acceptodds
Under review as a conference paper at ICLR 2027

CPGN: A Coordination Polyhedron Graph Network for Multi-Scale Crystal Property Prediction Using Dual-Level Atomic and Geometric Representations

Abstract

Accurate prediction of crystal properties remains a fundamental challenge in computational materials science. Although graph neural networks (GNNs) such as CGCNN, MEGNet, ALIGNN, and SchNet have achieved consistent performance, they represent crystals solely at the atomic level, relying on message passing to implicitly capture local environments. However, many material properties are governed by coordination polyhedra rather than individual atoms. To address this limitation, we propose the Coordination Polyhedron Graph Network (CPGN), a multi-scale GNN that jointly learns from atom-, bond-, and coordination-polyhedron-level representations. CPGN integrates an atom graph, a line graph for angular interactions, and a coordination polyhedron graph connected through shared polyhedral relationships. Polyhedra are described using physically meaningful geometric descriptors, while an interleaved message-passing framework with bidirectional cross-attention enables effective information exchange across structural scales. CPGN is evaluated on three benchmark datasets: Materials Project (69,239 structures), JARVIS-DFT (75,993 structures), and QM9 (130,831 molecules) using standardized splits for fair comparison. On Materials Project, CPGN achieves a formation energy MAE of 0.054 eV/atom and the band-gap MAE of 0.251 eV, outperforming CGCNN, SchNet, MEGNet, and ALIGNN. On JARVIS-DFT, CPGN provides accurate multi-task predictions, achieving MAEs of 0.0925 eV/atom for formation energy, 0.227 eV for optical band gap, 14.01 GPa for bulk modulus, 11.28 GPa for shear modulus, and 0.088 eV/atom for stability. On QM9, CPGN achieves a HOMO MAE of 0.030 eV and competitive performance for LUMO (0.023 eV) and HOMO–LUMO gap (0.038 eV), demonstrating its effectiveness across both crystalline and molecular systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multiscale graph neural networks, Materials Informatics, Crystal Property Prediction, Geometric Deep Learning, Multi-Task Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_e6fPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CPGN, a graph neural network for crystal (and molecular) property prediction that augments the standard atom graph + line graph formulation of ALIGNN with a third "coordination polyhedron graph" GPG_P. Polyhedra are extracted via Pymatgen's VoronoiNN, represented as nodes with 7 hand-crafted geometric/chemical descriptors (coordination number, bond-length distortion, bond-angle variance, volume proxy, mean bond length, electronegativity mismatch, coordination type), and connected by edges weighted by sharing type. The three graphs are processed by interleaved message passing, coupled by a "bidirectional cross-attention" in which each stream attends over the crystal-level mean of the other stream, followed by mean pooling, fusion, and multi-task heads (formation energy, band gap, stability classification). The paper reports results on MP 2018.6.1, JARVIS-DFT, QM9, with ablations framing CGCNN →\to ALIGNN →\to CPGN as progressively adding GLG_L and GPG_P, a scaling study over training-set sizes, inference/memory profiling, and a theoretical section.

Critical strengths

Writing is clear and easy to follow.

The evaluation benchmark follows mature baselines, and is well-covered on evaluation datasets.

Inductive bias is physically well-motivated.

Critical weaknesses

The headline SOTA claims rest on substantially under-tuned reproductions, and the paper's own appendix contradicts them. There is internal inconsistency on the single most important baseline number. See Tables 1, 11, 19 and the digits reported by ALIGNN's original article. Please check the implementation of ALIGNN thoroughly.

Given that your cross-attention has a single key, why does it need $0.326 * 4.85 $ M parameters? Have you compared it against (a) a small MLP injecting the pooled other-stream vector, and (b) per-atom-to-own-polyhedron attention?

The performance of this work is substantially outdated. All the compared methods are from 2018-2021. Please consider some up-to-date works, including PotNet (ICML23), ComFormer (ICLR24), Crystalformer (ICLR24), PDDFormer (IJCAI25), CrysLDNet (ICML26) and so on.

Questions for authors

See Weaknesses.

Reviewer_n6U9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CPGN as a multi-scale graph neural network for predicting crystal and molecular properties. It combines an atom graph, a line graph for bond-angle information, and a coordination-polyhedron graph that captures higher-level local structure. Bidirectional cross-attention and interleaved message passing are used to exchange information between the atomic and polyhedral representations.

The model is evaluated on Materials Project, JARVIS-DFT, and QM9, with strong results for formation energy, band gap, elastic properties, and related targets. Ablation studies show that adding line-graph information improves over atom-only models, while adding the polyhedron graph with cross-attention gives further gains, especially for band-gap prediction. The paper also provides a theoretical argument that CPGN can distinguish some structures that atom- and line-graph models such as ALIGNN cannot.

Critical strengths
  1. The strongest contribution is the explicit coordination-polyhedron graph, which complements atom-level and bond-angle information with higher-level geometric and topological structure. This is consequential because the gains are particularly strong for band-gap prediction, where coordination geometry is plausibly important.

  2. The method is evaluated across Materials Project, JARVIS-DFT, and QM9, and shows strong or competitive performance across several prediction tasks. The empirical results are further supported by a theoretical argument that CPGN can distinguish a class of structures that atom- and line-graph models such as ALIGNN cannot distinguish.

Critical weaknesses
  1. The empirical comparison does not include sufficiently strong recent baselines. The main and extended comparisons are dominated by older models such as CGCNN, MEGNet, SchNet, and ALIGNN, while highly relevant top-venue crystal models such as Matformer, PotNet, and ComFormer are not evaluated. On the same Materials Project benchmark setting, substantially lower formation-energy/band-gap MAEs than CPGN's have already been reported: by Matformer by PotNet, and by ComFormer [1, 2, 3]. Consequently, the evidence that CPGN is competitive with the current state of the art is substantially weakened, which is important because improved predictive accuracy is a central empirical claim.

  2. The ablation does not isolate the paper's central contribution. The coordination-polyhedron graph and bidirectional cross-attention are introduced together in the main ablation, so the reported improvement cannot be attributed specifically to the proposed polyhedral representation rather than cross-attention, additional parameters, or their combination. More diagnostic component-wise ablations have been provided in closely related top-venue work, where periodic encoding, potential-based interactions, infinite-range interactions, and geometric components were separately removed or controlled [1, 2, 3]. Since explicit coordination-polyhedron modeling is the main claimed novelty, the lack of a G_P only, cross-attention-only, and preferably parameter-matched comparison makes the key causal claim insufficiently validated.

References

[1] Periodic Graph Transformers for Crystal Material Property Prediction, NeurIPS'22

[2] Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction, ICML'23

[3] Complete and Efficient Graph Transformers for Crystal Material Property Prediction, ICLR'24

Questions for authors
  1. Could it be clarified whether the conclusions about empirical competitiveness would have changed if controlled comparisons with stronger recent models such as Matformer, PotNet, and ComFormer had been included under the same dataset splits and evaluation protocol?

  2. Could it be clarified whether the reported gains would still have been observed if the coordination-polyhedron graph and bidirectional cross-attention had been ablated separately, rather than introduced together?

Reviewer_xhYjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CPGN, a multi-graph neural network for crystal property prediction that extends the atom and line graphs of ALIGNN by introducing an additional coordination polyhedron graph. The coordination polyhedron graph is constructed using Voronoi tessellation and initialized with a set of precomputed, handcrafted features. The three representations interact through message passing and cross-attention mechanisms. The authors evaluate CPGN on the Materials Project, JARVIS-DFT, and QM9 datasets, demonstrating improvements over ALIGNN and other baseline models. Overall, the core idea and methodology of this paper are sound, intuitive, and easy to follow. Coordination polyhedra are indeed a reasonable structural scale to consider, as exemplified by metal clusters in MOFs. Although there is no fundamental reason to believe that conventional atom-based GNNs cannot implicitly learn such structural inductive biases, explicitly incorporating them may improve data efficiency, particularly in low-data regimes, and potentially yield more meaningful semantic representations. However, the evaluation of the proposed method is rather unconvincing, and the manuscript falls short of basic standards of scientific presentation. First, ALIGNN and DimeNet are already several years old. More recent models include, but are not limited to, GemNet, ViSNet, and MACE. Given the relatively limited methodological novelty of CPGN, and the fact that ALIGNN's dual-graph architecture has not been widely adopted by mainstream models, I am skeptical that CPGN can achieve state-of-the-art performance, although its core idea of introducing a polyhedron graph could potentially be incorporated into more recent architectures. Second, the ablation studies are incomplete. The authors treat cross-attention, message passing, and the additional polyhedron graph as a single component in their ablation experiments, making it difficult to assess the individual contributions of these components. Furthermore, the figures are blurry and cluttered, making them difficult to follow and interpret. For example, Fig. 3, which is supposed to illustrate the core methodology, unnecessarily exaggerates the complexity of the proposed approach. Table 9 also appears to contain issues with the presentation or interpretation of numerical results, as 0.022 < 0.030 and 0.00143 < 0.00280.

Critical strengths
  1. The incorporation of coordination polyhedra is well motivated by solid-state chemistry. Coordination geometry, polyhedral distortion, and connectivity are important structural characteristics associated with electronic and mechanical properties. Explicitly representing these features provides a potentially useful inductive bias beyond conventional atom- and bond-level message passing.
  2. The combination of classical feature such as voronoi is interesting.
Critical weaknesses
  1. The ablation studies do not sufficiently isolate the contributions of the proposed components.
  2. The baselines are insufficient. Recent models like GemNet, ViSNet, MACE are not considered.
  3. The writing is unsatisfied, including vague pictures, tables with wrong annotation.
Questions for authors
  1. Please provide comparison with other baselines
  2. Please ablate each components
  3. Please isolate the influence of computational cost, since an additional graph is added in CPGN

Comments on the paper

1 comment
Area_Chair_dEJWOpenReview ↗

Official Comment by Area Chair dEJW

Dear reviewers,

Thank you very much for your time and effort in reviewing this submission. The authors have now submitted their rebuttal, and the discussion period is nearing its end. We would greatly appreciate it if you could take some time to carefully review the authors’ responses and share your assessment of whether your concerns have been adequately addressed. Thank you again for your valuable time and contributions to the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.