acceptodds
Under review as a conference paper at ICLR 2027

Wallpaper Net: A general Blueprint for Wallpaper-Group-Equivariant Graph Neural Networks

Abstract

Equivariant graph neural networks are typically designed for continuous Euclidean symmetry groups such as or , where arbitrary global translations are symmetries and can be removed using relative coordinates. In many settings, however, a graph is defined relative to an underlying periodic structure, so its symmetry is no longer the full Euclidean group. In two dimensions, these symmetries are the 17 wallpaper groups. This structure creates an opportunity: rather than explicitly representing the entire periodic system as a large graph, one can represent only the smaller graph of interest while incorporating the symmetry of the surrounding periodic structure into the model. We present a general blueprint for wallpaper-group-equivariant message passing by separating the symmetry problem into two components: point-group equivariance and lattice-relative translational symmetry. We demonstrate how it can be applied to a molecular adsorption problem on a periodic graphene substrate, providing high accuracy at a fraction of the cost of full-system message passing.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
GNNs, Equivariants, Wallpaper Groups, Groups, Geometric Deep Learning, Representation Theory, Atomistic Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8tX3OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes wallpaper-group-equivariant message passing to model graphs relative to a periodic two-dimensional background. It introduces we3nn, a library for finite-group-equivariant neural networks, and evaluates the framework on synthetic benchmarks covering all 17 wallpaper groups. The framework is further applied to interatomic potential modeling for molecules adsorbed on graphene, using adsorbate-only graphs without explicitly representing substrate atoms.

Critical strengths

Strong theoretical contribution: wallpaper-group equivariance, Wyckoff-anchor encoding, nonsymmorphic translation-suborbit handling. A well-motivated approach to incorporating substrate symmetry. Encoding the fixed substrate through symmetry-aware positional features addresses a meaningful limitation of molecular graphs that otherwise omit the substrate environment. The combination of anchor displacements and member labels provides a concrete mechanism for retaining substrate-relative information.

Critical weaknesses

Frankly speaking I am not so familiar with your research area. Only a little suggestion is that I think the presentation is not clear enough to follow. This paper may lack a single overview figure that let readers quickly follow the problem and motivation. Besides, the main experiments relies on Tables, with useful result visualization deferred to the Appendix. I think it is useful to visulize the results clearly in the main paper.

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Overview image request

Dear Reviewer,

Thank you for your feedback and for your openness about your specialization in a different area of representation learning. We see your expertise as an advantage for our paper. We have seen that researchers from different backgrounds are finding applications of geometric deep learning in their areas, such as combining it with LLMs or flow-matching models, or connecting it more closely to physics. So it is in our interest to make our paper as accessible as possible. We are already working on restructuring Section 3, and we think that introducing an overview figure is a great idea! If you don’t mind, we will run it by you before Sunday to make sure that it connects well with the theoretical part of the paper. We are planning to base it on one of the illustrations in Appendix C, showing two graphs: the original and a transformed version, with their corresponding Wyckoff anchors. We will show which elements go into which building blocks described in the paper (Wyckoff-anchor features, translation-suborbit member labels, and interatomic distances).

Is there something in particular that in your opinion would help make paper more straight-forward?

AuthorsOpenReview ↗

Figure 1 updated

Dear Reviewer,

To improve the readability and accessibility of the paper, we have prepared a new overview figure (Figure 1) that visually summarizes the main building blocks of the WallpaperNet framework.

The figure illustrates:

  • (a) How wallpaper-group symmetries act on a molecular graph on a periodic substrate, including the transformation of Wyckoff-anchor (WA) vectors.
  • (b) Which geometric and node features are used for message construction.
  • (c–d) The two approaches to symmetry-preserving message passing: equivariant and invariant GNNs. For the equivariant approach, we illustrate how regular-representation channels are permuted under symmetry operations. For the invariant approach, we visualize an angular invariant with a periodicity of 60 and how invariant coefficients are used to construct equivariant vector features.
  • (e) How WA translation-suborbit labeling works for nonsymmorphic wallpaper groups, illustrated using glide reflections in pgg.

We would greatly appreciate your feedback on whether this new figure improves the accessibility of the paper and provides a more intuitive introduction to the concepts and building blocks presented in Section 3.

Additionally, we have added two illustrations related to our molecular adsorption experiments on graphene. Figure 4 (page 62) provides an atomistic visualization of the system depicted schematically in Figure 1(a), while Figure 5 (page 63) illustrates why the position of a molecular graph relative to the underlying graphene lattice is important for these simulations.

We have not yet revised the main text , but we are currently working on improving. Given the short rebuttal period, we wanted to share these new figures with you first and ask whether you think they are a step in the right direction.

AuthorsOpenReview ↗

Improving the presentation

Dear Reviewer,

Based on your feedback, we modified the main text of the paper to make it easier to follow, The text highlighted in blue contains revisions aimed at improving the clarity, readability, and straightforwardness of the presentation. We additionally revised the main figure.

We hope these revisions address your main concern. We still have a little bit of time, before the rebuttal period ends, so if there are any additional suggestions from your side, we have still time to add them.

Looking forward to any additional feedback!

Thanks for your rebuttal,I am satisfied with your effort and increase my confidence to 3.

Reviewer_M3KkOpenReview ↗
Rating
3
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WallpaperNet, a general framework for wallpaper-group-equivariant GNNs on graphs coupled to fixed periodic environments. The main idea is to avoid explicitly modeling the entire periodic substrate by using Wyckoff-anchor features to encode the relative position of graph nodes within the lattice, while maintaining the appropriate equivariance properties. The paper also develops we3nn, a finite-group-equivariant neural network library, and provides a theoretical justification for the proposed construction.

The authors evaluate the framework on a synthetic atomistic benchmark covering all 17 wallpaper groups and on a molecular adsorption task involving graphene. The experiments show that WallpaperNet can achieve comparable or better energy and force prediction accuracy than the tested full-cell GNN baselines, while using substantially smaller graphs and less GPU memory.

Critical strengths
  1. insightful problem formulation and technical design.
  2. convincing experimental results.
  3. code released anonymously.
Critical weaknesses
  1. Presentation and motivation could be improved: I find some parts of the paper relatively hard to follow, especially Section 3. There are long chunks of mathematical definitions, and it takes some effort to understand how each part contributes to the overall framework and eventually Theorem 1, and mostly sometimes why I need to pay attention to certain part of the theories. It is also slightly abrupt to end section 3 without explaining how this relates to the design of architectures and experiments and in my opinions they deserve to be in the main body of the paper instead of hiding somewhere in the appendix, although this is partly addressed by directly reading the code.

  2. Ablations: for table 1 and table 2, I think the full-cell invariant GNN and e3nn baselines are reasonable. However, I wonder whether comparing against a simpler compact alternative would help better understand the contribution of the proposed architecture. For example, a molecule-only GNN using local substrate information or periodic positional features could be a simpler comparison. This might help clarify how much of the improvement comes from reducing the graph size simply versus the proposed wallpaper-equivariant representations. I understand that such alternatives may require additional implementation effort and are not necessarily standard off-the-shelf baselines for this setting.

Questions for authors
  1. ablation: I am curious whether the authors have considered comparing against a simpler compact GNN baseline, such as a molecule-only geometric GNN with periodic positional features or a small local substrate environment. The current full-cell baselines seem reasonable, but such a comparison might help clarify the benefit of the wallpaper-specific architecture.
2 comments
AuthorsOpenReview ↗

Section 3 changes: the initial plan

Dear Reviewer,

We are thankful for your feedback. Indeed, the motivation and organization of the section 3 can be improved, and we plan to use the additional page available during the discussion period to make the presentation more accessible. We want to fight it on two fronts: by introducing a better overview figure and restructuring Section 3. We would begin with the translational problem: why relative edge vectors cannot distinguish different registries, and what information the Wyckoff-anchor features supply. We would then introduce the neural-network operations that process these features, explaining the purpose of each component before its mathematical definition and making its connection to the new overview figure. We are also planning to draw clearer connections (in that section) to familiar geometric GNNs, so that the potential reader can be guided through the manuscript by comparison with already established ideas. This restructuring will take a bit of time, we are planning to be done by Monday the 16th.

We would like your guidance on the matter. Do you think that these changes will improve the readability of the paper? Is there something in particular in that section that would benefit from a more detailed or simpler explanation?

AuthorsOpenReview ↗

Ablation: Simpler GNN with local information

In the meantime, we would like to showcase the experiment that we recently did based on your feedback. The question of how much of the accuracy improvement comes from the fact that we operate on a smaller graphs is very intriguing. First of all, we were not expecting to see any improvement with WPNet: we aimed to match the full-cell baseline at a lower cost. Therefore, the size difference improvement is a good hypothesis to test. To address this, we implemented a Local-surface EGNN-like baseline that performs message passing only on the adsorbate graph. Before message passing, it constructs a descriptor for every atom of small graph, by sum-aggregating features derived from its distances to all nearby surface atoms (within radius of 6 Å) and their atomic types. We trained the model on the molecular adsorption on graphene dataset using the same training protocol as the existing models. We will include the architecture, training details, and individual-run results in the supplementary material with the planned revision on Monday.

Interestingly, the Local-surface EGNN achieves a mean direct-force component MAE of (60.1 ± 4.3) meV/Å, compared with (83.6 ± 12.2) meV/Å for the full-cell invariant GNN. However, this improvement is specific to direct-force prediction: its energy and conservative-force errors are higher than those of the full-cell invariant GNN.

WP-O3 (our best performing model) retains a substantial advantage over this compact baseline. Its mean direct-force component MAE is (31.3 ± 0.2) meV/Å, approximately 48% lower than that of the Local-surface EGNN. For conservative forces, the corresponding errors are (32.9 ± 0.9) and (132.7 ± 3.2) meV/Å. This shows that a reduction in graph size can contribute to the improvement, but is not completely responsible for it.

We plan to summarize this comparison briefly in the main text and provide the full discussion in the supplementary material. The extended table below includes the new baseline alongside the existing results for a better overview. (It's Table 2 in the current version of the manuscript)

Test-set performance comparison

Energy MAE (E. MAE) is reported in meV/atom, force-component MAE (F. MAE) in meV/A, and relative force L2 (F. L2) in %. C denotes force-weighted conservative training (forces derived from energy), while V denotes direct vector-head force prediction. Values are mean ± sample standard deviation over three random seeds. Lower is better for error metrics; higher is better for R2R^2.

ModelE. MAE ↓F. L2 C ↓F. L2 V ↓F. MAE C ↓F. MAE V ↓R2R^2 C ↑R2R^2 V ↑
Wallpaper-symmetry models
WP-Inv1.73 ± 0.0280.79 ± 1.3541.35 ± 7.79139.9 ± 0.675.0 ± 16.30.35 ± 0.020.82 ± 0.07
WP-InvEq4.75 ± 5.0978.68 ± 1.1525.67 ± 0.63137.1 ± 2.043.5 ± 1.00.38 ± 0.020.93 ± 0.00
WP-Equi1.90 ± 0.1780.59 ± 0.9427.95 ± 0.80140.7 ± 1.448.8 ± 0.90.35 ± 0.020.92 ± 0.00
WP-VecInv2.24 ± 0.0662.49 ± 15.9526.62 ± 0.93109.5 ± 25.444.8 ± 0.80.59 ± 0.210.93 ± 0.01
WP-Reg2.09 ± 0.1230.70 ± 1.9027.15 ± 1.0754.0 ± 2.745.7 ± 1.80.91 ± 0.010.93 ± 0.01
WP-RegTP1.86 ± 0.2430.49 ± 1.3326.21 ± 1.1452.7 ± 2.344.6 ± 1.10.91 ± 0.010.93 ± 0.01
WP-O31.80 ± 0.0819.78 ± 0.7019.22 ± 0.7532.9 ± 0.931.3 ± 0.20.96 ± 0.000.96 ± 0.00
Full-cell baselines
Invariant GNN2.58 ± 0.2367.51 ± 1.0347.22 ± 6.46119.9 ± 2.583.6 ± 12.20.54 ± 0.010.77 ± 0.06
e3nn (l_max=1)1.92 ± 0.0725.89 ± 0.8925.31 ± 3.5044.7 ± 2.143.9 ± 7.20.93 ± 0.000.94 ± 0.02
e3nn (l_max=2)2.04 ± 0.2221.43 ± 0.7525.55 ± 5.8236.2 ± 1.245.0 ± 11.10.95 ± 0.000.93 ± 0.03
Local-surface baseline (new experiment)
Local-surface EGNN2.94 ± 0.0776.39 ± 1.8733.78 ± 2.34132.7 ± 3.260.1 ± 4.30.42 ± 0.030.89 ± 0.02

We would appreciate your view on whether this additional comparison helps clarify the role of the compact graph representation.

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

This paper proposes a graph neural network for molecules interacting with a fixed periodic surface. It encodes the molecule’s position relative to the surface without explicitly including all surface atoms in message passing. The method is evaluated on wallpaper-group benchmarks and molecular adsorption on graphene.

Critical strengths
  1. The representation captures information about position relative to the surface that molecular coordinates alone would miss.

  2. The experiments cover all 17 wallpaper groups and include a more realistic graphene application.

Critical weaknesses
  1. The nearest-anchor representation can change discontinuously when an atom crosses an anchor boundary. It is unclear whether the predicted energy and forces remain continuous.

  2. Some benchmark labels only approximately satisfy the assumed periodic symmetry. The magnitude of this mismatch is not quantified.

Questions for authors
  1. What happens to predicted energies and forces when an atom crosses a nearest-anchor boundary?

  2. How large are the symmetry violations in the reference labels compared with the model’s prediction errors?

3 comments
AuthorsOpenReview ↗

Approximate periodic symmetry investigation

Dear Reviewer,

We agree on both points and thank you for your suggestions. We believe that answering both questions will lead to a stronger paper.

In the first experiment (Table 1 in the current version of the paper) we benchmark our models on approximately symmetric systems, without the periodic boundary conditions. This design was based on set-up that is common in computational simulations, and we agree that quantification of the symmetry violation is necessary. We have therefore performed a direct symmetry-consistency audit of the finite-cluster xTB labels.

For each wallpaper group, we selected one existing molecular configuration containing 3, 4, and 5 atoms. We rigidly transformed each molecule by every symmetry representative used in the unit-cell construction, while keeping the explicit (3x3) support cluster unchanged. A single lattice translation, applied to the entire molecule, returned its center of mass to the central unit cell. We measured the absolute energy difference relative to the identity calculation, normalized by the total number of support and molecule atoms, and compared the molecular forces with the covariantly transformed identity forces.

Across the 16 groups(excluding p1) with non-identity symmetry representatives, the mean discrepancies are 1.28 meV / atom for energy, 66.3 meV / Å) for force components, and 2.53% for relative force (L2L_2). To assess their scale, we compare them with the lowest model error of the averaged over each group:

MetricMean symmetry discrepancyMean of lowest model errorsDiscrepancy / model error
Energy MAE (meV / atom)1.34.528.4%
Force-component MAE, combined training (meV / Å)66.3412.416%
Force-component MAE, vector-head training (meV / Å)66.3250.726%
Relative force L2L_2, combined training (%)2.518.314%
Relative force L2L_2, vector-head training (%)2.511.921%

Overall, we observed that label miss-match due to the symmetry violation is under the model error for every best performing model (the lowest error) consistency across all groups.

We will add an extended discussion on the symmetry violation of the synthetic benchmark in the report, with dedicated tables for every group. We hope that this experiment addressed your concern, and if you have any additional suggestion, we are open for your advice in the additional study of this problem.

AuthorsOpenReview ↗

The nearest-anchor representation discontinuity

The proposed representation is indeed discontinuous when crossing the anchor boundaries. To quantify the sensitivity to anchor reassignment of already trained models, we performed a controlled boundary-swap experiment on 100 randomly selected test structures. For each structure, we placed the molecular atom closest to boundary between two Wyckoff-anchors exactly on that boundary. We compared predictions using its original and the second-closest anchors, keeping atomic coordinates, all other WA vectors, and graph edges identical. We perform this experiment on the graphene-molecule dataset, since it's the one that mimics a realistic application scenario. (Table 2 in the original manuscript)

The table reports differences between these two predictions. Energy swap MAE is normalized by the number of full-cell atoms. Force-component swap MAE and relative force L2L_2 are evaluated over all molecular atoms, over the 100 structures. C denotes force-weighted energy-derived force training, and V denotes direct vector-head training.

ModelEnergy swap MAE (meV / Atom)Rel. Force L2L_2, C (%)Rel. Force L2L_2, V (%)Force-component MAE, C (meV / Å)Force-component swap MAE, V (meV / Å)
WP-Inv0.353 ± 0.0337.57 ± 1.263.25 ± 0.593.61 ± 0.882.29 ± 0.31
WP-InvEq0.247 ± 0.1106.15 ± 0.483.73 ± 0.272.99 ± 0.213.48 ± 0.03
WP-Equi0.342 ± 0.0317.02 ± 2.504.58 ± 0.354.29 ± 0.943.78 ± 0.38
WP-VecInv0.653 ± 0.0909.74 ± 1.066.96 ± 0.199.86 ± 0.829.20 ± 0.19
WP-Reg0.454 ± 0.0911.33 ± 0.252.05 ± 0.181.84 ± 0.502.00 ± 0.08
WP-RegTP0.492 ± 0.0653.28 ± 0.172.66 ± 0.134.41 ± 0.172.72 ± 0.15
WP-O30.426 ± 0.0781.24 ± 0.580.98 ± 0.531.48 ± 0.331.49 ± 0.77

To put these differences in context, we compare their magnitude with each model’s reported test MAE. Across the seven wallpaper models, the force-component swap differences amount to 2.2–9.0% of the corresponding test force-component MAE in conservative mode and 3.1–20.5% in direct-head mode. The per-atom energy error for the anchor-swap-experiment corresponds to the interval between 5.2-29.1% of the model respective error. We also observe some tendency for larger test-force errors to accompany larger force changes under anchor reassignment, most clearly for the conservative models, although this is not universal. We hypothesize that this error can be reduced, if during training, for atoms placed on the boundary, the WAs position would randomly swap (as a form of data-augmentation).

We will add a detailed report and tables for all 3 seeds evaluation of this experiment with a short discussion of the trends in the updated manuscript of the paper. We would appreciate your feedback on whether this additional boundary-swap experiment addresses your concern about changes in the predicted energies and forces at nearest-anchor boundaries and we are open to suggestions.

Dear Reviewer, we've updated the manuscript in the Appendix with your proposed experiments. (Highlighted in purple). Pages 29 and 65!

We hope that we addresses all of your concerns.

Comments on the paper

8 comments

Thank you for raising this question. For context, AI4Mat-NeurIPS 2025 is a non-archival workshop without official proceedings. We will consult the AC on how best to clarify the relationship between the manuscripts while respecting the ICLR policies. We would like to add that we are confident in the novelty of the scientific contributions of our submission and this situation is quite common in our domain. While we await the AC guidance, we would politely ask to assess the manuscript as an original research contribution, without discounting its novelty on the basis of the non-archival workshop report.

AuthorsOpenReview ↗

Figure 1 updated

Dear Reviewer,

We really appreciate your fast feedback! We are working towards improving the accessibility of the paper. First, we would like to show you our updated Figure 1 to make sure that this new visualization is an improvement. The idea here was to visualize the concepts from Section 3 that are possible to visualize (angular invariants, Wyckoff-anchor transformations, and regular representations), while also showing how they fit into the overall design, with Figure 1(a) being essentially a merge of the previous Figures 1(a) and 1(b) into one concept. Then we show how the nodes, edges, and vectors from that picture becoming part of the message-passing scheme. We have not modified the main text yet, and we will also make sure to provide the context of the paper that you mentioned.

In case the connection of the framework to a real problem was not straightforward, we have added two illustrations related to our molecular adsorption experiments on graphene. Figure 4 (page 62) provides an atomistic visualization of the system depicted schematically in Figure 1(a), while Figure 5 (page 63) illustrates why the position of a molecular graph relative to the underlying graphene lattice is important for these simulations.

We are working on the Section 3 text now, so any feedback on the images before the end of the rebuttal period would be highly appreciated.

Dear Reviewer,

We have revised the main text of the manuscript and highlighted the changes. The paragraph highlighted in green provides additional motivation from the application domain and emphasizes the importance of the problem. The text highlighted in blue contains revisions aimed at improving the clarity, readability, and straightforwardness of the presentation.

We hope these revisions address your main concern and make the manuscript easier to follow. If you feel that further improvements are needed, please do not hesitate to let us know. We would greatly appreciate any additional suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.