acceptodds
Withdrawn from ICLR 2027

Graph Reach Sets the Remote-Response Limit of Local Interatomic Potentials

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Force accuracy cannot certify physical fidelity when a response depends on atoms beyond a model's graph reach. For finite-range local-energy message-passing models, we prove an exact architectural law relating interaction depth L to source-target graph distance d_G. When 2L - d_G < 0, attribute-to-force and coordinate force-response cross-derivatives vanish for every parameter setting, so training cannot recover the missing response. In Au2/MgO, increasing MACE depth from two to three blocks crosses this boundary, improves held-out DFT doped-undoped contrast, and activates aligned Al-to-Mg sensitivity. By contrast, a parameter-matched width increase leaves graph reach unchanged and does not recover the response. Increasing cutoff closes 57.0% of the predictive gap (95% interval, 55.2-58.8%), while a SchNet-style model reproduces the depth-width ordering. Ag-Pd exposes the complementary failure: graph-permitted models retain the wrong response direction or scale despite competitive force errors. The graph-reach law therefore provides a pre-training test of representability and separates architectural impossibility from failures of supervision or optimization.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
interatomic potentials, message-passing neural networks, graph reach, long-range response, MACE, atomistic machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_wxrdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the architectural limitations of MLIPs in representing remote atomic responses. Specifically, for an L-layer model, the force response between atoms separated by more than 2L graph hops is identically zero. The authors validate this finding by demonstrating that model depth affects force responses in the Au₂/MgO system. Furthermore, even when the architecture has sufficient expressive capacity, it does not necessarily produce physically accurate responses. Overall, this paper examines the influence of receptive fields on MLIPs, particularly in the context of remote atomic interactions.

Critical strengths
  1. Clear theoretical formulation. The theoretical framework is well formulated, particularly in distinguishing the general concept of receptive fields from the actual interatomic interactions that a model can represent.
  2. Solid experimental analysis. The experimental analysis is convincing, and the conclusions are clear and well supported.
Critical weaknesses
  1. I have some reservations about the overall contribution of this paper. Receptive fields have long been studied in MLIPs, and numerous existing works have introduced additional architectures or dedicated branches to capture long-range interactions. Moreover, in many systems, such as small molecules and relatively small crystals, the receptive field is of limited relevance, as the atomic graph may even be fully connected.
  2. Insufficient evaluation of practical significance. Although the experiments clearly demonstrate the 2L-hop threshold effect, the practical significance of this remote response has not been sufficiently evaluated. For example, how does the inability to capture such long-range responses affect the accuracy and reliability of actual atomistic simulations?
Questions for authors

Could the authors clarify the practical significance of the identified graph-reach limitations for actual atomistic simulations? To what extent do these missing remote responses affect simulation accuracy and reliability? Are existing approaches designed to capture long-range interactions in MLIPs sufficient to address the limitations identified in this paper? If so, what additional insights or practical benefits does the proposed analysis offer?

Reviewer_qhm3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the representation limitation of finite-range local-energy message-passing models for predicting remote atomic responses. It identifies an architectural limitation that cannot be overcome through additional training or model width. The authors study this boundary with two systems, and made some very interesting observations: 1) Increasing model width does not recover the missing response; 2) increasing cutoff provides another means of extending graph reach; 3) architectural reachability alone does not guarantee correct response prediction.

Critical strengths
  1. This paper identifies a very meaningful theoretical result. Authors identified an architectural limitation that cannot be overcome through additional training or model width. This provides a useful distinction between representational impossibility and optimization failure.

  2. Not only theory but the paper also shows empirical support. The paper carefully studies the depth, width, and cutoff comparisons. And the parameter-matched width control helps separate increased capacity from increased receptive-field support.

  3. I found this paper very valuable to the ai for science community in general. It highlights why low average force error may be insufficient for evaluating interatomic potentials when downstream scientific applications depend on specific nonlocal responses.

Critical weaknesses

1.For the theory part, because the graph reach bound minly applies to finite-range local-energy message-passing models, it does not directly apply to the archtectures includes global interaction, charge equilibration or other nonlocal machanisms.

  1. It is not a new discovery that GNN suffers from the problem that it cannot propage information beyond its receptive field. It would be great if the authors can furthur clarify the unique contribution in this paper, which mainly is its application to force-response derivatives and physically meaningful interventions.

  2. So far the empirical results is mainly with two systems, and it remains unclear how broadly the finidings translate to other chemical environments, materials and respnse types.

Questions for authors

It would be great if the authors could answer the questions brought in the weakness, and address each concern separately. I will consider raising the scores if those get properly addressed.

Reviewer_arySOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a graph-reach criterion for identifying structurally excluded remote force responses in local-energy message-passing interatomic potentials. It shows that the relevant cross-response vanishes beyond graph hops under the stated assumptions. Au₂/MgO experiments compare depth, width, and cutoff changes, with an additional SchNet-style validation, while Ag–Pd experiments demonstrate that permitted reach does not guarantee accurate response. The intended contribution is a diagnostic for screening architectures before expensive training.

Critical strengths
  • Architecture search for MLIPs and their targeting domain is a community-wide beneficial subject. This paper aims to resolve the problem preceding heavy model training, and the proposed methods are simple.
Critical weaknesses
  • The proposed methods do not provide significant insight or methodological decisiveness; rather, they are close to general insight that MLIP researchers already have about locality coverage related to GNN. While mathematically valid, the manuscript does not establish why this constitutes a substantive theoretical advance rather than a useful restatement of receptive-field limitations.
  • The empirical scope does not establish the broader significance of the proposed diagnostic. The experiments focus on depth–width comparisons in (\mathrm{Au}_2/\mathrm{MgO}), cutoff expansion in the same system, and a SchNet-style repetition of the same task. These only provide limited evidence across distinct physical problems. The (\mathrm{Ag}\text{–}\mathrm{Pd}) experiment illustrates that permitted reach does not ensure accurate response, without establishing how the criterion guides successful architecture selection. Overall, the experiments support the selected observations but do not sufficiently demonstrate broader practicality.
  • The paper presentation does not sufficiently support its contribution in theoretical novelty or empirical generality. Figure 1 largely restates the proposed workflow without explaining what additional modeling decisions it enables. Several figures illustrate the support argument or revisit the same (\mathrm{Au}_2/\mathrm{MgO}) result, while the distinction between group-contrast prediction, model-only sensitivity, and paired physical response requires careful reconstruction from the text. A consolidated presentation connecting each claim to its evidence, assumptions, and demonstrated scope is missing. This makes the central contribution difficult to assess beyond the individual reported results.
Questions for authors
  • What substantive theoretical insight does the (2L) criterion add beyond existing locality and receptive-field arguments?
  • Which reported result demonstrates that this criterion enables a useful architecture-selection decision beyond the selected (\mathrm{Au}_2/\mathrm{MgO}) setting?
  • Which physical-response claim is supported by each evaluation: the unpaired (\mathrm{Au}_2/\mathrm{MgO}) group contrast, the model-only identity intervention, and the paired (\mathrm{Ag}\text{–}\mathrm{Pd}) force constants? Clarifying this mapping would help establish the precise scope of the contribution.
Reviewer_BKmTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper concerns two failure modes of local machine-learned interatomic potentials (MLIPs). The first failure mode is an architectural one: for a model that represents the total energy as a sum of learned local energies, if the source and target sites of interest are dGd_G hops apart, an LL-layer model can only have a non-vanishing response to perturbations of the source (spatial or chemical) if m=2L−dG≥0m = 2L − dG \ge 0. This is demonstrated by considering a slab of MgO, with an Au$_2$ dimer on the surface, in which the response to transmutation of subsurface atoms is investigated. The authors demonstrate that a model with m≥0m \ge 0 significantly improves performance across all metrics compared with models with m<0m < 0, even if the number of parameters is kept fixed. Nevertheless, m≥0m \ge 0 (combined with a good force accuracy) is not sufficient to guarantee a correct response. An example is shown in Ag-Pd calculations, in which graph-permitted models predict the wrong response.

Critical strengths

The author provides a very detailed and systematic review of calculations that test the importance of the 2L−g2L - g inequality for calculating physical responses. This includes cases where mm being non-negative is key, and cases where it is not enough to reproduce a correct response. I expect this to be useful for some readers.

Critical weaknesses

It seems like the main conclusion from this paper is the 2L−dG2L - d_G indicator. However, this is already a well-known result, although usually phrased as the range of force responses being 2rL2rL. While the authors do discuss a slightly different (and stronger) statement, very similar discussions can be found in the literature. For example,

it can be shown that Eq. (1) results in an effective environment-dependence of the forces corresponding up to twice the environment radius defining the atomic energies in MLPs \cite{herbold2022hessian} (cited by authors)

Our approach is based on the insight that, as detailed below, the analytic atomic forces corresponding to the total energy expression eqn (1) depend on up to twice the cutoff radius \cite{herbold2023machine}

Due to the finite receptive field of GNN models, Hessian matrix elements between atoms that are farther apart than twice the receptive field radius (which is given by the atomic environment cutoff multiplied by the number of message-passing steps) are strictly zero \cite{gonnheimer2025beyond}

As a result, GNN-IPs require a broader region for communication, reaching up to $r_c$ multiplied by the number of message-passing steps \cite{park2024scalable}, More generally, in a message-passing implementation, backpropagation ensures that force predictions benefit from an effective receptive radius that is twice that of atom-centered energy predictions \cite{bigi2024dark}.

The bibtex for the above references can be found below:

@article{herbold2022hessian, title={A hessian-based assessment of atomic forces for training machine learning interatomic potentials}, author={Herbold, Marius and Behler, J{"o}rg}, journal={The Journal of Chemical Physics}, volume={156}, number={11}, year={2022}, publisher={AIP Publishing} }

@article{herbold2023machine, title={Machine learning transferable atomic forces for large systems from underconverged molecular fragments}, author={Herbold, Marius and Behler, J{"o}rg}, journal={Physical Chemistry Chemical Physics}, volume={25}, number={18}, pages={12979--12989}, year={2023}, publisher={The Royal Society of Chemistry} }

@article{gonnheimer2025beyond, title={Beyond numerical Hessians: higher-order derivatives for machine learning interatomic potentials via automatic differentiation}, author={Gönnheimer, Nils and Reuter, Karsten and Margraf, Johannes T}, journal={Journal of Chemical Theory and Computation}, volume={21}, number={9}, pages={4742--4752}, year={2025}, publisher={ACS Publications} }

@article{park2024scalable, title={Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations}, author={Park, Yutack and Kim, Jaesun and Hwang, Seungwoo and Han, Seungwu}, journal={Journal of chemical theory and computation}, volume={20}, number={11}, pages={4857--4868}, year={2024}, publisher={ACS Publications} }

@article{bigi2024dark, title={The dark side of the forces: assessing non-conservative force models for atomistic machine learning}, author={Bigi, Filippo and Langer, Marcel and Ceriotti, Michele}, journal={arXiv preprint arXiv:2412.11569}, year={2024} }

Questions for authors

Given the above concerns, could the authors clarify the novelty of the work presented in the paper?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.