pith. sign in

arxiv: 2505.01058 · v5 · submitted 2025-05-02 · ❄️ cond-mat.mtrl-sci · cond-mat.soft

Coarse-grained graph architectures for all-atom force predictions

Sungwoo Kang , Jinwoong Chae This is my paper

Pith reviewed 2026-05-22 17:51 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci cond-mat.soft
keywords coarse-grained graph neural networksall-atom force fieldsmachine learning interatomic potentialsequivariant modelsgrain embeddingelectrolytesRDXforce prediction
0
0 comments X

The pith

Grain embeddings let equivariant graphs predict all-atom forces from coarse-grained nodes.

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

The paper introduces the CGAA-FF framework, which performs message passing on grains rather than individual atoms while still returning forces for every atom. Grain embedding compresses atom coordinates into fewer nodes that preserve local geometry, allowing the model to use any equivariant graph architecture. Tests on EC/EMC electrolytes and RDX phases show force errors of 0.201 and 0.253 eV Å^{-1} respectively. The same setup delivers roughly ten-fold speed and five-fold memory savings over standard machine-learning interatomic potentials. The authors position the method as a general route to efficient all-atom simulations of soft-matter systems.

Core claim

The CGAA-FF model incorporates coarse-grained message passing inside an all-atom force field by encoding groups of atoms into grain nodes via grain embedding. This produces both grain-level energies and per-atom forces while exploiting the equivariant nature of the underlying graph model. On EC/EMC organic electrolytes the force error reaches 0.201 eV Å^{-1}; on RDX crystalline and disordered phases it reaches 0.253 eV Å^{-1}. These accuracies are obtained with approximately ten-fold higher computational speed and five-fold higher memory efficiency than conventional MLIPs. Because the coarse-graining step can be inserted into any equivariant architecture, the framework is presented as a drop

What carries the argument

Grain embedding, which maps atomistic coordinates into a reduced set of grain nodes that retain enough local geometry for accurate per-atom force recovery without full atom-level message passing.

If this is right

  • Per-atom forces remain usable for molecular dynamics even though message passing occurs only at the grain level.
  • Graph size shrinks with the number of grains, directly lowering memory and time costs for larger systems.
  • The same coarse-graining layer can be added to existing equivariant models without retraining their core layers.
  • Efficiency gains make longer-timescale all-atom simulations of soft-matter systems more practical.

Where Pith is reading between the lines

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

  • The approach may generalize to other organic or polymeric materials where local bonding motifs repeat.
  • Combining grain embeddings with existing multi-scale coarse-graining schemes could further reduce degrees of freedom.
  • Because the overhead of embedding is small, the method could be used to generate larger training datasets for the same compute budget.

Load-bearing premise

Mapping atomistic coordinates into a smaller set of grain nodes via grain embedding retains sufficient local geometric information to recover accurate per-atom forces without explicit atom-level message passing.

What would settle it

Force prediction errors exceeding 0.3 eV Å^{-1} on held-out EC/EMC or RDX configurations of similar size and disorder would show that the grain-node representation loses critical geometric detail.

Figures

Figures reproduced from arXiv: 2505.01058 by Jinwoong Chae, Sungwoo Kang.

Figure 1
Figure 1. Figure 1: Schematic of the CGAA-FF architecture. Illustration of the a, grain-based graph representation, b, node embedding. c, overall CGAA-FF model structure, d, interaction block, and e, convolution layers. The star marks in e indicate components that differ from the original NequIP model. 𝐹𝑗𝛼 = − 𝜕𝐸tot 𝜕𝑟𝑗𝛼 = − ∑ 𝜕𝐸tot 𝜕𝑟 ′ 𝑘𝛼 𝜕𝑟 ′ 𝑘𝛼 𝜕𝑟𝑗𝛼 𝑘∈{𝑛𝐽} − 𝜕𝐸tot 𝜕𝑅𝐽𝛼 𝜕𝑅𝐽𝛼 𝜕𝑟𝑗𝛼 . (4) where α denotes the directional index… view at source ↗
read the original abstract

We introduce a machine-learning framework termed coarse-grained all-atom force field (CGAA-FF), which incorporates coarse-grained message passing within an all-atom force field using equivariant nature of graph models. The CGAA-FF model employs grain embedding to encode atomistic coordinates into nodes representing grains rather than individual atoms, enabling predictions of both grain-level energies and atom-level forces. Tested on EC/EMC organic electrolytes and RDX crystalline and disordered phases, CGAA-FF achieves 0.201 and 0.253 eV A-1, respectively, while providing about 10-fold and 5-fold higher computational speed and memory efficiency, respectively, than conventional MLIPs. Since this CGAA framework can be integrated into any equivariant architecture, we believe this work opens the door to efficient all-atom simulations of soft-matter systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces the Coarse-Grained All-Atom Force Field (CGAA-FF) framework that incorporates coarse-grained message passing into an equivariant graph architecture for all-atom force prediction. Atomistic coordinates are mapped via grain embedding to a reduced set of grain nodes; the model then predicts grain-level energies and atom-level forces through grain-level equivariant message passing. Empirical results are reported on EC/EMC electrolytes (force error 0.201 eV Å^{-1}) and RDX phases (0.253 eV Å^{-1}), together with claimed 10-fold and 5-fold gains in speed and memory efficiency relative to conventional MLIPs. The approach is presented as architecture-agnostic and suitable for efficient soft-matter simulations.

Significance. If the performance numbers are backed by complete training protocols, held-out validation, error bars, and direct baseline comparisons, the work would demonstrate a practical route to reducing the cost of equivariant message passing while recovering per-atom forces, which could enable larger-scale all-atom simulations in materials and soft-matter modeling.

major comments (1)
  1. [§3] §3 (Grain Embedding): the central claim that grain-level equivariant message passing suffices for accurate atom-level forces rests on the assumption that the grain embedding operator preserves all local geometric information (pairwise distances, angles, orientations) required for force recovery. The manuscript provides no explicit invertibility argument, reconstruction step, or ablation that isolates the effect of intra-grain compression; without such evidence the mapping from grain features back to per-atom forces remains under-determined when multiple atoms map to one grain node.
minor comments (2)
  1. [Abstract] Abstract: numerical performance claims are stated without reference to training/validation splits, error bars, or the precise conventional MLIP baselines used for the speed-up factors; these details should be added or cross-referenced to the main text.
  2. [Abstract] Notation: units appear as 'eV A-1' in the abstract; adopt the standard 'eV Å^{-1}' consistently and define all symbols (e.g., grain embedding dimension) at first use.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive comments and for recognizing the potential of the CGAA-FF framework. We address the single major comment below with clarifications and a commitment to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§3] §3 (Grain Embedding): the central claim that grain-level equivariant message passing suffices for accurate atom-level forces rests on the assumption that the grain embedding operator preserves all local geometric information (pairwise distances, angles, orientations) required for force recovery. The manuscript provides no explicit invertibility argument, reconstruction step, or ablation that isolates the effect of intra-grain compression; without such evidence the mapping from grain features back to per-atom forces remains under-determined when multiple atoms map to one grain node.

    Authors: We thank the referee for this precise observation. In the CGAA-FF architecture the grain embedding is an equivariant aggregation that maps groups of atoms to grain nodes while retaining the original atomic coordinates and relative vectors for subsequent force evaluation. Grain-level message passing computes a coarse-grained energy; atom-resolved forces are then obtained by automatic differentiation of this energy with respect to the input atomic positions, which are never discarded. Because the embedding is constructed from equivariant tensor products and the differentiation is performed on the full set of atomic coordinates, the local geometric information required for force recovery is preserved by construction rather than reconstructed. We acknowledge that the original manuscript did not include an explicit invertibility argument or a dedicated ablation on intra-grain compression. In the revised version we will add (i) a concise mathematical description of the embedding operator showing how relative positions and orientations are maintained and (ii) an ablation study that varies grain size while reporting force errors and computational cost, thereby isolating the effect of the compression step. revision: yes

Circularity Check

0 steps flagged

No significant circularity; results are empirical on held-out data

full rationale

The paper introduces a CGAA-FF ML framework that uses grain embedding and equivariant message passing to predict atom-level forces from coarse-grained representations. Performance metrics (0.201 eV Å^{-1} on EC/EMC, 0.253 eV Å^{-1} on RDX) are reported as test-set outcomes on held-out systems, not as quantities derived by construction from fitted parameters or self-citations. No load-bearing step in the abstract or described architecture reduces a claimed prediction to an input by definition, renaming, or self-referential fitting. The central claim remains an empirical demonstration of efficiency gains over conventional MLIPs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The framework rests on standard equivariant graph neural network properties and on the unstated premise that a learned grain embedding can compress atomistic detail without destroying force accuracy. No explicit free parameters or invented physical entities are named in the abstract.

free parameters (1)
  • grain embedding dimension and mapping parameters
    The grain embedding step necessarily introduces trainable parameters whose values are determined during model fitting.
axioms (1)
  • standard math Equivariant graph models correctly preserve rotational and translational symmetries when operating on grain nodes.
    The abstract invokes the equivariant nature of graph models as a foundational property.
invented entities (1)
  • grain embedding no independent evidence
    purpose: To encode sets of atom coordinates into compact grain nodes for message passing.
    This is a newly introduced representational step whose independent validation is not provided in the abstract.

pith-pipeline@v0.9.0 · 5668 in / 1340 out tokens · 40881 ms · 2026-05-22T17:51:45.346979+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    & Parrinello, M

    Behler, J. & Parrinello, M. Generalized Neural -Network Representation of High-Dimensional Potential-Energy Surfaces. Phys. Rev. Lett. 98, 146401 (2007)

    work page 2007

  2. [2]

    P., Payne, M

    Bartók, A. P., Payne, M. C., Kondor, R. & Csányi, G. Gauss- ian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons. Phys. Rev. Lett. 104, 136403 (2010)

    work page 2010

  3. [3]

    Unke, O. T. et al. Machine Learning Force Fields. Chem. Rev. 121, 10142–10186 (2021)

    work page 2021

  4. [4]

    Cheng, B., Mazzola, G., Pickard, C. J. & Ceriotti, M. Evi- dence for supercritical behaviour of high-pressure liquid hydrogen. Na- ture 585, 217–220 (2020)

    work page 2020

  5. [5]

    Deringer, V. L. et al. Origins of structural and electronic transitions in disordered silicon. Nature 589, 59–64 (2021)

    work page 2021

  6. [6]

    & Parrinello, M

    Yang, M., Raucci, U. & Parrinello, M. Reactant-induced dy- namics of lithium imide surfaces during the ammonia decomposition process. Nat. Catal. 6, 829–836 (2023)

    work page 2023

  7. [7]

    & Deringer, V

    Zhou, Y., Zhang, W., Ma, E. & Deringer, V. L. Device-scale atomistic modelling of phase-change memory materials. Nat Electron 6, 746–754 (2023)

    work page 2023

  8. [8]

    & Kang, Y

    Kang, S., Han, S. & Kang, Y. First -Principles Calculations of Luminescence Spectra of Real-Scale Quantum Dots. ACS Mater. Au 2, 103–109 (2022)

    work page 2022

  9. [9]

    Dajnowicz, S. et al. High-Dimensional Neural Network Po- tential for Liquid Electrolyte Simulations. J. Phys. Chem. B 126, 6271– 6280 (2022)

    work page 2022

  10. [10]

    Magdău, I.-B. et al. Machine learning force fields for molec- ular liquids: Ethylene Carbonate/Ethyl Methyl Carbonate binary sol- vent. npj Comput. Mater. 9, (2023)

    work page 2023

  11. [11]

    & Cheng, J

    Wang, F. & Cheng, J. Understanding the solvation structures of glyme-based electrolytes by machine learning molecular dynamics. Chin. J. Struct. Chem. 42, 100061 (2023)

    work page 2023

  12. [12]

    Zhu, D. et al. Investigation of the Degradation of LiPF 6– in Polar Solvents through Deep Potential Molecular Dynamics. J. Phys. Chem. Lett. 15, 4024–4030 (2024)

    work page 2024

  13. [13]

    & Zahn, S

    Shayestehpour, O. & Zahn, S. Structure and transport prop- erties of LiTFSI-based deep eutectic electrolytes from machine-learned interatomic potential simulations. J. Chem. Phys. 161, (2024)

    work page 2024

  14. [14]

    Gong, S. et al. A predictive machine learning force -field framework for liquid electrolyte development. Nat. Mach. Intell. 7, 543–552 (2025)

    work page 2025

  15. [15]

    Hong, S. J. et al. First-Principles-Based Machine-Learning Molecular Dynamics for Crystalline Polymers with van der Waals In- teractions. J. Phys. Chem. Lett. 12, 6000–6006 (2021)

    work page 2021

  16. [16]

    L., Caro, M

    Deringer, V. L., Caro, M. A. & Csányi, G. Machine Learning Interatomic Potentials as Emerging Tools for Materials Science. Adv. Mater. 31, 1902765 (2019)

    work page 2019

  17. [17]

    Batzner, S. et al. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat. Commun. 13, 2453 (2022)

    work page 2022

  18. [18]

    P., Simm, G

    Batatia, I., Kovács, D. P., Simm, G. N. C., Ortner, C. & Csányi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. NeurIPS 35, 7138 (2022)

    work page 2022

  19. [19]

    How graph neural network interatomic potentials extrapolate: Role of the message-passing algorithm

    Kang, S. How graph neural network interatomic potentials extrapolate: Role of the message-passing algorithm. J. Chem. Phys. 161, 244102 (2024)

    work page 2024

  20. [20]

    Ju, S. et al. Application of pretrained universal machine - learning interatomic potential for physicochemical simulation of liquid electrolytes in Li -ion battery. Preprint at https://doi.org/10.48550/arXiv.2501.05211 (2025)

    work page doi:10.48550/arxiv.2501.05211 2025

  21. [21]

    & Han, S

    Park, Y., Kim, J., Hwang, S. & Han, S. Scalable Parallel Al- gorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations. J. Chem. Theory Comput. 20, 4857 –4868 (2024)

    work page 2024

  22. [22]

    Majewski, M. et al. Machine learning coarse-grained poten- tials of protein thermodynamics. Nat. Commun. 14, (2023)

    work page 2023

  23. [23]

    R., Vandermause, J., Molinari, N

    Duschatko, B. R., Vandermause, J., Molinari, N. & Kozinsky, B. Uncertainty driven active learning of coarse grained free energy models. npj Comput. Mater. 10, (2024)

    work page 2024

  24. [24]

    H., Larentzos, J

    Lee, B. H., Larentzos, J. P., Brennan, J. K. & Strachan, A. Graph neural network coarse-grain force field for the molecular crystal RDX. npj Comput. Mater. 10, (2024)

    work page 2024

  25. [25]

    & Goerigk, L

    Grimme, S., Ehrlich, S. & Goerigk, L. Effect of the damping function in dispersion corrected density functional theory. J. Comput. Chem. 32, 1456–1465 (2011)

    work page 2011

  26. [26]

    & Ceriotti, M

    Bigi, F., Langer, M. & Ceriotti, M. The dark side of the forces: assessing non-conservative force models for atomistic machine learning. Preprint at https://doi.org/10.48550/arXiv.2412.11569 (2025)

    work page doi:10.48550/arxiv.2412.11569 2025

  27. [27]

    Yu, Q. et al. Extending atomic decomposition and many - body representation with a chemistry -motivated approach to machine learning potentials. Nat. Comput. Sci. (2025) doi:10.1038/s43588-025- 00790-0

    work page doi:10.1038/s43588-025- 2025

  28. [28]

    Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and contin- uum scales. Comput. Phys. Commun. 271, 108171 (2022)

    work page 2022

  29. [29]

    Hjorth Larsen, A. et al. The atomic simulation environ- ment—a Python library for working with atoms. J. Phys. Condens. Matter 29, 273002 (2017)

    work page 2017

  30. [30]

    & Izumi, F

    Momma, K. & Izumi, F. VESTA 3 for three-dimensional vis- ualization of crystal, volumetric and morphology data. J. Appl. Crys- tallogr. 44, 1272–1276 (2011). Acknowledgement This work was supported by the Nano and Material Technology Development Programs through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (Gra...

    work page 2011