Pith. sign in

REVIEW 3 major objections 5 minor 26 references

ContinuouSP: Generative Model for Crystal Structure Prediction with Invariance and Continuity

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read ContinuouSP's energy function respects every re-description of a crystal and changes continuously with atomic motion.

desk verdict A promising formal framework for crystal generation with a load-bearing normalization flaw: the EBM density is not finite on the stated domain. read the letter →

arxiv 2502.02026 v1 pith:Y33ECS3Y submitted 2025-02-04 cs.LG cond-mat.mtrl-sci

Yuji Tone , Masatoshi Hanai , Mitsuaki Kawamura , Kenjiro Taura , Toyotaro Suzumura This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sci
keywords crystalstructurepredictionenergy-basedmodelinvariancecontinuityre-descriptionperiodicmaterialsgraphneuralnetworksMCMCsampling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a generative model for crystal structure prediction can satisfy two physically required properties at once: invariance under the symmetries of a crystal and continuity under small atomic displacements. It proposes ContinuouSP, an energy-based model whose energy function is a modified crystal graph convolutional network, and proves that this energy is strongly re-description invariant, invariant to translation and rotation, and continuous. From this it derives that the conditional sampling density $p_\theta(X,L\mid A)$ is weakly re-description invariant, invariant, and continuous, and that the training objective has the same properties. If the construction is correct, a crystal sampler would no longer depend on the arbitrary choice of unit cell, atomic ordering, or coordinate frame, and would respond smoothly to perturbations. The reported experiments are preliminary: the model outperforms predict-optimize baselines and matches several generative models, though not the strongest diffusion baseline.

What carries the argument

The machine that carries the argument is the modified graph convolution, Definition 16, whose update multiplies each message by $\cos^2(\pi \lVert x_j + Lk - x_i \rVert/(2D))$. The cosine-squared weight is one at zero distance, falls smoothly to zero as the distance approaches the cutoff $D$, and is exactly zero beyond it, so atoms entering or leaving the neighborhood no longer cause a jump in the output. Around this weight, the paper builds an energy-based model: $H_\theta = \mathrm{CGCNN}'_\theta$ is the energy, the conditional density is the Boltzmann factor divided by the partition function, and Metropolis-adjusted Langevin sampling draws coordinates and lattice bases. The proof of continuity uses Lemma 4, which transfers continuity of the energy on periodic units to continuity on the infinite solid material by choosing a periodic description with sufficiently many divisors of the period size.

What would settle it

Take a fixed species vector and evaluate the trained energy on periodic units with the same local atomic geometry but increasingly large lattice constants. If the energy decreases without bound as volume grows, then $\exp(-\beta H_\theta)$ is not integrable over $X\in\mathbb{R}^{3n}$, $L\in\mathbb{R}^{3\times 3}$, and the claimed probability density is not normalizable; this can be checked directly from the implementation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that replacing CGCNN's hard cutoff with a cosine-squared mask removes the only source of discontinuity without breaking the architecture's invariance, so the resulting energy $H_\theta = \mathrm{CGCNN}'_\theta$ can be plugged directly into the Boltzmann form $p_\theta(X,L\mid A)\propto \exp(-\beta H_\theta(A,X,L))$. Theorems 3 and 4 assert that $H_\theta$ is strongly re-description invariant (any two periodic units describing the same crystal give the same value) and that both the energy and the density are invariant to translation and rotation and continuous. The conditional density is claimed only in the weak form: invariance under re-descriptions that keep the number of atoms per period fixed. The method's training loss is also constructed, via a composition Reduce/Expand step and a geometric distribution over period sizes, so that it inherits strong re-description invariance, translation/rotation invariance, and continuity.

Load-bearing premise

The load-bearing premise is that the Boltzmann distribution over the unbounded space of coordinates and lattices is actually normalizable, so the density in Definition 18 exists as a probability density; the paper never proves the partition function is finite, and the implemented density penalty makes the energy fall without bound as the lattice volume grows.

Editorial extensions

If this is right

  • If Theorems 3 through 6 hold, a trained ContinuouSP energy gives the same value for any periodic unit describing the same crystal, so the model cannot be cheated by a different lattice-basis choice or a doubled cell.
  • The sampling density ignores global translations and rotations, so MCMC on fractional coordinates with Niggli reduction is not introducing symmetry-related bias.
  • Because the energy is continuous in atomic positions and lattice vectors, small displacements produce small energy changes, making gradient-based sampling and optimization on the energy landscape well behaved.
  • The training pseudo-loss is itself strongly re-description invariant and continuous, so invariant training data should yield an invariant and continuous model without data augmentation over unit-cell rewritings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: The paper's theoretical claims about the density presuppose that the Boltzmann integral is finite; if the energy is unbounded below under lattice expansion, the model may still serve as an unnormalized low-energy sampler, but the probabilistic interpretation in Theorem 4 would not hold as stated.
  • Editorial extension: Because re-description invariance is proven only weakly for the density (fixed number of atoms per period), a testable next step is to check numerically whether sampling from cells of different period sizes assigns consistent probabilities to the same infinite crystal; the Reduce/Expand loss is designed to help, but the paper does not demonstrate it.
  • Editorial extension: The cosine-squared window suggests a direct comparison with hard-cutoff CGCNN: measure the energy jump of a single atom crossing the cutoff in both models to confirm where the continuity gain comes from.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper formalizes translation/rotation invariance and continuity for crystal structure prediction, introduces ContinuouSP, an energy-based model built on a modified CGCNN with a cos^2 cutoff to ensure continuity, defines a conditional density pθ(X,L|A) over atomic coordinates and lattice parameters, proposes a pseudo-loss with geometric composition expansion, and reports CSP experiments on Perov-5, MP-20, and MPTS-52. The main theoretical claims are Theorem 3 (invariance/continuity of the energy) and Theorem 4 (same properties for the density). The paper also discusses the relation to data augmentation and diffusion models.

Significance. If the theorems were valid, the paper would provide a useful formal framework for invariant and continuous generative modeling of crystals, together with a concrete EBM implementation and public code. The cos^2 cutoff construction for continuity and the explicit treatment of re-description invariance are sensible contributions, and the experimental section is honestly reported as not achieving state-of-the-art results. However, the central probability-theoretic claim is not established: the proposed density is not normalizable on its stated domain, so the generative-model interpretation and the theoretical guarantees built on it are unsupported. The paper's value is therefore mainly in the formalization and the energy-function construction, not in the claimed conditional density model.

major comments (3)
  1. [Definition 18, Eq. (12)] The conditional density in Eq. (11) is not a probability density on the stated domain X ∈ R^{3n}, L ∈ R^{3×3}. Lemma 2 gives translation invariance, Hθ(A,X,L) = Hθ(A,X+t1,L), so exp(−βHθ) is constant along the three-dimensional translation direction and the X-integral in Eq. (12) contains a factor ∫_{R^3} dt = ∞. Independently, the implementation section states that the added density penalty is log(ρ/ρ0) with ρ = n/|det L| and ρ0 = 0.05 Å^{-3}; as |det L| → ∞ the interatomic distances all exceed the cutoff, the graph contribution approaches a constant, and Hθ ≈ const + log(n/(ρ0|det L|)) → −∞, making exp(−βHθ) ≈ C|det L|^β nonintegrable over L. Either reason alone makes Z(θ,β,A) infinite. Consequently pθ in Eq. (11) is improper, the expectations in Theorems 2 and 5 are undefined, and Theorem 4's claim concerns an object that has not been normalized.
  2. [Appendix: Properties of Probability Density Function in ContinuouSP] The proof of Theorem 4 only argues that the numerator and denominator of Eq. (11) transform in the same way under re-description. It never proves that Z(θ,β,A) is finite, and it does not address the fact that the integral in Eq. (12) is over absolute coordinates rather than a quotient by the translation symmetry. A proper density on the quotient would require an invariant measure and a different normalization, and the continuity claim would need to be proved with respect to that measure. As written, Theorem 4 is unsupported.
  3. [Definition 19 and Theorem 5, Eqs. (13)--(14)] The pseudo-loss J(θ) in Eq. (13) and its gradient formula in Eq. (14) both rely on log Z(θ,β,A') and on expectations under pθ(X,L|A'). Since Z is infinite for the reasons given above, the training objective is not well defined, and the claimed correspondence with standard EBM training is formal only. The infinite sum over the geometric distribution compounds the problem because each term requires a normalizing constant of an improper density.
minor comments (5)
  1. [Invariance of CGCNN, Lemma 1] Lemma 1 is stated for CGCNN′θ while the surrounding theorem concerns CGCNNθ; the notation should be made consistent, and the word 'invarant' should be corrected.
  2. [Definition 18 vs. Definition 5] Definition 18 integrates over all of R^{3n} × R^{3×3}, whereas the energy is defined on the non-overlapping periodic units Pn from Definition 5; the relationship between the integration domain and Pn should be stated explicitly.
  3. [Table 1 caption] The caption of Table 1 refers to 'MPTS-50' while the text and the table body use 'MPTS-52'; this should be fixed.
  4. [Conclusion] The conclusion claims a unique ability to maintain performance on extraordinary datasets, but no experiment on a clearly extraordinary dataset is reported; either provide such an experiment or soften the claim.
  5. [Appendix: Invariance of CGCNN] The proof of Lemma 1 is only a sketch; since strong re-description invariance is load-bearing for Theorem 3, the induction over graph-convolution steps should be written out in full.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the energy function, density, and loss are explicitly defined, and the derivations follow from those definitions rather than fitting or self-citing the target results.

full rationale

The claimed derivation chain is self-contained. The energy function Hθ is explicitly defined as a modified CGCNN (Definitions 16-17), with the cos^2 cutoff introduced to make the graph-convolution update continuous; Theorem 3 and its lemmas prove re-description invariance, translation/rotation invariance, and continuity directly from this definition. The probabilistic density is defined by the standard EBM normalization (Definition 18), and Theorem 4 transfers the invariance and continuity of Hθ to the density by invariance of both numerator and denominator; this is a direct consequence of the definition, not a fitted target. The pseudo-loss (Definition 19) and its gradient (Theorem 5) are derived by differentiating the explicit loss, and the evaluation is against held-out CSP benchmarks (Perov-5, MP-20, MPTS-52), so no fitted parameter is renamed as a prediction. The displacement-versus-energy experiment is presented as a qualitative check generated by the model, not as a quantity used to set constants. There are also no load-bearing self-citations: cited external methods are used as baselines or components, not to supply the paper's central claims. The main caveat, that finiteness of the partition function Z is assumed without proof and may fail due to translation zero modes and the logarithmic density penalty, is a mathematical-correctness risk rather than a circularity, because no target result is used to construct Hθ or pθ.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central theoretical claim rests on standard EBM machinery plus two unexamined assumptions: the Gibbs density is normalizable over unbounded periodic-unit space, and MCMC sampling reaches the target distribution. Both are likely false or unverified as stated; the density penalty introduced in implementation pushes the energy to -∞ as volume grows, so Z(θ,β,A) diverges. No free parameters are fitted to a target result; all listed hyperparameters are hand-chosen design choices.

free parameters (6)
  • Cutoff distance D = 3.0 × (atomic density)^(-1/3)
    Sets the graph neighborhood in Eq. (2); chosen per crystal, no sensitivity analysis.
  • Reference atomic density ρ0 = 0.05 Å^-3
    Used in the density penalty added to the energy output; hand-chosen.
  • Inverse temperature schedule β = exponential from 1 to 1000
    Annealing schedule for MALA sampling; arbitrary, no ablation.
  • MALA step size α = exponential from 0.5 to 0.0005
    Step-size schedule for MALA; no convergence diagnostics.
  • Geometric distribution parameter q = 0.5
    Controls the expansion of species vectors in the pseudo-loss; chosen by hand.
  • MALA iterations = 1000
    Number of sampling steps per gradient estimate; mixing not verified.
assumptions (6)
  • standard math Boltzmann-Gibbs form pθ ∝ exp(-βHθ) and detailed balance of MH/MALA
    Definitions 12-14; standard energy-based model formalism.
  • domain assumption Every crystal can be represented by a periodic unit (A,X,L) and PtoS is surjective
    Definitions 5-6; standard crystallographic assumption.
  • ad hoc to paper Partition function Z(θ,β,A) is finite over R^{3n} × R^{3×3}
    Needed for Definition 18 and Theorem 4; never proved and likely false with the density penalty used in implementation.
  • ad hoc to paper MALA sampling after 1000 annealed steps approximates pθ
    Implementation Details; no burn-in or convergence diagnostics.
  • domain assumption CGCNN graph convolution features from the periodic unit equal those from the infinite crystal graph
    Appendix Lemma 1; depends on well-defined mean pooling over infinite graphs.
  • domain assumption Training on stable crystals is sufficient for EBM learning
    Discussion section; argued qualitatively, not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ContinuouSP: Generative Model for Crystal Structure Prediction with Invariance and Continuity." pith.science (2026). https://pith.science/paper/Y33ECS3Y

@misc{pith2026250202026,
  author       = {Pith},
  title        = {Pith review of: ContinuouSP: Generative Model for Crystal Structure Prediction with Invariance and Continuity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y33ECS3Y}},
  note         = {Machine review of arXiv:2502.02026}
}
read the original abstract

The discovery of new materials using crystal structure prediction (CSP) based on generative machine learning models has become a significant research topic in recent years. In this paper, we study invariance and continuity in the generative machine learning for CSP. We propose a new model, called ContinuouSP, which effectively handles symmetry and periodicity in crystals. We clearly formulate the invariance and the continuity, and construct a model based on the energy-based model. Our preliminary evaluation demonstrates the effectiveness of this model with the CSP task.

Figures

Figures reproduced from arXiv: 2502.02026 by the authors.

Figure 1
Figure 1. Diagram illustrating the training workflow of ContinuouSP: For each periodic unit of the crystal included in the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The results of the displacement vs. energy experiment: For crystals with stable structures, an atom is selected from [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 22 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Behler, J.; and Parrinello, M. 2007. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces. Phys. Rev. Lett., 98: 146401

    work page 2007

  4. [4]

    E.; Landis, D

    Castelli, I. E.; Landis, D. D.; Thygesen, K. S.; Dahl, S.; Chorkendorff, I.; Jaramillo, T. F.; and Jacobsen, K. W. 2012 a . New cubic perovskites for one-and two-photon water splitting using the computational materials repository. Energy & Environmental Science, 5(10): 9034--9043

    work page 2012

  5. [5]

    E.; Olsen, T.; Datta, S.; Landis, D

    Castelli, I. E.; Olsen, T.; Datta, S.; Landis, D. D.; Dahl, S.; Thygesen, K. S.; and Jacobsen, K. W. 2012 b . Computational screening of perovskite metal oxides for optimal solar light capture. Energy & Environmental Science, 5(2): 5814--5819

    work page 2012

  6. [6]

    Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; and Ong, S. P. 2019. Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals. Chemistry of Materials, 31(9): 3564--3572

    work page 2019

  7. [7]

    Cheng, G.; Gong, X.-G.; and Yin, W.-J. 2022. Crystal structure prediction by combining graph network and optimization algorithm. Nature Communications, 13(1): 1492

    work page 2022

  8. [8]

    Choudhary, K.; and DeCost, B. 2021. Atomistic Line Graph Neural Network for improved materials property predictions. npj Computational Materials, 7(1): 185

    work page 2021

Show all 26 references
  1. [9]

    Du, Y.; Meier, J.; Ma, J.; Fergus, R.; and Rives, A. 2020. Energy-based models for atomic-resolution protein conformations. In International Conference on Learning Representations

    2020

  2. [10]

    Du, Y.; and Mordatch, I. 2019. Implicit Generation and Generalization in Energy-Based Models. CoRR, abs/1903.08689

    2019 arXiv

  3. [11]

    E.; and Sejnowski, T

    Hinton, G. E.; and Sejnowski, T. J. 1986. Learning and relearning in Boltzmann machines, 282–317. Cambridge, MA, USA: MIT Press. ISBN 026268053X

    1986

  4. [12]

    Jiao, R.; Huang, W.; Lin, P.; Han, J.; Chen, P.; Lu, Y.; and Liu, Y. 2023. Crystal Structure Prediction by Joint Equivariant Diffusion. In Oh, A.; Naumann, T.; Globerson, A.; Saenko, K.; Hardt, M.; and Levine, S., eds., Advances in Neural Information Processing Systems, volume...

    2023

  5. [13]

    Jiao, R.; Huang, W.; Liu, Y.; Zhao, D.; and Liu, Y. 2024. Space Group Constrained Crystal Generation. In The Twelfth International Conference on Learning Representations

    2024

  6. [14]

    Lin, P.; Chen, P.; Jiao, R.; Mo, Q.; Jianhuan, C.; Huang, W.; Liu, Y.; Huang, D.; and Lu, Y. 2024. Equivariant Diffusion for Crystal Structure Prediction. In Salakhutdinov, R.; Kolter, Z.; Heller, K.; Weller, A.; Oliver, N.; Scarlett, J.; and Berkenkamp, F., eds., Proceedings ...

    2024

  7. [15]

    Lin, Y.; Yan, K.; Luo, Y.; Liu, Y.; Qian, X.; and Ji, S. 2023. Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction. In Krause, A.; Brunskill, E.; Cho, K.; Engelhardt, B.; Sabato, S.; and Scarlett, J., eds., Proceedings of the 40th Intern...

    2023

  8. [16]

    Luo, Y.; Liu, C.; and Ji, S. 2023. Towards Symmetry-Aware Generation of Periodic Materials. In Oh, A.; Naumann, T.; Globerson, A.; Saenko, K.; Hardt, M.; and Levine, S., eds., Advances in Neural Information Processing Systems, volume 36, 53308--53329. Curran Associates, Inc

    2023

  9. [17]

    Nouira, A.; Sokolovska, N.; and Crivello, J.-C. 2019. CrystalGAN: Learning to Discover Crystallographic Structures with Generative Adversarial Networks. arXiv:1810.11203

    2019 arXiv

  10. [18]

    Ren, Z.; Tian, S. I. P.; Noh, J.; Oviedo, F.; Xing, G.; Li, J.; Liang, Q.; Zhu, R.; Aberle, A. G.; Sun, S.; Wang, X.; Liu, Y.; Li, Q.; Jayavelu, S.; Hippalgaonkar, K.; Jung, Y.; and Buonassisi, T. 2022. An invertible crystallographic representation for general inverse design o...

    2022

  11. [19]

    u tt, K.; Kindermans, P.-J.; Sauceda Felix, H. E.; Chmiela, S.; Tkatchenko, A.; and M\

    Sch\" u tt, K.; Kindermans, P.-J.; Sauceda Felix, H. E.; Chmiela, S.; Tkatchenko, A.; and M\" u ller, K.-R. 2017. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. In Guyon, I.; Luxburg, U. V.; Bengio, S.; Wallach, H.; Fergus, R.; Vish...

    2017

  12. [20]

    Taniai, T.; Igarashi, R.; Suzuki, Y.; Chiba, N.; Saito, K.; Ushiku, Y.; and Ono, K. 2024. Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding. In The Twelfth International Conference on Learning Representations

    2024

  13. [21]

    Xiang, S. 2021. Eliminating topological errors in neural network rotation estimation using self-selecting ensembles. ACM Trans. Graph., 40(4)

    2021

  14. [22]

    Xie, J.; Xu, Y.; Zheng, Z.; Gao, R.; Wang, W.; Song-Chun, Z.; and Wu, Y. N. 2021. Generative PointNet: Deep Energy-Based Learning on Unordered Point Sets for 3D Generation, Reconstruction and Classification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2021

  15. [23]

    Xie, T.; Fu, X.; Ganea, O.-E.; Barzilay, R.; and Jaakkola, T. S. 2022. Crystal Diffusion Variational Autoencoder for Periodic Material Generation. In International Conference on Learning Representations

    2022

  16. [24]

    Xie, T.; and Grossman, J. C. 2018. Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties. Phys. Rev. Lett., 120: 145301

    2018

  17. [25]

    Yan, K.; Liu, Y.; Lin, Y.; and Ji, S. 2022. Periodic Graph Transformers for Crystal Material Property Prediction. In The 36th Annual Conference on Neural Information Processing Systems

    2022

  18. [26]

    Zhou, Y.; Barnes, C.; Jingwan, L.; Jimei, Y.; and Hao, L. 2019. On the Continuity of Rotation Representations in Neural Networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.