Accurate and transferable mul- titask prediction of chemical properties with an atoms-in-molecules neural network

    Roman Zubatyuk, Justin S. Smith, Jerzy Leszczynski, and Olexandr Isayev. “Accurate and transferable mul- titask prediction of chemical properties with an atoms-in-molecules neural network”. In:Science Advances 5.8 (2019), eaav6490

    2019

  19. [19]

    Extending the Applicability of the ANI Deep Learning Molecular Potential to Sulfur and Halogens

    Christian Devereux et al. “Extending the Applicability of the ANI Deep Learning Molecular Potential to Sulfur and Halogens”. In:Journal of Chemical Theory and Computation16.7 (2020), pp. 4192–4202

    2020

  20. [20]

    SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects

    Oliver T Unke, Stefan Chmiela, Michael Gastegger, Kristof T Sch¨ utt, Huziel E Sauceda, and Klaus-Robert M¨ uller. “SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects”. In: Nature Communications12.1 (2021), pp. 1–14

    2021

  21. [21]

    Gemnet: Universal directional graph neural networks for molecules

    Johannes Gasteiger, Florian Becker, and Stephan G¨ unnemann. “Gemnet: Universal directional graph neural networks for molecules”. In:Advances in Neural Information Processing Systems34 (2021), pp. 6790–6802

    2021

  22. [22]

    Multi-scale approach for the prediction of atomic scale properties

    Andrea Grisafi, Jigyasa Nigam, and Michele Ceriotti. “Multi-scale approach for the prediction of atomic scale properties”. In:Chemical Science12.6 (2021), pp. 2078–2090

    2021

  23. [23]

    Learning local equivariant representations for large-scale atomistic dynamics

    Albert Musaelian et al. “Learning local equivariant representations for large-scale atomistic dynamics”. In: Nature Communications14.1 (2023), p. 579

    2023

  24. [24]

    Force-field-enhanced neural network interactions: from local equivariant embedding to atom-in-molecule properties and long-range effects

    Thomas Pl´ e, Louis Lagard` ere, and Jean-Philip Piquemal. “Force-field-enhanced neural network interactions: from local equivariant embedding to atom-in-molecule properties and long-range effects”. In:Chemical Science14 (44 2023), pp. 12554–12569

    2023

  25. [25]

    Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments

    Oliver T. Unke et al. “Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments”. In:Science Advances10.14 (2024), eadn4397

    2024

  26. [26]

    Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields

    Adil Kabylda et al. “Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields”. In:Journal of the American Chemical Society147.37 (2025), pp. 33723–33734. 15

    2025

  27. [27]

    SchNetPack 2.0: A neural network toolbox for atomistic machine learning

    Kristof T Sch¨ utt, Stefaan S P Hessmann, Niklas W A Gebauer, Jonas Lederer, and Michael Gastegger. “SchNetPack 2.0: A neural network toolbox for atomistic machine learning”. en. In:J. Chem. Phys.158.14 (Apr. 2023), p. 144801

    2023

  28. [28]

    DeePMD-kit v2: A software package for deep potential models

    Jinzhe Zeng et al. “DeePMD-kit v2: A software package for deep potential models”. en. In:J. Chem. Phys. 159.5 (Aug. 2023)

    2023

  29. [29]

    MLatom 3: A platform for machine learning-enhanced computational chemistry simu- lations and workflows

    Pavlo O Dral et al. “MLatom 3: A platform for machine learning-enhanced computational chemistry simu- lations and workflows”. en. In:J. Chem. Theory Comput.20.3 (Feb. 2024), pp. 1193–1213

    2024

  30. [30]

    FeNNol: An efficient and flexible library for building force-field-enhanced neural network potentials

    Thomas Pl´ e, Olivier Adjoua, Louis Lagard` ere, and Jean-Philip Piquemal. “FeNNol: An efficient and flexible library for building force-field-enhanced neural network potentials”. In:The Journal of Chemical Physics 161.4 (July 2024), p. 042502.issn: 0021-9606

    2024

  31. [31]

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

    Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and G´ abor Cs´ anyi. “MACE: Higher order equivariant message passing neural networks for fast and accurate force fields”. In:Advances in Neural Information Processing Systems35 (2022), pp. 11423–11436

    2022

  32. [32]

    MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules

    D´ avid P´ eter Kov´ acs et al. “MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules”. In:Journal of the American Chemical Society147.21 (2025), pp. 17598–17611

    2025

  33. [33]

    A Foundation Model for Accurate Atomistic Simulations in Drug Design

    Thomas Pl´ e et al. “A Foundation Model for Accurate Atomistic Simulations in Drug Design”. In:preprint ChemRxiv, DOI: 10.26434/chemrxiv-2025-f1hgn-v4(2025)

    work page doi:10.26434/chemrxiv-2025-f1hgn-v4(2025 2025

  34. [34]

    Pushing the Accuracy Limit of Foundation Neural Network Models with Quantum Monte Carlo Forces and Path Integrals

    Anouar Benali et al. “Pushing the Accuracy Limit of Foundation Neural Network Models with Quantum Monte Carlo Forces and Path Integrals”. In:preprint arXiv: 2504.07948, DOI: 10.48550/arXiv.2504.07948 (2025).doi:10.48550/arXiv.2504.07948.url:https://arxiv.org/abs/2504.07948

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.07948 2025

  35. [35]

    Robust and efficient configurational molecular sampling via Langevin dynamics

    Benedict Leimkuhler and Charles Matthews. “Robust and efficient configurational molecular sampling via Langevin dynamics”. In:The Journal of Chemical Physics138.17 (May 2013), p. 174102.issn: 0021-9606. doi:10.1063/1.4802990. eprint:https://pubs.aip.org/aip/jcp/article- pdf/doi/10.1063/1. 4802990/15462311/174102_1_online.pdf.url:https://doi.org/10.1063/1.4802990

    doi:10.1063/1.4802990 2013

  36. [36]

    Rational construction of stochastic numerical methods for molecular sampling

    B Leimkuhler and C Matthews. “Rational construction of stochastic numerical methods for molecular sampling”. In:Appl. Math. Res. Express(June 2012)

    2012

  37. [37]

    Reversible multiple time scale molecular dynamics

    M. Tuckerman, B. J. Berne, and G. J. Martyna. “Reversible multiple time scale molecular dynamics”. In:The Journal of Chemical Physics97.3 (Aug. 1992), pp. 1990–2001.issn: 0021-9606. eprint:https : //pubs.aip.org/aip/jcp/article-pdf/97/3/1990/19000200/1990\_1\_online.pdf

    1992

  38. [38]

    Efficient multiple time step method for use with Ewald and particle mesh Ewald for large biomolecular systems

    Ruhong Zhou, Edward Harder, Huafeng Xu, and BJ Berne. “Efficient multiple time step method for use with Ewald and particle mesh Ewald for large biomolecular systems”. In:The Journal of chemical physics 115.5 (2001), pp. 2348–2358

    2001

  39. [39]

    Pushing the Limits of Multiple-Time-Step Strate- gies for Polarizable Point Dipole Molecular Dynamics

    Louis Lagard` ere, F´ elix Aviat, and Jean-Philip Piquemal. “Pushing the Limits of Multiple-Time-Step Strate- gies for Polarizable Point Dipole Molecular Dynamics”. In:The Journal of Physical Chemistry Letters10.10 (2019), pp. 2593–2599.doi:10.1021/acs.jpclett.9b00901. eprint:https://doi.org/10.1021/acs. jpclett.9b00901.url:https://doi.org/10.1021/acs.jpc...

    work page doi:10.1021/acs.jpclett.9b00901 2019

  40. [40]

    Dangers of Multiple Time Step Methods

    J. J. Biesiadecki and R. D. Skeel. “Dangers of Multiple Time Step Methods”. In:Journal of Computational Physics109.2 (1993), pp. 318–328.issn: 0021-9991

    1993

  41. [41]

    Verlet-I/R-RESPA/Impulse is Limited by Nonlinear Instabilities

    Q. Ma, J. Izaguirre, and R. Skeel. “Verlet-I/R-RESPA/Impulse is Limited by Nonlinear Instabilities”. In: SIAM Journal on Scientific Computing24.6 (2003), pp. 1951–1973

    2003

  42. [42]

    Efficient multiple time scale molecular dynamics: Using colored noise thermostats to stabilize resonances

    Joseph A. Morrone, Thomas E. Markland, Michele Ceriotti, and B. J. Berne. “Efficient multiple time scale molecular dynamics: Using colored noise thermostats to stabilize resonances”. In:The Journal of Chemical Physics134.1 (2011), p. 014103.doi:10.1063/1.3518369

    work page doi:10.1063/1.3518369 2011

  43. [43]

    Combining Iteration-Free Polarization with Large Time Step Stochastic-Isokinetic Integration

    A. Albaugh, M. E. Tuckerman, and T. Head-Gordon. “Combining Iteration-Free Polarization with Large Time Step Stochastic-Isokinetic Integration”. In:Journal of Chemical Theory and Computation15.4 (2019), pp. 2195–2205.doi:10.1021/acs.jctc.9b00072

    work page doi:10.1021/acs.jctc.9b00072 2019

  44. [44]

    Stochastic, resonance-free multiple time-step algo- rithm for molecular dynamics with very large time steps

    B. Leimkuhler, D. T. Margul, and M. E. Tuckerman. “Stochastic, resonance-free multiple time-step algo- rithm for molecular dynamics with very large time steps”. In:Molecular Physics111.22-23 (2013), pp. 3579– 3594. 16

    2013

  45. [45]

    A Stochastic, Resonance-Free Multiple Time-Step Algorithm for Polarizable Models That Permits Very Large Time Steps

    D. T. Margul and M. E. Tuckerman. “A Stochastic, Resonance-Free Multiple Time-Step Algorithm for Polarizable Models That Permits Very Large Time Steps”. In:Journal of Chemical Theory and Computation 12.5 (2016), pp. 2170–2180.doi:10.1021/acs.jctc.6b00188

    work page doi:10.1021/acs.jctc.6b00188 2016

  46. [46]

    Velocity Jumps for Molecular Dynamics

    Nicola ¨ ı Gouraud, Louis Lagard` ere, Olivier Adjoua, Thomas Pl´ e, Pierre Monmarch´ e, and Jean-Philip Pique- mal. “Velocity Jumps for Molecular Dynamics”. In:Journal of Chemical Theory and Computation21.6 (2025), pp. 2854–2866.issn: 1549-9618.doi:10.1021/acs.jctc.5c00023.url:https://doi.org/10. 1021/acs.jctc.5c00023

    work page doi:10.1021/acs.jctc.5c00023.url:https://doi.org/10 2025

  47. [47]

    A Versatile Multiple Time Step Scheme for Efficient ab Initio Molecular Dynamics Simulations

    Elisa Liberatore, Rocco Meli, and Ursula Rothlisberger. “A Versatile Multiple Time Step Scheme for Efficient ab Initio Molecular Dynamics Simulations”. In:Journal of Chemical Theory and Computation14.6 (2018). PMID: 29624388, pp. 2834–2842.doi:10.1021/acs.jctc.7b01189. eprint:https://doi.org/10.1021/ acs.jctc.7b01189.url:https://doi.org/10.1021/acs.jctc.7b01189

    work page doi:10.1021/acs.jctc.7b01189 2018

  48. [48]

    Scalable hybrid deep neural networks/polarizable potentials biomolecular sim- ulations including long-range effects

    Th´ eo Jaffrelot Inizan et al. “Scalable hybrid deep neural networks/polarizable potentials biomolecular sim- ulations including long-range effects”. In:Chemical Science14 (20 2023), pp. 5438–5452

    2023

  49. [49]

    Learning Inter- atomic Potentials at Multiple Scales

    Xiang Fu, Albert Musaelian, Anders Johansson, Tommi Jaakkola, and Boris Kozinsky. “Learning Inter- atomic Potentials at Multiple Scales”. In:NeurIPS 2023 AI for Science Workshop. 2023.url:https : //openreview.net/forum?id=qFIs4hYZaZ

    2023

  50. [50]

    Toward Improving Multiple Time Step QM/MM Simula- tions with ∆-Machine Learning

    Reilly Osadchey, Kwangho Nam, and Qiang Cui. “Toward Improving Multiple Time Step QM/MM Simula- tions with ∆-Machine Learning”. In:The Journal of Physical Chemistry B129.40 (2025). PMID: 40997780, pp. 10451–10466.doi:10 . 1021 / acs . jpcb . 5c05255. eprint:https : / / doi . org / 10 . 1021 / acs . jpcb . 5c05255.url:https://doi.org/10.1021/acs.jpcb.5c05255

    work page doi:10.1021/acs.jpcb.5c05255 2025

  51. [51]

    Machine learning-enhanced multiple time-step ab initio molecular dynamics

    Fran¸ cois Mouvet, Nicholas J. Browning, Pablo Baudin, Elisa Liberatore, and Ursula Rothlisberger. “Machine learning-enhanced multiple time-step ab initio molecular dynamics”. In:The Journal of Chemical Physics 163.18 (Nov. 2025), p. 184115.issn: 0021-9606.doi:10.1063/5.0288262. eprint:https://pubs.aip. org / aip / jcp / article - pdf / doi / 10 . 1063 / ...

    work page doi:10.1063/5.0288262 2025

  52. [52]

    Accelerating Molecular Dynamics Simulations with Foundation Neural Network Models Using Multiple Time Steps and Distillation

    Cˆ ome Cattin, Thomas Pl´ e, Olivier Adjoua, Nicola ¨ ı Gouraud, Louis Lagard` ere, and Jean-Philip Piquemal. “Accelerating Molecular Dynamics Simulations with Foundation Neural Network Models Using Multiple Time Steps and Distillation”. In:The Journal of Physical Chemistry Letters17.5 (2026). PMID: 41562233, pp. 1288–1295.doi:10.1021/acs.jpclett.5c03720....

    doi:10.1021/acs.jpclett.5c03720 2026

  53. [53]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. “Distilling the knowledge in a neural network”. In:arXiv preprint arXiv:1503.02531(2015)

    Pith/arXiv arXiv 2015

  54. [54]

    Knowledge distillation: A survey

    Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. “Knowledge distillation: A survey”. In:International journal of computer vision129.6 (2021), pp. 1789–1819

    2021

  55. [55]

    Forcenet: A graph neural network for large-scale quantum calculations

    Weihua Hu et al. “Forcenet: A graph neural network for large-scale quantum calculations”. In:arXiv preprint arXiv:2103.01436(2021)

    Pith/arXiv arXiv 2021

  56. [56]

    Equiformerv2: Improved equivariant trans- former for scaling to higher-degree representations

    Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. “Equiformerv2: Improved equivariant trans- former for scaling to higher-degree representations”. In:arXiv preprint arXiv:2306.12059(2023)

    Pith/arXiv arXiv 2023

  57. [57]

    Orb: A fast, scalable neural network potential

    Mark Neumann et al. “Orb: A fast, scalable neural network potential”. In:arXiv preprint arXiv:2410.22570 (2024)

    Pith/arXiv arXiv 2024

  58. [58]

    Orb-v3: atomistic simulation at scale

    Benjamin Rhodes et al. “Orb-v3: atomistic simulation at scale”. In:arXiv preprint arXiv:2504.06231(2025)

    Pith/arXiv arXiv 2025

  59. [59]

    How simple can you go? An off-the-shelf transformer approach to molecular dynamics

    Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T Unke, Klaus-Robert M¨ uller, and Stefan Gugler. “How simple can you go? An off-the-shelf transformer approach to molecular dynamics”. In:arXiv preprint arXiv:2503.01431(2025)

    arXiv 2025

  60. [60]

    The dark side of the forces: assessing non-conservative force models for atomistic machine learning

    Filippo Bigi, Marcel F. Langer, and Michele Ceriotti. “The dark side of the forces: assessing non-conservative force models for atomistic machine learning”. In:Forty-second International Conference on Machine Learn- ing. 2025.url:https://openreview.net/forum?id=OEl3L8osas

    2025

  61. [61]

    Directional message passing for molecular graphs

    Johannes Gasteiger, Janek Groß, and Stephan G¨ unnemann. “Directional message passing for molecular graphs”. In:arXiv preprint arXiv:2003.03123(2020). 17

    Pith/arXiv arXiv 2003

  62. [62]

    Repulsive interatomic potentials calculated at three levels of theory

    Kai Nordlund, Susi Lehtola, and Gerhard Hobler. “Repulsive interatomic potentials calculated at three levels of theory”. In:Phys. Rev. A111 (3 Mar. 2025), p. 032818.doi:10.1103/PhysRevA.111.032818. url:https://link.aps.org/doi/10.1103/PhysRevA.111.032818

    work page doi:10.1103/physreva.111.032818 2025

  63. [63]

    SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials

    Peter Eastman et al. “SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials”. In:Scientific Data10.1 (2023), p. 11

    2023

  64. [64]

    Nutmeg and SPICE: Models and Data for Biomolecular Machine Learning

    Peter Eastman, Benjamin P. Pritchard, John D. Chodera, and Thomas E. Markland. “Nutmeg and SPICE: Models and Data for Biomolecular Machine Learning”. In:Journal of Chemical Theory and Computation 20.19 (2024), pp. 8583–8593

    2024

  65. [65]

    2024.url:https : / / kellerjordan.github.io/posts/muon/

    Keller Jordan et al.Muon: An optimizer for hidden layers in neural networks. 2024.url:https : / / kellerjordan.github.io/posts/muon/

    2024

  66. [66]

    Improving efficiency of large time-scale molecular dynamics simulations of hydrogen-rich systems

    K. Anton Feenstra, Berk Hess, and Herman J. C. Berendsen. “Improving efficiency of large time-scale molecular dynamics simulations of hydrogen-rich systems”. In:Journal of Computational Chemistry20.8 (1999), pp. 786–798. eprint:https://onlinelibrary.wiley.com/doi/pdf/10.1002/%28SICI%291096- 987X%28199906%2920%3A8%3C786%3A%3AAID- JCC5%3E3.0.CO%3B2- B.url:h...

    1999

  67. [67]

    Optimal friction matrix for underdamped Langevin sampling

    Martin Chak, Nikolas Kantas, Tony Leli` evre, and Grigorios A. Pavliotis. “Optimal friction matrix for underdamped Langevin sampling”. In:ESAIM: Mathematical Modelling and Numerical Analysis57.6 (Nov. 2023), pp. 3335–3371.issn: 2804-7214.doi:10.1051/m2an/2023083.url:http://dx.doi.org/10.1051/ m2an/2023083

    arXiv 2023

  68. [68]

    Fast-forward Langevin dynamics with momentum flips

    Mahdi Hijazi, David M. Wilkins, and Michele Ceriotti. “Fast-forward Langevin dynamics with momentum flips”. In:The Journal of Chemical Physics148.18 (May 2018).issn: 1089-7690.doi:10.1063/1.5029833. url:http://dx.doi.org/10.1063/1.5029833

    work page doi:10.1063/1.5029833 2018

  69. [69]

    Tinker-HP: a massively parallel molecular dynamics package for multiscale simula- tions of large complex systems with advanced point dipole polarizable force fields

    Louis Lagard` ere et al. “Tinker-HP: a massively parallel molecular dynamics package for multiscale simula- tions of large complex systems with advanced point dipole polarizable force fields”. In:Chemical Science9 (4 2018), pp. 956–972

    2018

  70. [70]

    Tinker-hp: Accelerating molecular dynamics simulations of large complex systems with advanced point dipole polarizable force fields using gpus and multi-gpu systems

    Olivier Adjoua et al. “Tinker-hp: Accelerating molecular dynamics simulations of large complex systems with advanced point dipole polarizable force fields using gpus and multi-gpu systems”. In:Journal of Chemical Theory and Computation17.4 (2021), pp. 2034–2053

    2021

  71. [71]

    Using collective variables to drive molecular dynamics simulations

    Giacomo Fiorin, Michael L. Klein, and J´ erˆ ome H´ enin. “Using collective variables to drive molecular dynamics simulations”. In:Molecular Physics111.22–23 (Dec. 2013), pp. 3345–3362.issn: 1362-3028.doi:10.1080/ 00268976.2013.813594.url:http://dx.doi.org/10.1080/00268976.2013.813594

    arXiv 2013

  72. [72]

    Lambda-ABF: Simplified, Portable, Accurate, and Cost-Effective Alchemical Free- Energy Computation

    Louis Lagard` ere et al. “Lambda-ABF: Simplified, Portable, Accurate, and Cost-Effective Alchemical Free- Energy Computation”. In:Journal of Chemical Theory and Computation20.11 (2024), pp. 4481–4498

    2024

  73. [73]

    Tuckerman.Statistical Mechanics: Theory and Molecular Simulation

    M. Tuckerman.Statistical Mechanics: Theory and Molecular Simulation. Oxford Graduate Texts. OUP Oxford, 2010. 18

    2010