Pith. sign in

REVIEW 2 major objections 30 references

Orientation features on ellipsoidal beads are necessary for equivariant machine-learning coarse-grained potentials to reproduce liquid-water structure.

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 · grok-4.5

2026-07-14 01:06 UTC pith:GYBUCJ3Y

load-bearing objection Workable equivariant CG for ellipsoids with torques; anisotropy beats isotropic on water structure, but the ablation confounds shape with orientation and the evidence stays thin. the 2 major comments →

arxiv 2607.10002 v1 pith:GYBUCJ3Y submitted 2026-07-10 cond-mat.mtrl-sci

Anisotropic representations for E(3)-equivariant machine learning coarse-grained potentials

Varun Shankar , Emil Annevelink This is my paper
classification cond-mat.mtrl-sci
keywords coarse-grainingmachine learning potentialsequivariant neural networksanisotropic beadsliquid waterorientation featuresforce and torque matching
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.

This paper claims that coarse-graining polar molecules as isotropic points discards directional information that is required for correct liquid structure, and that representing each molecule as an oriented ellipsoid whose shape and principal axes enter an E(3)-equivariant message-passing network allows the model to learn energies, forces, and torques directly from atomistic data. On liquid water the anisotropic model matches radial and angular distribution functions and relative orientation correlations; an otherwise identical isotropic baseline systematically shifts nearest-neighbor peaks and degrades angular order. The authors therefore argue that orientation-dependent features are essential, not optional, for systems governed by directional interactions. Even a three-atom-to-one-bead reduction already yields 7–27× wall-clock speedups while exposing rotational observables that isotropic beads cannot define. The result matters for anyone who wants mesoscale simulations of molecular liquids, polymers, or biomolecules that retain structural fidelity without hand-tuned classical force fields.

Core claim

An E(3)-equivariant message-passing network whose node features include ellipsoid semi-axes and spherical-harmonic projections of principal axes learns a coarse-grained potential that quantitatively recovers radial, angular, and orientational structure of liquid water from atomistic forces and torques, whereas the same architecture stripped of orientation features produces systematic errors in short-range order and angular correlations.

What carries the argument

The anisotropic node embedding that concatenates each bead’s semi-axes with real spherical-harmonic projections (up to ℓ=2) of two principal axes derived from its quaternion; forces and torques are then obtained by energy-conserving automatic differentiation with respect to center-of-mass position and quaternion coordinates.

Load-bearing premise

The rigid-body ellipsoid fitted from the molecular inertia tensor is assumed to be a sufficient representation of each water molecule, so that discarding intramolecular flexibility does not destroy the liquid structure the network must learn.

What would settle it

Train anisotropic and isotropic models on identical water trajectories and check whether the isotropic first RDF peak remains shifted and the ADFs at 3–5 Å remain visibly wrong while the anisotropic model stays on the atomistic reference; if improved data or architecture make the isotropic model match equally well, the necessity claim fails.

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

If this is right

  • Isotropic bead models will systematically misplace nearest-neighbor distances and angular correlations in polar or shape-anisotropic liquids.
  • Once beads carry orientation, rotational structural observables become well-defined diagnostics of coarse-grained fidelity.
  • Even a minimal three-to-one reduction already multiplies simulation throughput by roughly an order of magnitude, with larger gains expected at bigger system sizes.
  • The same equivariant anisotropic construction supplies a route to mesoscale models of polymers and biomolecules where shape anisotropy is stronger.
  • Oriented beads tighten the reverse-mapping step required by active-learning loops that refine the potential from atomistic oracles.

Where Pith is reading between the lines

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

  • For molecules much larger than water the degrees-of-freedom reduction compounds, so wall-clock gains should widen substantially for polymers or proteins.
  • The rigid inertia-tensor mapping will need multi-site or flexible generalizations before conformationally soft biomolecules can be treated faithfully.
  • The paper’s own unsatisfactory uncertainty quantification implies that anisotropic CG models will require purpose-built UQ before fully autonomous active learning is reliable.
  • Many existing isotropic machine-learning CG potentials for hydrogen-bonding liquids may already be underfitted on angular structure even when their RDFs look acceptable.

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

2 major / 0 minor

Summary. The manuscript introduces an anisotropic machine-learning coarse-grained potential (MLCGP) that maps each molecule to a rigid ellipsoidal bead whose semi-axes and principal-axis orientations are obtained from the inertia tensor (Eqs. 1–3). Node embeddings concatenate the semi-axes with l≤2 spherical-harmonic projections of the principal axes and are processed by an E(3)-equivariant message-passing network that predicts bead energies; forces and torques are recovered by automatic differentiation, the latter via a quaternion-to-Cartesian map T(q) (Eqs. 7–9). On liquid water (64 molecules, 10 ps NVT trajectories driven by a MACE all-atom potential) the anisotropic model reproduces radial distribution functions, angular distribution functions at several cut-offs, and relative-orientation projections of the coarse-grained reference. An isotropic baseline that removes both shape and orientation features shows clear RDF peak shifts and ADF degradation. Wall-clock speed-ups of 7–27 imes relative to the all-atom simulation are reported for 64- and 512-molecule systems.

Significance. If the necessity claim holds, the work supplies a concrete, equivariant route to anisotropic CG that simultaneously learns energies, forces and torques and recovers rotational observables inaccessible to spherical beads. The explicit comparison against an isotropic ablation, the energy-conserving torque formulation, and the demonstration of non-negligible speed-ups even for a three-atom molecule are genuine technical contributions that would interest the CG and MLIP communities. The framework is also positioned for active-learning pipelines that exploit orientation for reverse mapping, a practical advantage over isotropic models.

major comments (2)
  1. §4.2 and Figs. 3–4: the isotropic baseline “excludes the orientation- and shape-dependent features,” so the ablation simultaneously removes both the spherical-harmonic orientation embedding and the semi-axes (a,b,c). Because water’s inertia tensor is non-spherical, any RDF/ADF degradation cannot be attributed cleanly to the absence of orientation; it could arise from missing shape parameters alone. A pure-orientation ablation (shape retained, orientation removed) is required to isolate the necessity claim that is central to the abstract and §5.2.
  2. §5.1–5.2: all structural comparisons rest on 10 ps trajectories of a 64-molecule system with no reported statistical uncertainties, block averages or independent replicas. At this length the first few coordination shells are only sparsely sampled; without error bars it is impossible to judge whether the visual agreement of the anisotropic model (or the degradation of the isotropic model) is statistically significant. Longer production runs or bootstrap estimates are needed before the quantitative-match claim can be regarded as established.

Circularity Check

0 steps flagged

No significant circularity: structural observables emerge from force/torque-matched dynamics, not by construction from the training targets or mapping.

full rationale

The paper's load-bearing chain is standard bottom-up CG force-matching. Atomic energies/forces from a MACE reference are aggregated into bead E/F/τ via the rigid inertia-tensor mapping (Eqs. 1–4); an E(3)-equivariant network is trained to reproduce those targets; CG MD is then run and the resulting RDFs, ADFs and orientation distributions are compared to the identically mapped reference trajectories. The structural agreement is therefore an independent dynamical prediction, not a quantity that is fitted or definitionally forced by the inputs. The isotropic ablation (node embedding stripped of shape and orientation features) is an imperfect control, but that is an experimental-design limitation, not a circular reduction. No free-energy parameters are fitted and then re-predicted, no uniqueness theorem is imported from overlapping authors, and no ansatz is smuggled via self-citation. The derivation is self-contained against the external AA reference.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claim rests on the adequacy of the rigid-ellipsoid mapping, the force-matching principle, and the particular orientation embedding; all are domain assumptions or paper-specific constructions rather than free parameters fitted to the final structural metrics. Hyperparameters exist but are secondary to the modeling choices.

free parameters (3)
  • multi-task loss coefficients for energies, forces, torques, total energy and stress
    Coefficients are scaled to normalize magnitude disparities across observables; exact values are not reported but chosen by the authors to balance the loss.
  • network and training hyperparameters (lr schedule, early-stopping patience, spherical-harmonic cutoff l=2)
    Standard choices (lr=1e-2, patience=500 epochs, l=2) that affect final accuracy and are not derived from first principles.
  • active-learning high-error selection thresholds
    Unspecified numerical criteria used to decide which configurations to re-label and retrain on.
axioms (4)
  • domain assumption Force- and torque-matching of aggregated atomistic forces yields a thermodynamically consistent CG potential for structural observables.
    Standard bottom-up CG assumption invoked throughout §4; known to be approximate for dynamics and free energies.
  • domain assumption The rigid ellipsoidal mapping via the inertia tensor (Eqs. 1–3) is a sufficient representation of a water molecule for liquid-structure purposes.
    Core modeling choice in §4.1; water is flexible and polar, so the rigid-body approximation may discard relevant degrees of freedom.
  • domain assumption E(3)-equivariant message passing with learned tensor products preserves physical symmetries and improves data efficiency.
    Inherited from the MLIP literature cited in §2 and treated as given.
  • domain assumption The MACE universal potential provides sufficiently accurate ground-truth forces for liquid water at 300 K.
    Used as the sole reference calculator; any systematic bias in MACE is inherited by the CG model.
invented entities (2)
  • anisotropic node embedding that concatenates semi-axes with l≤2 spherical-harmonic projections of principal axes no independent evidence
    purpose: to inject shape and orientation information into the equivariant GNN while remaining E(3)-equivariant
    Constructed specifically for this architecture; no external experimental signature beyond the water RDF/ADF matches shown.
  • MLCGP with automatic differentiation of torques via the quaternion-to-Cartesian transformation T(q) no independent evidence
    purpose: to enable energy-conserving prediction of both forces and torques on oriented beads
    The overall model is the paper’s contribution; validation is internal to the water test case.

pith-pipeline@v1.1.0-grok45 · 15576 in / 3076 out tokens · 41243 ms · 2026-07-14T01:06:46.812241+00:00 · methodology

0 comments
read the original abstract

Coarse-graining (CG) lowers the computational cost of atomistic simulations by representing groups of atoms as effective interaction sites, reducing the degrees of freedom of the system but often compromising structural fidelity or requiring system-specific parameterization. Here, we introduce a novel anisotropic machine learning CG potential that extends the point particle representation of atomic nuclei to massive ellipsoidal beads with orientation-dependent features, enabling the learning of energies, forces, and torques directly from atomistic data. The anisotropic representation is physically motivated for polar and asymmetric molecules, where directional interactions and shape anisotropy play important roles in determining structure and dynamics. Using an equivariant message-passing neural network, the model accurately reproduces radial and angular distribution functions as well as relative orientation correlations in liquid water, demonstrating that both translational and rotational dynamics are well captured. Comparison with an isotropic baseline reveals that the lack of orientation information leads to systematic errors in short and long range order and degradation of angular correlations, proving orientation features are essential for accurate coarse-graining. The anisotropic model also exposes rotational structural observables fundamentally inaccessible to isotropic representations, with minimal computational overhead. Even for coarse-graining just three degrees of freedom, CG simulations achieve 7-27$\times$ speedups while preserving structural fidelity, highlighting the efficiency gains of this systemic reduction. This framework establishes the feasibility and necessity of learned equivariant representations for anisotropic CG modeling and provides a path towards accurate and efficient mesoscopic simulations of complex molecular liquids, polymers, and biomolecular systems.

Figures

Figures reproduced from arXiv: 2607.10002 by Emil Annevelink, Varun Shankar.

Figure 1
Figure 1. Figure 1: Presents an overview of the coarse-graining procedure for an ethylene carbonate molecule. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Initial configurations of the 64-molecule liquid water system in atomistic (left) and beaded [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Radial distribution function g(r) of coarse-grained bead centers for liquid water at 300K. The anisotropic MLCGP accurately reproduces the reference coarse-grained data (AA) derived from all-atom simulations using the MACE potential, with agreement in both peak positions and magnitudes. In contrast, the isotropic baseline (MLCGP-ISO) shows a clear rightward shift of the first coordination peak and reduced … view at source ↗
Figure 4
Figure 4. Figure 4: Angular distribution functions (ADFs) for triplets of coarse-grained beads computed at [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Depiction of a characteristic pair of ellipsoids, with local coordinate axes [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distributions of relative orientation projections between neighboring coarse-grained el [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set

    G. Kresse and J. Furthmüller. “Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set”. In:Computational Materials Science6.1 (1996), pp. 15–50.issn: 0927-0256.doi:https://doi.org/10.1016/0927-0256(96)00008-0

    doi:10.1016/0927-0256(96)00008-0 1996

  2. [2]

    Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

    Volker L. Deringer, Miguel A. Caro, and Gábor Csányi. “Machine Learning Interatomic Potentials as Emerging Tools for Materials Science”. In:Advanced Materials31.46 (2019), p. 1902765.doi:https://doi.org/10.1002/adma.201902765

    doi:10.1002/adma.201902765 2019

  3. [3]

    Neural Network Potentials: A Concise Overview of Methods

    Emir Kocer, Tsz Wai Ko, and Jörg Behler. “Neural Network Potentials: A Concise Overview of Methods”. In:Annual Review of Physical Chemistry73.Volume 73, 2022 (2022), pp. 163–186. issn: 1545-1593.doi:https://doi.org/10.1146/annurev-physchem-082720-034254

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

  4. [4]

    CHIPS-FF: Evaluating Universal Machine Learning Force Fields for Material Properties

    Daniel Wines and Kamal Choudhary. “CHIPS-FF: Evaluating Universal Machine Learning Force Fields for Material Properties”. In:ACS Materials Letters7.6 (2025), pp. 2105–2114. doi:10.1021/acsmaterialslett.5c00093

    doi:10.1021/acsmaterialslett.5c00093 2025

  5. [5]

    MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields

    Ilyes Batatia et al. “MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields”. In:Advances in Neural Information Processing Systems. Ed. by Alice H. Oh et al. 2022.url:https://openreview.net/forum?id=YPpSngE-ZU

    2022

  6. [6]

    Wood et al.UMA: A Family of Universal Models for Atoms

    Brandon M. Wood et al.UMA: A Family of Universal Models for Atoms. 2025.url:https: //arxiv.org/abs/2506.23971

    arXiv 2025

  7. [7]

    2025.url:https://arxiv

    Benjamin Rhodes et al.Orb-v3: atomistic simulation at scale. 2025.url:https://arxiv. org/abs/2504.06231

    Pith/arXiv arXiv 2025

  8. [8]

    A review of advancements in coarse-grained molecular dynamics simulations

    Soumil Y. Joshi and Sanket A. Deshmukh. “A review of advancements in coarse-grained molecular dynamics simulations”. In:Molecular Simulation47.10-11 (2021), pp. 786–803.doi: 10.1080/08927022.2020.1828583

    doi:10.1080/08927022.2020.1828583 2021

  9. [9]

    Perspective: Coarse-grained models for biomolecular systems

    William G. Noid. “Perspective: Coarse-grained models for biomolecular systems”. In:J. Chem. Phys.139.9 (2013), p. 090901.doi:10.1063/1.4818908

    doi:10.1063/1.4818908 2013

  10. [10]

    Two decades ofMartini: Better beads, broader scope

    Siewert J. Marrink et al. “Two decades ofMartini: Better beads, broader scope”. In:WIREs Comp. Mol. Sci.13.1 (2023).doi:10.1002/wcms.1620

    work page doi:10.1002/wcms.1620 2023

  11. [11]

    Perspective: Dissipative Particle Dynamics

    P. Español and P. B. Warren. “Perspective: Dissipative Particle Dynamics”. In:J. Chem. Phys. 146.15 (2017), p. 150901.doi:10.1063/1.4979514

    doi:10.1063/1.4979514 2017

  12. [12]

    Modification of the overlap potential to mimic a linear site–site potential

    J. G. Gay and B. J. Berne. “Modification of the overlap potential to mimic a linear site–site potential”. In:J. Chem. Phys.74.6 (1981), pp. 3316–3319.doi:10.1063/1.441483

    doi:10.1063/1.441483 1981

  13. [13]

    Extension and generalization of the Gay-Berne potential

    Douglas J. Cleaver et al. “Extension and generalization of the Gay-Berne potential”. In:Phys. Rev. E54 (1 July 1996), pp. 559–567.doi:10.1103/PhysRevE.54.559

    work page doi:10.1103/physreve.54.559 1996

  14. [14]

    Generalizedneural-networkrepresentationofhigh-dimensional potential-energy surfaces

    JörgBehlerandMicheleParrinello.“Generalizedneural-networkrepresentationofhigh-dimensional potential-energy surfaces”. In:Phys. Rev. Lett.98.14 (2007), p. 146401.doi:10 . 1103 / PhysRevLett.98.146401

    2007

  15. [15]

    Gaussian approximation potentials: The accuracy of quantum me- chanics, without the electrons

    Albert P. Bartók et al. “Gaussian approximation potentials: The accuracy of quantum me- chanics, without the electrons”. In:Phys. Rev. Lett.104.13 (2010), p. 136403.doi:10.1103/ PhysRevLett.104.136403. 16

    2010

  16. [16]

    E(3)-equivariant graph neural networks for data-efficient and accu- rate interatomic potentials

    Sebastian Batzner et al. “E(3)-equivariant graph neural networks for data-efficient and accu- rate interatomic potentials”. In:Nat. Commun.13 (2022), p. 2909.doi:10.1038/s41467- 022-29939-5

    doi:10.1038/s41467- 2022

  17. [17]

    Approaching coupled cluster accuracy with a general-purpose neural network potential through active learning

    John S. Smith et al. “Approaching coupled cluster accuracy with a general-purpose neural network potential through active learning”. In:Nature Commun. Eng.1 (2022), p. 47.doi: 10.1088/2632-2153/abc9fe

    doi:10.1088/2632-2153/abc9fe 2022

  18. [18]

    Uncertainty-aware dynamics for machine learning interatomic potentials

    Xiang Li et al. “Uncertainty-aware dynamics for machine learning interatomic potentials”. In: Nature Mach. Intell.5 (2023), pp. 332–343.doi:10.1038/s43588-023-00406-5

    doi:10.1038/s43588-023-00406-5 2023

  19. [19]

    Machine learning for coarse-grained molecular simulation: a survey of methods, models, and applications

    Yaqoot Sultan et al. “Machine learning for coarse-grained molecular simulation: a survey of methods, models, and applications”. In:Curr. Opin. Struct. Biol.81 (2023), p. 102533.doi: 10.1016/j.sbi.2023.102533

    work page doi:10.1016/j.sbi.2023.102533 2023

  20. [20]

    DeePCG: Constructing coarse-grained models via deep neural networks

    Lixin Zhang et al. “DeePCG: Constructing coarse-grained models via deep neural networks”. In:J. Chem. Phys.154.4 (2021), p. 044123.doi:10.1063/5.0026133

    doi:10.1063/5.0026133 2021

  21. [21]

    Coarse-graining molecular dynamics with graph neural networks

    Brendan E. Husic et al. “Coarse-graining molecular dynamics with graph neural networks”. In:J. Chem. Phys.158.6 (2023), p. 064109.doi:10.1063/5.0143724

    work page doi:10.1063/5.0143724 2023

  22. [22]

    Active learning a coarse-grained neural network model for bulk water from sparse training data

    James Loeffler et al. “Active learning a coarse-grained neural network model for bulk water from sparse training data”. In:Molecular Systems Design & Engineering5.4 (2020), pp. 995– 1006.doi:10.1039/C9ME00184K

    work page doi:10.1039/c9me00184k 2020

  23. [23]

    On-the-fly active learning of interpretable Bayesian force fields for atomistic rare events

    Jonathan Vandermause et al. “On-the-fly active learning of interpretable Bayesian force fields for atomistic rare events”. In:npj Computational Materials6.1 (2020), p. 20.issn: 2057-3960. doi:10.1038/s41524-020-0283-z

    doi:10.1038/s41524-020-0283-z 2020

  24. [24]

    Uncertainty Driven Active Learning of Coarse Grained Free Energy Models

    Robert Duschatko et al. “Uncertainty driven active learning of coarse-grained free-energy models”. In:arXiv preprint(2022).doi:10.48550/arXiv.2210.16364

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.16364 2022

  25. [25]

    Deep Potential Molecular Dynamics: A Scalable Model with the Ac- curacy of Quantum Mechanics

    Linfeng Zhang et al. “Deep Potential Molecular Dynamics: A Scalable Model with the Ac- curacy of Quantum Mechanics”. In:Phys. Rev. Lett.120 (14 Apr. 2018), p. 143001.doi: 10.1103/PhysRevLett.120.143001

    doi:10.1103/physrevlett.120.143001 2018

  26. [26]

    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.issn: 2041-1723.doi:10.1038/ s41467-023-36329-y

    2023

  27. [27]

    Coarse-Graining with Equivariant Neural Networks: A Path Towards Accurate and Data-Efficient Models

    Oliver Loose et al. “Coarse-graining with equivariant neural networks: a path towards accurate and data-efficient models”. In:arXiv preprint(2023).doi:10.48550/arXiv.2309.09290

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.09290 2023

  28. [28]

    Anisotropic molecular coarse-graining by force and torque matching with neural networks

    Caleb J. Wilson and Liang Huang. “Anisotropic molecular coarse-graining by force and torque matching with neural networks”. In:J. Chem. Phys.159.2 (2023), p. 024110.doi:10.1063/ 5.0143724

    2023

  29. [29]

    2022.doi:10.48550/arXiv.2205.06643

    Ilyes Batatia et al.The Design Space of E(3)-Equivariant Atom-Centered Interatomic Poten- tials. 2022.doi:10.48550/arXiv.2205.06643

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.06643 2022

  30. [30]

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

    A. P. Thompson et al. “LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales”. In:Comp. Phys. Comm.271 (2022), p. 108171.doi:10.1016/j.cpc.2021.108171. 17

    doi:10.1016/j.cpc.2021.108171 2022