Machine Learning Multiscale Interactions

Albert Mosella-Montoro; Alexandre Tkatchenko; \`Alex Sol\'e; Eliseo Ruiz; Javier Ruiz-Hidalgo; Sergio Su\'arez-Dou; Silvia G\'omez-Coca

arxiv: 2605.25710 · v1 · pith:IYLM3RI6new · submitted 2026-05-25 · ⚛️ physics.chem-ph · cond-mat.mtrl-sci· cs.LG· physics.comp-ph

Machine Learning Multiscale Interactions

\`Alex Sol\'e , Sergio Su\'arez-Dou , Albert Mosella-Montoro , Silvia G\'omez-Coca , Eliseo Ruiz , Alexandre Tkatchenko , Javier Ruiz-Hidalgo This is my paper

Pith reviewed 2026-06-29 19:39 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cond-mat.mtrl-scics.LGphysics.comp-ph

keywords machine learning force fieldsmultiscale modelingcoarse graininglong-range interactionsquantum mechanical interactionsmolecular dynamicshierarchical representations

0 comments

The pith

MuSE uses soft coarse-graining to let standard ML force fields capture quantum interactions across length scales.

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

The paper introduces a hierarchical model called MuSE that addresses the inability of most machine learning force fields to handle emergent interactions spanning multiple scales. It constructs coarse representations through soft fractional assignments of atoms to higher-level nodes, allowing existing MLFF architectures to process information at different resolutions simultaneously. Demonstrations on Hessian benchmarks, biomolecular folding paths, and nanostructure energy profiles show that this approach recovers accurate quantum-mechanical energies and forces where other long-range models fall short. A sympathetic reader would care because many physical and chemical systems exhibit precisely these cross-scale effects, limiting the reliability of current predictive simulations.

Core claim

MuSE employs Soft Coarse-Graining Pooling to build hierarchical representations from smooth fractional atom-to-node assignments, enabling architecture-agnostic coupling with MLFF modules such as SO3krates, MACE, and PaiNN so that they can operate across multiple scales and capture quantum-mechanical interactions that message-passing layers alone miss.

What carries the argument

Soft Coarse-Graining Pooling, the operation that maps atoms to coarse nodes via continuous fractional assignments and thereby supplies multiscale inputs to downstream ML force field modules.

If this is right

Existing MLFF architectures can be extended to model long-range many-body effects without redesigning their core layers.
Folding trajectories of biomolecules become feasible with near-quantum accuracy at multiple scales.
Energy profiles of molecule-graphene nanostructures can be computed reliably across relevant length scales.
Hessian-based benchmarks for multiscale systems show improved agreement with quantum references.

Where Pith is reading between the lines

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

The same pooling construction could be tested on systems with explicit time-scale separation, such as reactive processes coupled to slow conformational changes.
Because the method is architecture-agnostic, it may allow direct comparison of how different base MLFFs respond to the same multiscale input representations.
If the fractional assignments prove stable under modest perturbations, MuSE could serve as a drop-in module for existing simulation codes without requiring new training data at every scale.

Load-bearing premise

Soft fractional assignments of atoms to coarse nodes preserve enough quantum-mechanical detail that later modules can recover accurate energies and forces without uncontrolled errors.

What would settle it

A test system dominated by long-range many-body quantum effects in which MuSE-computed forces or energies deviate from high-level quantum reference calculations by more than the reported error bars while conventional long-range models remain closer.

read the original abstract

Realistic physical systems are characterised by emergent interactions across multiple length and time scales, posing a significant challenge for predictive machine learning (ML) models. Most scientific ML models focus on a narrow range of interactions. While machine learning force fields (MLFFs) offer near-quantum accuracy, the ubiquitous message-passing layers miss long-range many-body effects. Here we introduce the Multiscale Structural Ensemble (MuSE), a hierarchical model that uses Soft Coarse-Graining Pooling to construct coarse representations from smooth fractional assignments of atoms to coarse nodes, enabling MLFF modules to operate across multiple scales. MuSE is architecture-agnostic and coupled with SO3krates, MACE, and PaiNN MLFFs for both molecules and materials. We demonstrate the power of MuSE through Hessian-based benchmarks, folding trajectories for biomolecules, and energy profiles in molecule-graphene nanostructures, where MuSE accurately captures quantum-mechanical interactions at relevant scales -- unlike other recent long-range ML models.

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.

Desk Editor's Note private letter to a colleague

MuSE adds soft coarse-graining pooling to standard MLFFs for multiscale work, but the abstract gives no numbers, baselines, or implementation details so the superiority claims cannot be checked.

read the letter

The one or two things to know are that the paper introduces MuSE, which inserts a soft coarse-graining pooling step using fractional atom-to-node assignments so that existing MLFF modules can run at multiple scales, and that the abstract claims this works better than other recent long-range models on Hessian, folding, and nanostructure tasks.

What is actually new is the soft pooling construction itself. The model is kept architecture-agnostic and is shown coupled to SO3krates, MACE, and PaiNN on both molecules and materials. That combination and the fractional assignment approach look like a distinct technical step from the long-range extensions referenced in the abstract.

The soft spots sit in the evidence. The abstract states that MuSE accurately captures quantum-mechanical interactions at relevant scales, yet it supplies no error values, no baseline comparisons, no error bars, and no description of the pooling equations or training details. Without those, the performance claims stay unverified. The central assumption—that smooth fractional assignments preserve enough quantum information for the downstream modules without uncontrolled artifacts—remains plausible but untested in the given text.

This paper is aimed at people already working on machine learning force fields who need to move beyond single-scale limits. A reader focused on hierarchical or long-range modeling could extract the pooling idea for their own experiments, but anyone wanting to judge whether the method delivers on its claims will need the full methods and results sections.

I would send it to peer review. The target problem is real and the proposed mechanism is a reasonable extension worth checking, provided the full paper supplies the quantitative comparisons and implementation details that the abstract omits.

Referee Report

2 major / 0 minor

Summary. The paper introduces the Multiscale Structural Ensemble (MuSE), a hierarchical model using Soft Coarse-Graining Pooling to construct coarse representations from smooth fractional assignments of atoms to coarse nodes. This enables existing MLFF modules (coupled here with SO3krates, MACE, and PaiNN) to operate across multiple scales in an architecture-agnostic manner. The approach is demonstrated on Hessian-based benchmarks, folding trajectories for biomolecules, and energy profiles in molecule-graphene nanostructures, with the claim that MuSE accurately captures quantum-mechanical interactions at relevant scales unlike other recent long-range ML models.

Significance. If the central claims are substantiated with quantitative evidence, MuSE could address a recognized limitation of message-passing MLFFs in capturing long-range many-body effects while remaining compatible with established architectures. The use of soft fractional assignments for coarse-graining is a potentially generalizable idea for multiscale modeling in chemistry and materials science.

major comments (2)

[Abstract] Abstract: The abstract asserts superior performance on Hessian, folding, and nanostructure benchmarks but supplies no numerical results, error bars, baseline comparisons, or exclusion criteria. Without these, the central claim that MuSE 'accurately captures quantum-mechanical interactions at relevant scales' cannot be evaluated or compared to other long-range ML models.
The manuscript provides no equations, definitions, or ablation studies for the Soft Coarse-Graining Pooling operation or the preservation of quantum-mechanical information under soft fractional assignments. This leaves the weakest assumption (that such assignments introduce no uncontrolled artifacts for downstream MLFF modules) untested in the provided text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below and indicate where revisions will be made to strengthen the presentation of our results.

read point-by-point responses

Referee: [Abstract] Abstract: The abstract asserts superior performance on Hessian, folding, and nanostructure benchmarks but supplies no numerical results, error bars, baseline comparisons, or exclusion criteria. Without these, the central claim that MuSE 'accurately captures quantum-mechanical interactions at relevant scales' cannot be evaluated or compared to other long-range ML models.

Authors: We agree that the abstract would benefit from quantitative highlights to support the claims. In the revised manuscript we will insert specific metrics (e.g., MAE reductions on Hessian eigenvalues relative to MACE and PaiNN baselines, Pearson correlations for folding trajectories, and energy-profile RMSEs for the graphene systems) together with a brief statement of the main baselines and dataset filters. The complete numerical results, error bars, and exclusion criteria are already reported in Sections 3.1–3.3 and the Supplementary Information; the abstract revision will simply make these accessible at a glance. revision: yes
Referee: The manuscript provides no equations, definitions, or ablation studies for the Soft Coarse-Graining Pooling operation or the preservation of quantum-mechanical information under soft fractional assignments. This leaves the weakest assumption (that such assignments introduce no uncontrolled artifacts for downstream MLFF modules) untested in the provided text.

Authors: The Soft Coarse-Graining Pooling operation and the soft fractional assignment matrix are defined mathematically in Section 2.2 (Eq. 2) and the equivariance proof appears in Appendix A. Nevertheless, we accept that a dedicated ablation study is missing. We will add a new subsection (and corresponding figure) that compares soft versus hard assignments on the biomolecule folding and nanostructure benchmarks, reporting the resulting differences in predicted energies and forces. This will directly quantify any artifacts introduced by the soft assignments. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and context describe MuSE as a hierarchical model using Soft Coarse-Graining Pooling with fractional assignments, coupled to existing MLFFs like SO3krates, MACE, and PaiNN. No equations, fitted parameters, predictions, or self-citations are presented that reduce any claimed result to an input by construction. The central claim of improved multiscale capture is stated as an empirical demonstration via benchmarks, without any self-definitional, fitted-input, or uniqueness-imported steps visible. Full manuscript equations are referenced but unavailable here, precluding any load-bearing circularity identification per the rules requiring explicit quotes and reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Only the abstract is available; the ledger is therefore limited to the high-level modeling choice stated there.

axioms (1)

domain assumption Existing ML force-field architectures remain accurate when supplied with hierarchically coarsened atomic representations
The paper relies on this to claim that MuSE modules can be coupled to SO3krates, MACE, and PaiNN without loss of quantum fidelity.

invented entities (1)

Soft Coarse-Graining Pooling no independent evidence
purpose: To produce smooth fractional assignments of atoms to coarse nodes
New component introduced to enable the hierarchical model; no independent evidence supplied in the abstract.

pith-pipeline@v0.9.1-grok · 5730 in / 1181 out tokens · 33159 ms · 2026-06-29T19:39:36.670098+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references

[1]

https://pytorch.org/

PyTorch Foundation: PyTorch. https://pytorch.org/. Accessed: 2026-02-19 (2026)

2026
[2]

https://pyg.org/

PyG Team: PyTorch Geometric. https://pyg.org/. Accessed: 2026-02-19 (2026)

2026
[3]

Journal of Physics: Condensed Matter29(27), 273002 (2017)

Larsen, A.H., Mortensen, J.J., Blomqvist, J., Castelli, I.E., Christensen, R., Dułak, M., Friis, J., Groves, M.N., Hammer, B., Hargus, C.,et al.: The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter29(27), 273002 (2017)

2017
[4]

The Annals of Mathe- matical Statistics35(1), 73–101 (1964)

Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathe- matical Statistics35(1), 73–101 (1964)

1964
[5]

Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms268(11–12), 1818–1823 (2010)

Ziegler, J.F., Ziegler, M.D., Biersack, J.P.: Srim—the stopping and range of ions in matter (2010). Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms268(11–12), 1818–1823 (2010)

2010
[6]

Science Advances9(2), 0873 (2023)

Chmiela, S., Vassilev-Galindo, V., Unke, O.T., Kabylda, A., Sauceda, H.E., Tkatchenko, A., Müller, K.-R.: Accurate global machine learning force fields for molecules with hundreds of atoms. Science Advances9(2), 0873 (2023)

2023
[7]

Poltavsky, I., Charkin-Gorbulin, A., Puleva, M., Fonseca, G., Batatia, I., Brown- ing, N.J., Chmiela, S., Cui, M., Frank, J.T., Heinen, S., Huang, B., Käser, S., Kabylda, A., Khan, D., Müller, C., Price, A.J.A., Riedmiller, K., Töpfer, K., Ko, T.W., Meuwly, M., Rupp, M., Csányi, G., Lilienfeld, O.A., Margraf, J.T., Müller, K.-R., Tkatchenko, A.: Crash tes...

2023
[8]

Nature Communications15(1), 6539 (2024)

Frank, T., Unke, O., Müller, K.-R., Chmiela, S.: A euclidean transformer for fast and stable machine learned force fields. Nature Communications15(1), 6539 (2024)

2024
[9]

In: 2019 IEEE 31st International Conference on Tools with Artificial Intelligence (ICTAI), pp

Tan, T., Yin, S., Liu, K., Wan, M.: On the convergence speed of amsgrad and beyond. In: 2019 IEEE 31st International Conference on Tools with Artificial Intelligence (ICTAI), pp. 464–470 (2019)

2019
[10]

Science Advances10(14), 4397 (2024)

Unke, O.T., Stöhr, M., Ganscha, S., Unterthiner, T., Maennel, H., Kashu- bin, S., Ahlin, D., Gastegger, M., Medrano Sandonás, L., Berryman, J.T., Tkatchenko, A., Müller, K.-R.: Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments. Science Advances10(14), 4397 (2024)

2024
[11]

Nature Communications12(1), 7273 (2021)

Unke, O.T., Chmiela, S., Gastegger, M., Schütt, K.T., Sauceda, H.E., Müller, K.-R.: SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nature Communications12(1), 7273 (2021)

2021
[12]

https://github.com/jax-ml/jax (2018)

Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Katariya, Y., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs. https://github.com/jax-ml/jax (2018)

2018
[13]

https://github.com/ general-molecular-simulations/so3lr

general-molecular-simulations: SO3LR: SO3krates and Universal Pair- wise Force Field for Molecular Simulation. https://github.com/ general-molecular-simulations/so3lr. GitHub repository (2025) 7

2025

[1] [1]

https://pytorch.org/

PyTorch Foundation: PyTorch. https://pytorch.org/. Accessed: 2026-02-19 (2026)

2026

[2] [2]

https://pyg.org/

PyG Team: PyTorch Geometric. https://pyg.org/. Accessed: 2026-02-19 (2026)

2026

[3] [3]

Journal of Physics: Condensed Matter29(27), 273002 (2017)

Larsen, A.H., Mortensen, J.J., Blomqvist, J., Castelli, I.E., Christensen, R., Dułak, M., Friis, J., Groves, M.N., Hammer, B., Hargus, C.,et al.: The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter29(27), 273002 (2017)

2017

[4] [4]

The Annals of Mathe- matical Statistics35(1), 73–101 (1964)

Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathe- matical Statistics35(1), 73–101 (1964)

1964

[5] [5]

Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms268(11–12), 1818–1823 (2010)

Ziegler, J.F., Ziegler, M.D., Biersack, J.P.: Srim—the stopping and range of ions in matter (2010). Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms268(11–12), 1818–1823 (2010)

2010

[6] [6]

Science Advances9(2), 0873 (2023)

Chmiela, S., Vassilev-Galindo, V., Unke, O.T., Kabylda, A., Sauceda, H.E., Tkatchenko, A., Müller, K.-R.: Accurate global machine learning force fields for molecules with hundreds of atoms. Science Advances9(2), 0873 (2023)

2023

[7] [7]

Poltavsky, I., Charkin-Gorbulin, A., Puleva, M., Fonseca, G., Batatia, I., Brown- ing, N.J., Chmiela, S., Cui, M., Frank, J.T., Heinen, S., Huang, B., Käser, S., Kabylda, A., Khan, D., Müller, C., Price, A.J.A., Riedmiller, K., Töpfer, K., Ko, T.W., Meuwly, M., Rupp, M., Csányi, G., Lilienfeld, O.A., Margraf, J.T., Müller, K.-R., Tkatchenko, A.: Crash tes...

2023

[8] [8]

Nature Communications15(1), 6539 (2024)

Frank, T., Unke, O., Müller, K.-R., Chmiela, S.: A euclidean transformer for fast and stable machine learned force fields. Nature Communications15(1), 6539 (2024)

2024

[9] [9]

In: 2019 IEEE 31st International Conference on Tools with Artificial Intelligence (ICTAI), pp

Tan, T., Yin, S., Liu, K., Wan, M.: On the convergence speed of amsgrad and beyond. In: 2019 IEEE 31st International Conference on Tools with Artificial Intelligence (ICTAI), pp. 464–470 (2019)

2019

[10] [10]

Science Advances10(14), 4397 (2024)

Unke, O.T., Stöhr, M., Ganscha, S., Unterthiner, T., Maennel, H., Kashu- bin, S., Ahlin, D., Gastegger, M., Medrano Sandonás, L., Berryman, J.T., Tkatchenko, A., Müller, K.-R.: Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments. Science Advances10(14), 4397 (2024)

2024

[11] [11]

Nature Communications12(1), 7273 (2021)

Unke, O.T., Chmiela, S., Gastegger, M., Schütt, K.T., Sauceda, H.E., Müller, K.-R.: SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nature Communications12(1), 7273 (2021)

2021

[12] [12]

https://github.com/jax-ml/jax (2018)

Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Katariya, Y., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs. https://github.com/jax-ml/jax (2018)

2018

[13] [13]

https://github.com/ general-molecular-simulations/so3lr

general-molecular-simulations: SO3LR: SO3krates and Universal Pair- wise Force Field for Molecular Simulation. https://github.com/ general-molecular-simulations/so3lr. GitHub repository (2025) 7

2025