Pith. sign in

REVIEW 4 major objections 4 minor 40 references

By partitioning equivariant graph neural networks according to a crystal's point group, this paper shows that most predictive power for elastic and dielectric tensors concentrates in the trivial irreducible-representation blocks, so an A1-r

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:40 UTC pith:YBFEY4CI

load-bearing objection A genuinely new architectural idea with a plausible but under-supported central claim; worth engaging, needs stronger capacity control and significance testing. the 4 major comments →

arxiv 2607.16871 v1 pith:YBFEY4CI submitted 2026-07-18 cond-mat.dis-nn cond-mat.mtrl-sci

Point Group Equivariant Graph Neural Networks for Materials

Alexander J. Heilman , Qimin Yan This is my paper
classification cond-mat.dis-nn cond-mat.mtrl-sci
keywords equivariant graph neural networkspoint group symmetrymaterials property predictionelastic tensordielectric tensorirreducible representationssymmetry-adapted basiscrystal symmetry principle
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that explicitly encoding a crystal's point group—not just its full three-dimensional rotation symmetry—changes what an equivariant graph neural network can learn about tensorial material properties, and that most of what it learns sits in the trivial symmetry channel. The authors introduce a network whose features, weights, and filters are indexed by both rotational order and point-group irreducible representation, then compare four variants at matched parameter counts on four elastic and four dielectric tensor datasets. They find that an A1-restricted model, which keeps only the invariant blocks, matches or improves on both the full point-group model and the conventional SO(3)-equivariant model while training fewer active parameters. The reason, they argue, is the symmetry principle that a crystal property tensor must be invariant under the crystal's point group, so the non-trivial blocks contribute only indirectly and are mostly unnecessary at current dataset sizes. If correct, this gives materials machine learning a way to build leaner models of equal accuracy and pinpoints where anisotropic signal is genuinely learnable.

Core claim

The paper's central claim is that the predictive content of equivariant graph networks for elastic and dielectric tensors concentrates in the trivial irreducible-representation blocks of a point-group-adapted basis. It constructs this basis by a unitary change from standard spherical harmonics, which block-diagonalizes every point-group operation and labels each component by an irreducible representation α of the crystal's point group in addition to its rotational order ℓ. The architecture, PGEqNN, uses this (ℓ, α) labelling to give distinct weights to components that a purely SO(3)-equivariant network must share. In a synthetic D4h classification task, the point-group-partitioned network se

What carries the argument

The central object is the point-group-adapted basis of point-group harmonics, obtained by applying a unitary matrix U_G^{(ℓ)} to the standard spherical harmonics of order ℓ. In this basis, the action of every point-group operation is block-diagonal: components belonging to different irreducible representations α are never mixed. That makes α an invariant label that can index learnable weights, so a convolution can assign distinct radial filters and self-interactions to each (ℓ, α) block, achieving sub-irrep specificity that a conventional SO(3)-equivariant layer, which must treat all 2ℓ+1 components of an ℓ block identically, cannot express. The A1-restricted variant then zeroes every filter

Load-bearing premise

The comparison relies on the assumption that two models with the same number of trainable weights have the same learning capacity, so that any accuracy difference is due to how symmetry is encoded; if that assumption is wrong, the main conclusion does not follow.

What would settle it

On the trigonal D3d dielectric dataset where the full point-group model is clearly best, retrain the A1-only variant with its parameter budget increased to the full model's budget (instead of forfeiting the budgets of non-trivial orders). If the A1-only model then matches the full model's total MAE, the paper's claim that non-trivial blocks are the source of the anisotropic-signal advantage is falsified; if the full model retains its lead, that claim survives.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • On all four elastic datasets, the A1-only model posts the lowest total mean absolute error while using roughly half to two-thirds the active parameters of the SO(3) reference; the full point-group partition separates from SO(3) only on D2h.
  • On the dielectric datasets, the three symmetry-aware partitions are statistically tied in total error for D2h, C2v, and C2h; only on D3d, where the ℓ=2 deviatoric block has strong learnable signal, does the full point-group partition clearly beat both SO(3) and A1-only.
  • The invariant scalar baseline trails every symmetry-aware model on every dataset and sometimes lands above the no-skill floor, showing that invariant angular content of local environments carries information beyond interatomic distances.
  • The point-group partition helps exactly when three conditions coincide: the target populates ℓ>0 subspaces, the point group resolves those subspaces into multiple irreps or repeated trivial copies, and the dataset carries anisotropic signal above its no-skill floor.
  • A synthetic D4h vector-orientation task demonstrates that a purely SO(3)-equivariant model cannot distinguish components within the same rotational order that transform under different point-group irreps, while the point-group-partitioned model solves it perfectly.

Where Pith is reading between the lines

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

  • Inference: If the concentration in trivial blocks holds for symmetry-constrained tensor targets, then the full point-group partition should show larger gains on targets that genuinely populate non-trivial irreps—electronic Hamiltonians, Berry curvature, magnetic tensors—where the extra blocks are not discarded by symmetry. The paper itself points toward Hamiltonians as the natural next test.
  • Inference: The parameter-matching protocol uses binary masks to equalize active counts per rotational order. Whether this truly equalizes capacity is not established; a more direct test would sweep model width and mask budget independently to see whether the A1-only advantage survives at equal total parameter count and equal training epochs.
  • Inference: The within-ℓ specificity mechanism is not limited to message-passing architectures; it could be attached to attention-based or other equivariant backbones, and to weight sharing across point groups via correlation tables, which the paper's discussion of repeated A1 copies suggests.
  • Inference: The finding that full point-group models are most useful where anisotropic signal exceeds the floor suggests a practical screening rule: compute the no-skill floor of the anisotropic block, and only invest in the full partition if the floor is far enough below the target error; otherwise the leaner A1-only model is the rational choice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces point-group-symmetry-aware equivariant graph neural networks (PGEqNN), which partition equivariant weights and features not only by SO(3) rotational order ℓ but also by the irreducible representations of the material's point group in a symmetry-adapted basis. The authors compare four models—an invariant scalar baseline, an SO(3)-partitioned equivariant model, a full point-group-partitioned model, and an A1-restricted variant—on elastic and dielectric tensor prediction from Materials Project data, using active parameter counts matched per rotational order through binary masks. The central claim is that predictive content concentrates in trivial-irrep (A1) blocks, so the A1-restricted variant matches or improves on the full PG and SO(3) models while training fewer active parameters. A synthetic D4h vector-orientation task demonstrates that PG partitioning can distinguish components within a single ℓ that SO(3) equivariance cannot.

Significance. If the concentration claim were firmly established, the paper would provide a practical design principle: for tensor targets governed by Neumann's principle, equivariant models can be restricted to trivial-irrep blocks, yielding leaner models of equal accuracy. The synthetic task cleanly demonstrates the architectural capability, and the use of MAD floors as no-skill references is a good practice. The paper also makes a useful conceptual contribution by connecting deterministic crystallographic canonicalization to point-group-adapted equivariant architectures. However, the central empirical conclusion is currently not supported by the reported statistics: most headline comparisons are within cross-validation scatter, no significance tests are provided, and the parameter-matching protocol is not a validated capacity control.

major comments (4)
  1. [Tensor Prediction (Tables 2 and 3)] The headline comparisons are within cross-validation scatter. For example, Table 2 Oh: PG 22.6±1.1, A1-only 22.1±1.0, SO(3) 22.5±1.0; D6h: 39.9±3.0 vs 40.8±2.5; D4h: 31.9±1.7 vs 32.0±2.1. In Table 3, the D2h total for A1-only (0.435±0.042) is numerically worse than PG (0.428±0.048) and SO(3) (0.428±0.042), and the C2h total is similarly tied or slightly worse. Even the one separation highlighted in the text, D3d PG 0.396±0.028 vs A1-only 0.425±0.046, is within combined fold scatter. No paired per-fold tests, multi-seed repetitions, or block-level error bars are reported. Consequently, the claims that the A1-restricted model 'matches or improves' and that predictive content 'concentrates' in trivial blocks are not supported by the reported statistics.
  2. [Tensor Prediction and A1-Restricted Point-Group Model] The parameter-matching protocol does not equalize capacity. Only self-interaction and radial-filter weights are masked order-by-order; normalization/scaling and readout heads are outside the matched protocol, and the A1-only variant deliberately gives each repeated A1 copy its own readout. Total active parameter counts differ substantially (e.g., Oh A1-only 133.5k vs SO(3) 287.0k in Table 2). Zeroing non-trivial blocks acts as a hard regularizer, and with L2 λ=1e-4 the observed equivalence could reflect regularization rather than trivial-irrep concentration. A control is needed: match total active parameters by width reduction, or include random block-mask ablations with identical readout structure and normalization, and compare with multi-seed runs.
  3. [A1-Restricted Point-Group Model and Results] The causal interpretation is weakened by architecture selection on the same datasets. The A1-restricted variant is motivated by inspecting learned filter couplings from full PG models trained on these same elastic datasets, and the same datasets are then used to 'confirm causally' that A1 content dominates. This is a selection-on-the-test-data concern. To support the claim, the A1-only restriction should be pre-registered, selected on a held-out point group, or at least the number of architecture variants tried after weight inspection should be reported.
  4. [Elastic Tensor Prediction (Table 2, D2h)] The 'below MAD floor' claims lack uncertainty quantification. For the D2h symmetric ℓ=2 block, the floor is 38.4 GPa and the PG variants report 37.3 and 36.4 GPa—differences of only 1–2 GPa. For symmetric ℓ=4, PG 45.9/45.4 vs floor 56.1, but no block-level error bars are given; the total cross-fold σ is ±1.2–1.9 GPa. The conclusion that PG variants 'recover a measurable signal' while SO(3) does not is therefore not established. Block-level MAE should be accompanied by uncertainties or significance tests.
minor comments (4)
  1. [General] There are several typographical errors: 'it’s' should be 'its' in multiple places (e.g., 'it’s corresponding tensor' in Standard Bases), and 'material’s' appears in the Abstract. The manuscript would benefit from a copyedit.
  2. [Standard Bases] The footnote '1These properties were one-hot encoded.' appears as a dangling footnote marker in the text. It should be integrated or placed as a proper table/figure caption.
  3. [Equation (3)] The notation in Eq. (3) is under-specified: the object U^γ_n_{αiβj} is introduced as a Clebsch–Gordan coefficient, but its indices and dependence on the point-group chain (and on the edge's point group) should be defined explicitly. Also, the superscript/subscript placement of γ_n, αi, βj is confusing.
  4. [Tables 2 and 3] Table 3 does not include MAD floor rows, although the text reports floors for each block. Including them in the table would make the comparison against no-skill references more transparent, as done in Table 2.

Circularity Check

0 steps flagged

No significant circularity: the central result is a held-out ablation comparison; self-citations are contextual, not load-bearing.

full rationale

The paper's central claim—that predictive content in the point-group partition concentrates in trivial-irrep blocks—is supported by a test-set comparison of an A1-restricted model against full PG and SO(3) baselines, not by construction. The A1-restricted model is an ablation that zeroes non-trivial blocks; it could have performed worse, so its matching or improving performance is an empirical finding rather than a tautology. Neumann's principle motivates the hypothesis but does not determine the relative performance of the architectures. The only self-citations ([5], [10]) are contextual references to prior GNN and tensor-prediction work and are not invoked to justify the paper's novel partition or to forbid alternatives. The parameter-matching protocol raises a capacity-control concern, but that is a validity issue, not circularity: no fitted parameter is renamed as a prediction, and the target tensors are external Materials Project data. Therefore no circular step is identified.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central empirical claim rests on several unstated model choices (width, ℓmax, layers, λ, normalization statistics, parameter budgets) and on standard crystallographic/ML assumptions: standard frames via spglib/pymatgen, MultiPie PGH matrices, Neumann's principle, CrystalNN edge selection, and Materials Project DFT labels. No new physical entities are introduced; the A1-restricted model is an architectural hypothesis rather than a fitted constant. The main scientific burden is the unverified capacity-equalization assumption and the small, selectively reported dataset cohorts.

free parameters (7)
  • L2 regularization strength λ = 1e-4
    Used throughout training (Fig. 2 caption); chosen to suppress spurious weights and directly shapes which filter couplings decay, motivating the A1-restricted model.
  • ℓ_max for edge spherical harmonics = not stated
    Edge features are spherical harmonics 0≤ℓ≤ℓ_max (Methods, Wyckoff Graphs), but the maximum order is never given; this controls feature resolution and parameter budgets.
  • convolutional block width / hidden channels = not stated
    A 'common block width' is mentioned but its value is not reported; capacity of all compared models depends on it.
  • number of convolutional layers = 2 (implied)
    Figure 2 caption refers to 'the second layer of 2'; no explicit architecture table is provided.
  • per-rank-ℓ target normalization constants = fitted per block
    Targets are normalized per rank-ℓ block (Invariant Scalar Baseline); normalization statistics are computed from data and affect reported MAE values.
  • parameter-matching masks and budgets = not fully specified
    PG and A1 variants are budget-matched 'within every rotational order' via binary masks reapplied after optimizer steps; the exact allocation is not specified beyond active parameter counts, making capacity equalization hard to audit.
  • point-group cohort selection = Oh, D6h, D4h, D2h (elastic); D2h, D3d, C2v, C2h (dielectric)
    Only four point groups per property are reported, without a stated selection criterion; small cohorts (n=330 to 3,663) limit the generality of the 'matches or improves' claim.
axioms (6)
  • domain assumption Neumann's principle: symmetric property tensors have non-zero components only in the trivial-irrep (A1) blocks of the point-group-adapted basis.
    Invoked in Methods, Standard Bases, and Discussion; it justifies predicting only A1 coefficients and motivates the A1-restricted model.
  • domain assumption spglib/pymatgen standardization places structures in the International Tables standard setting, making IR labels physically meaningful.
    Used throughout Methods; if standardization is incorrect or non-unique, the A1 decomposition and masked targets are wrong.
  • standard math MultiPie point-group-harmonic matrices U_G^(ℓ) correctly block-diagonalize point-group representations.
    The entire PGH basis change in Eq. 2 rests on these external matrices, cited to [36].
  • standard math Clebsch-Gordan tensor products decomposed by point-group irreps are complete and correctly implemented.
    Message passing in Eq. 3 relies on the PG-irreducible decomposition of tensor products; no implementation verification is provided.
  • domain assumption CrystalNN edge selection produces edges that reflect site symmetry and are suitable for Wyckoff-graph message passing.
    Methods state CrystalNN is chosen both to reduce overfitting and to 'encourage the formation of edges that most directly reflect the symmetry of the site'; this is a heuristic assumption.
  • domain assumption Materials Project DFT elastic and dielectric tensors are accurate ground truth for the masked non-zero components.
    All benchmark conclusions inherit the accuracy of the underlying DFT calculations in [38,39].

pith-pipeline@v1.3.0-alltime-deepseek · 15240 in / 11798 out tokens · 124090 ms · 2026-08-01T19:40:30.144458+00:00 · methodology

0 comments
read the original abstract

Equivariant graph neural networks have proven effective tools for inference of material's properties directly from their structure. Traditionally, these have been applied such that they respect full $O(3)$ equivariance, so that any rotation or reflection of the input structure is respected in the model's output. While this works for general arrangements of atoms, additional symmetries of atomistic systems are left unleveraged. Furthermore, any symmetries of the filter functions are implicitly learned from the full dataset and not strictly enforced. In this work, we introduce point-group symmetry aware equivariant graph neural networks (PGEqNN) for materials science, with filter functions aligned with symmetry-aware indices for greater granularity in predictive tasks. With this architecture, we show that most of the predictive power of equivariant networks for tensorial elastic and dielectric datasets lies in the trivial subspaces of the point-group adapted bases. Exploiting this, an $A_1$-restricted variant matches or improves on its full point-group and $SO(3)$-partitioned counterparts while training fewer active parameters, yielding leaner models of equal accuracy.

Figures

Figures reproduced from arXiv: 2607.16871 by Alexander J. Heilman, Qimin Yan.

Figure 1
Figure 1. Figure 1: Symmetry-labelled subspaces of MoS2 and it’s corresponding Wyckoff graph. In the present work, symmetric sites are mapped onto one node representing their Wyckoff position in a reduced Wyckoff graph. The central molybdenum sites of the material molybdenum disulfide have D6h symmetry, which may be used to partition the full 3D space into a 2D symmetry plane labelled by the IR E, and a trivial axis labelled … view at source ↗
Figure 2
Figure 2. Figure 2: (a) Workflow for the results presented in this paper. Materials’ axis alignment with a standard [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Point Group Equivariant Model Architecture. Schematic depicting architecture of equiv￾ariant models depicted in this work. Atomic embeddings follow that used in CGCNN [4] for the scalar block, while edge embeddings are a gaussian expansion of distance with dimension equal to block size. For each convolutional block, a radial filter function is evaluated for the edge feature, self interaction is applied to … view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the synthetic D4h vector-orientation task. Vector node features are either uniformly along the A2u principal axis or uniformly in the Eu symmetry plane. The SO(3)-equivariant network, weighting the ℓ = 1 block as a whole, has no channel that separates the two cases; the PG-partitioned network distinguishes them by routing through separate A2u and Eu weight paths. The SO(3) network performs at … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

40 extracted references · 5 canonical work pages

  1. [1]

    Machine learning in materials science

    Jing Wei et al. “Machine learning in materials science”. In:InfoMat1.3 (2019), pp. 338–358

    2019

  2. [2]

    Machine learning in materials genome initiative: A review

    Yingli Liu et al. “Machine learning in materials genome initiative: A review”. In:J. Mater. Sci. Tech. 57 (2020), pp. 113–122.issn: 1005-0302.doi:https://doi.org/10.1016/j.jmst.2020.01.067.url: https://www.sciencedirect.com/science/article/pii/S1005030220303327

    work page doi:10.1016/j.jmst.2020.01.067.url: 2020

  3. [3]

    Scope of machine learning in materials research—A review

    Md Hosne Mobarak et al. “Scope of machine learning in materials research—A review”. In:Appl. Surf. Sci. Adv.18 (2023), article 100523.issn: 2666-5239.doi:https://doi.org/10.1016/j.apsadv. 2023.100523.url:https://www.sciencedirect.com/science/article/pii/S2666523923001575

    arXiv 2023

  4. [4]

    Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties

    Tian Xie and Jeffrey C Grossman. “Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties”. In:PRL120.14 (2018), article 145301

    2018

  5. [5]

    Crystal hypergraph convolutional networks

    Alexander J Heilman, Weiyi Gong, and Qimin Yan. “Crystal hypergraph convolutional networks”. In: npj Computational Materials11.1 (2025), p. 336

    2025

  6. [6]

    Graph networks as a universal machine learning framework for molecules and crystals

    Chi Chen et al. “Graph networks as a universal machine learning framework for molecules and crystals”. In:Chem. Mater.31.9 (2019), pp. 3564–3572

    2019

  7. [7]

    A universal graph deep learning interatomic potential for the periodic table

    Chi Chen and Shyue Ping Ong. “A universal graph deep learning interatomic potential for the periodic table”. In:Nat. Comput. Sci.2.11 (2022), pp. 718–728

    2022

  8. [8]

    e3nn: Euclidean neural networks

    Mario Geiger and Tess Smidt. “e3nn: Euclidean neural networks”. In:arXiv preprint arXiv:2207.09453 (2022)

    Pith/arXiv arXiv 2022

  9. [9]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    Nathaniel Thomas et al. “Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds”. In:arXiv preprint arXiv:1802.08219(2018). arXiv:1802.08219 [cs.LG].url: https://arxiv.org/abs/1802.08219

    Pith/arXiv arXiv 2018

  10. [10]

    Alex Heilman, Claire Schlesinger, and Qimin Yan.Equivariant Graph Neural Networks for Prediction of Tensor Material Properties of Crystals. 2024. arXiv:2406.03563 [physics.comp-ph].url:https: //arxiv.org/abs/2406.03563

    Pith/arXiv arXiv 2024

  11. [11]

    Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures

    Nguyen Tuan Hung et al. “Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures”. In:Advanced Materials36.46 (2024), p. 2409175

    2024

  12. [12]

    Equivariance with Learned Canonicalization Functions

    S´ ekou-Oumar Kaba et al. “Equivariance with Learned Canonicalization Functions”. In:Proceedings of the 40th International Conference on Machine Learning (ICML). Vol. 202. Proceedings of Machine Learning Research. 2023, pp. 15546–15566

    2023

  13. [13]

    Spatial Transformer Networks

    Max Jaderberg et al. “Spatial Transformer Networks”. In:Advances in Neural Information Processing Systems (NeurIPS). Vol. 28. 2015

    2015

  14. [14]

    Equivariant Adaptation of Large Pretrained Models

    Arnab Kumar Mondal et al. “Equivariant Adaptation of Large Pretrained Models”. In:Advances in Neural Information Processing Systems (NeurIPS). Vol. 36. 2023

    2023

  15. [15]

    Spglib: a software library for crystal symmetry search

    Atsushi Togo, Kohei Shinohara, and Isao Tanaka. “Spglib: a software library for crystal symmetry search”. In:Science and Technology of Advanced Materials: Methods4.1 (2024), p. 2384822.doi: 10.1080/27660400.2024.2384822. eprint:https://doi.org/10.1080/27660400.2024.2384822. url:https://doi.org/10.1080/27660400.2024.2384822

    arXiv 2024

  16. [16]

    Revisiting the Canonicalization for Fast and Accurate Crystal Tensor Property Prediction

    Haowei Hua et al. “Revisiting the Canonicalization for Fast and Accurate Crystal Tensor Property Prediction”. In:Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. 1. 2026, pp. 417– 425

    2026

  17. [17]

    SchNet – A deep learning architecture for molecules and materials

    K. T. Sch¨ utt et al. “SchNet – A deep learning architecture for molecules and materials”. In:J. Chem. Phys.148.24 (Mar. 2018), article 241722.issn: 0021-9606.doi:10.1063/1.5019779. eprint:https: //pubs.aip.org/aip/jcp/article-pdf/doi/10.1063/1.5019779/16655678/241722\_1\_online. pdf.url:https://doi.org/10.1063/1.5019779

    doi:10.1063/1.5019779 2018

  18. [18]

    Atomistic Line Graph Neural Network for improved materials property predictions

    Kamal Choudhary and Brian DeCost. “Atomistic Line Graph Neural Network for improved materials property predictions”. In:npj Comput. Mater.7.1 (2021), pp. 1–8. 16

    2021

  19. [19]

    Developing an improved crystal graph convolutional neu- ral network framework for accelerated materials discovery

    Cheol Woo Park and Chris Wolverton. “Developing an improved crystal graph convolutional neu- ral network framework for accelerated materials discovery”. In:Phys. Rev. Mater.4.6 (2020), article 063801

    2020

  20. [20]

    A geometric-information-enhanced crystal graph network for predicting properties of materials

    Jiucheng Cheng, Chunkai Zhang, and Lifeng Dong. “A geometric-information-enhanced crystal graph network for predicting properties of materials”. In:Commun. Mater.2.1 (2021), article 92

    2021

  21. [21]

    SE(3)-Transformers: 3D roto-translation equivariant attention networks

    Fabian B. Fuchs et al. “SE(3)-Transformers: 3D roto-translation equivariant attention networks”. In: Advances in Neural Information Processing Systems (NeurIPS). Vol. 33. 2020

    2020

  22. [22]

    Cormorant: Covariant molecular neural net- works

    Brandon Anderson, Truong Son Hy, and Risi Kondor. “Cormorant: Covariant molecular neural net- works”. In:Advances in Neural Information Processing Systems (NeurIPS). Vol. 32. 2019

    2019

  23. [23]

    E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials

    Simon Batzner et al. “E(3)-equivariant graph neural networks for data-efficient and accurate inter- atomic potentials”. In:Nature Communications13.1 (2022), p. 2453.doi:10 . 1038 / s41467 - 022 - 29939-5

    2022

  24. [24]

    MACE: Higher order equivariant message passing neural networks for fast and accurate force fields

    Ilyes Batatia et al. “MACE: Higher order equivariant message passing neural networks for fast and accurate force fields”. In:Advances in Neural Information Processing Systems (NeurIPS). Vol. 35. 2022

    2022

  25. [25]

    Learning local equivariant representations for large-scale atomistic dynamics

    Albert Musaelian et al. “Learning local equivariant representations for large-scale atomistic dynamics”. In:Nature Communications14.1 (2023), p. 579

    2023

  26. [26]

    EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations

    Yi-Lun Liao et al. “EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations”. In:International Conference on Learning Representations (ICLR). 2024.url:https: //openreview.net/forum?id=mCOBKZmrzD

    2024

  27. [27]

    An equivariant graph neural network for the elasticity tensors of all seven crystal systems

    Mingjian Wen et al. “An equivariant graph neural network for the elasticity tensors of all seven crystal systems”. In:Digital Discovery3.5 (2024), pp. 869–882

    2024

  28. [28]

    Discovery of Highly Anisotropic Dielectric Crystals with Equiv- ariant Graph Neural Networks

    Yuchen Lou and Alex M. Ganose. “Discovery of Highly Anisotropic Dielectric Crystals with Equiv- ariant Graph Neural Networks”. In:Faraday Discussions256.0 (2025), pp. 255–274.doi:10.1039/ D4FD00096J.url:https://pubs.rsc.org/en/content/articlelanding/2025/fd/d4fd00096j

    2025

  29. [29]

    Frame averaging for invariant and equivariant network design

    Omri Puny et al. “Frame averaging for invariant and equivariant network design”. In:International Conference on Learning Representations (ICLR). 2022

    2022

  30. [30]

    Connectivity optimized nested line graph networks for crystal structures

    Robin Ruff et al. “Connectivity optimized nested line graph networks for crystal structures”. In:Digit. Discov.3.3 (2024), pp. 594–601

    2024

  31. [31]

    Remarks on Wyckoff positions

    Ulrich M¨ uller. “Remarks on Wyckoff positions”. In:International Tables for Crystallography Volume A1: Symmetry relations between space groups. Springer, 2006, pp. 24–26

    2006

  32. [32]

    Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis

    Shyue Ping Ong et al. “Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis”. In:Computational Materials Science68 (2013), pp. 314–319

    2013

  33. [33]

    Neural message passing for quantum chemistry

    Justin Gilmer et al. “Neural message passing for quantum chemistry”. In:34th Int. Conf. Mach. Learn. Vol. 70. PMLR. 2017, pp. 1263–1272

    2017

  34. [34]

    Complete and Efficient Graph Transformers for Crystal Material Property Pre- diction

    Keqiang Yan et al. “Complete and Efficient Graph Transformers for Crystal Material Property Pre- diction”. In:arXiv preprint arXiv:2403.11857(2024). arXiv:2403.11857 [cs.LG].url:https:// arxiv.org/abs/2403.11857

    Pith/arXiv arXiv 2024

  35. [35]

    Bradley and Arthur P

    Christopher J. Bradley and Arthur P. Cracknell.The Mathematical Theory of Symmetry in Solids: Representation Theory for Point Groups and Space Groups. Clarendon Press, Oxford, 1972.isbn: 9780198519201

    1972

  36. [36]

    Multipole expansion for magnetic structures: A generation scheme for a symmetry- adapted orthonormal basis set in the crystallographic point group

    M.-T. Suzuki et al. “Multipole expansion for magnetic structures: A generation scheme for a symmetry- adapted orthonormal basis set in the crystallographic point group”. In:Physical Review B99.17 (2019), p. 174407.doi:10.1103/PhysRevB.99.174407. 17

    work page doi:10.1103/physrevb.99.174407 2019

  37. [37]

    Commentary: The Materials Project: A materials genome approach to accelerating materials innovation

    Anubhav Jain et al. “Commentary: The Materials Project: A materials genome approach to accelerating materials innovation”. In:APL Mater.1.1 (July 2013), article 011002.issn: 2166-532X.doi:10.1063/ 1.4812323. eprint:https://pubs.aip.org/aip/apm/article- pdf/doi/10.1063/1.4812323/ 13163869/011002\_1\_online.pdf.url:https://doi.org/10.1063/1.4812323

    work page doi:10.1063/1.4812323/ 2013

  38. [38]

    Charting the complete elastic properties of inorganic crystalline compounds

    Maarten de Jong et al. “Charting the complete elastic properties of inorganic crystalline compounds”. In:Scientific Data2.1 (2015), p. 150009.issn: 2052-4463.doi:10.1038/sdata.2015.9.url:https: //doi.org/10.1038/sdata.2015.9

    work page doi:10.1038/sdata.2015.9.url:https: 2015

  39. [39]

    High-throughput screening of inorganic compounds for the discovery of novel dielectric and optical materials

    Ioannis Petousis et al. “High-throughput screening of inorganic compounds for the discovery of novel dielectric and optical materials”. In:Scientific Data4.1 (2017), p. 160134.issn: 2052-4463.doi:10. 1038/sdata.2016.134.url:https://doi.org/10.1038/sdata.2016.134

    doi:10.1038/sdata.2016.134 2017

  40. [40]

    A database to enable discovery and design of piezoelectric materials

    Maarten de Jong et al. “A database to enable discovery and design of piezoelectric materials”. In: Scientific Data2.1 (2015), p. 150053.issn: 2052-4463.doi:10.1038/sdata.2015.53.url:https: //doi.org/10.1038/sdata.2015.53. 18

    work page doi:10.1038/sdata.2015.53.url:https: 2015