Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a diffusion model can output a crystal's complete symmetry specification—space group and Wyckoff positions included—by reversing learned symmetry-breaking jumps from the lowest-symmetry group P1, and that this…

desk verdict Novel jump-diffusion machinery for space-group transitions, but the headline claim of complete crystallographic specifications is not actually validated; Appendix E concedes the coordinate constraint is open. read the letter →

arxiv 2608.13457 v1 pith:ZDMD2HY6 submitted 2026-08-13 cs.LG

Van Khoa Nguyen , Alexandros Kalousis This is my paper · ORCID
classification cs.LG
keywords crystalgenerationdiffusionmodelsspacegroupsWyckoffpositionssymmetrybreakingMarkovianjumpsitematerialsdiscovery
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 proposes SbCD, a diffusion framework that generates crystals complete with their symmetry specifications: the space group and each atom's Wyckoff position are produced by the model rather than sampled from the training distribution. Its central move is to run the forward process as a symmetry-breaking chain that sends every crystal's space group through random jumps down to P1, the lowest-symmetry group, and to learn the reverse process that climbs back up. Because the space group changes during diffusion, the lattice and site-symmetry constraints change with it; the paper derives closed-form forward and reverse kernels for both continuous and discrete components. On de novo generation over MP-20 and MPTS-52, the model beats its symmetry-preserving counterpart on key metrics including thermodynamic stability, composition validity, and space-group distribution distance. If correct, this is the first generative framework whose output includes the full crystallographic specification, which matters because downstream property prediction and synthesis planning depend on symmetry.

What carries the argument

The central object is a Markovian jump diffusion—a continuous-time Markov process that holds a state for a random duration and then jumps—run over the 230 crystallographic space groups, with a forward rate matrix that sends every group to the absorbing lowest-symmetry group P1. Around that jump chain, two adaptive kernels carry the argument: a masked Gaussian diffusion over the six-dimensional lattice representation (mask–bias pairs per crystal family switch when the space group jumps), and a categorical diffusion over site symmetry whose one-hot labels are the 81 distinct oriented site-symmetry symbols, with a prior that interpolates linearly between the old and new space group's site-symmetry priors. The reverse process starts at P1 with trivial site symmetry and samples space-group jumps via τ-leaping, then uses the sampled space group to choose the lattice mask and site-symmetry prior for the next step.

What would settle it

Estimate the true log-likelihood of held-out crystals by importance sampling over the complete reverse path and compare it with the variational bound used in training; a systematic gap that grows with sample size would show the assumed factorization is not met. A cheaper check is to condition the atom-type predictor on the predicted space group and see whether validation likelihood improves.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that generating a complete crystal specification—not just atoms in a cell but the space group and the Wyckoff positions of its sites—can be learned, without an empirical space-group prior, by training a reverse Markovian jump-diffusion process that starts at the lowest-symmetry space group P1. The forward process jumps any space group down to P1 at a constant rate; the reverse process learns the jump rates back up, and the space group's evolution changes the constraints applied to the lattice and site symmetries. Closed-form kernels are derived for the discrete space-group chain, the continuous lattice under mask–bias constraints, and a categorical chain over 81 site-symmetry symbols. On the MP-20 and MPTS-52 benchmarks, the claimed result is that SbCD variants outperform their symmetry-preserving counterpart across structural and compositional validity, property statistics, and thermodynamic stability, matching or approaching models that condition on empirical space-group distributions.

Load-bearing premise

The training loss is a valid upper bound on the model's log-likelihood only if the reverse generative process factorizes into independent components (space group, lattice and site symmetry given the space group, atom types, coordinates); if those components actually interact during reversal, the model may be optimizing something other than the data likelihood.

Editorial extensions

If this is right

  • Generated crystals can carry space groups and Wyckoff positions that are rare or absent in the training distribution, because generation no longer requires an empirical space-group prior.
  • The learned inter-space-group transitions make few-step generation efficient: validity at 50 function evaluations is over 65 percent for SbCD against about 25 percent for its symmetry-preserving counterpart.
  • Site-symmetry learning over 81 symbols removes the need to project generated symmetries onto the nearest valid point group, which stabilizes training and downstream Wyckoff assignment.
  • The posterior holding-time formula gives an adaptive schedule for when symmetry breaking occurs during generation; fixed windows produce higher structural validity but poorer density statistics.
  • Lattices and site symmetries are constrained exactly at every generation step by the mask–bias pair of the current space-group family, so sampled crystals respect crystallographic constraints without external templates.

Reading between the lines

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

  • Beyond the paper: the variational bound's assumed factorization could be stress-tested by importance-sampling the true likelihood; a large gap would point to coupling between the space group and the atom-type or coordinate components.
  • Beyond the paper: because every forward jump goes directly to P1, the model only learns one-step symmetry breaking; intermediate transitions through subgroups are a concrete extension that could make generated paths physically more realistic.
  • Beyond the paper: the 81-symbol site-symmetry encoding is separable from the jump-diffusion machinery and could be reused by any asymmetric-unit crystal generator.
  • Beyond the paper: the fixed-window ablation suggests the symmetry-breaking schedule is a tunable trade-off between structural validity and crystal density, which could be exploited for density-targeted screening.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 proposes SbCD, a generative model for crystals that jointly models the space group, site symmetries, lattice, atom types, and fractional coordinates of an asymmetric unit. The key idea is a Markovian jump-diffusion framework in which a forward process corrupts a crystal by repeatedly transitioning its space group to the lowest-symmetry group P1, while the reverse process learns to restore symmetry by generating space groups, site symmetries, and the associated structural components. The authors derive a variational upper bound on the negative log-likelihood, split into continuous and discrete diffusion terms, and provide analytical forward/reverse kernels for the lattice and site-symmetry processes under symmetry breaking. On MP-20 and MPTS-52, SbCD outperforms its symmetry-preserving counterpart and several baselines in stable-structure generation and space-group distribution matching. The paper's central claim is that SbCD is the first framework to generate complete crystallographic structure specifications, including space groups and Wyckoff positions.

Significance. If the central claims are borne out, SbCD is a useful step: it replaces empirical space-group sampling with a learned distribution, unifies continuous and discrete diffusion objectives, and demonstrates consistent gains over a symmetry-preserving baseline. The paper is also careful to provide analytical forms and explicit training and sampling algorithms, and the ablation against the no-symmetry-breaking variant supports the contribution of the symmetry-breaking mechanism. However, the central claim about complete crystallographic specifications is not verified: the evaluation never checks whether generated coordinates actually respect the declared site symmetry and space group. The theoretical bound also rests on an unverified factorization assumption. These issues do not invalidate the empirical method, but they need to be addressed before the paper's claims can be accepted.

major comments (3)
  1. [Section 5 / Appendix B.3 / Appendix E] The paper's central claim that SbCD produces complete crystallographic structure specifications is not supported by the evaluation. None of the metrics in Tables 2 and 3 checks whether a generated fractional coordinate F0 lies on the Wyckoff manifold of the declared site-symmetry group S0, or whether the replicated structure actually has the declared space group G0. Algorithm 1 line 9 generates F with the unconstrained DiffCSP diffusion, and Appendix B.3 assigns a Wyckoff position only a posteriori via search_closest_wp; Appendix E explicitly states that enforcing dynamic symmetry constraints on fractional coordinates is left to future work. Consequently, a generated (G,S,F) triple can be internally inconsistent, and the 'complete specification' claim outruns the evidence. I request either a per-sample consistency check (e.g., using PyXtal or Spglib to determine the actual space group of the generated structure and comparing it with the declared G, or checking that F0 lies on the appropriate Wyckoff subspace) or a softened claim that does not assert full crystallographic consistency.
  2. [Appendix D.1, Eqs. (10)-(11)] The variational bound in Proposition 4.1 assumes that the exact reverse bridge measure factorizes as Q_G times Q_{k|G} times Q_{S|G} times Q_A times Q_F. For crystalline asymmetric units, this factorization is questionable because the fractional coordinates F are constrained by the site symmetry S and the space group G: a Wyckoff position determines the support and symmetry constraints on F. If Eq. (10) fails, the KL decomposition in Eq. (11) omits non-negative mutual-information terms, so the claimed upper bound on the negative log-likelihood is not established. The proof simply states 'We assume...' without justification. This is load-bearing for the theoretical contribution; the authors should either prove the bound under a weaker assumption, add the missing terms, or reframe the objective as a heuristic approximation supported by empirical validation.
  3. [Algorithm 1 line 7 vs. Proposition 4.4] The reverse sampling step for the lattice parameters is inconsistent between Algorithm 1 and Proposition 4.4. Algorithm 1 line 7 gives k_{t-1} = m⊙( sqrt(alpha_bar_{t-1})/(1-alpha_bar_t)(1-alpha_t) times the predicted k_0 + m⊙ sqrt(alpha_t)/(1-alpha_bar_t)(1-alpha_bar_{t-1}) k_t + m_b + sigma_t epsilon ) + m_b, whereas Proposition 4.4 states the same expression but with the mask and bias of the target higher-symmetry space group G_{t-1} (denoted em and em_b) instead of the current group G_t. Using m in place of em and adding m_b instead of em_b means the sampled lattice is projected onto the wrong crystal family (the one being left, not the one being entered). If the implementation follows Algorithm 1, the method differs from the described model; if it follows Proposition 4.4, the pseudocode is incorrect. This must be fixed for reproducibility.
minor comments (5)
  1. [Figure 1] The notation in Figure 1 is confusing: the right panel uses 'g' and 'g' (presumably 'g' with a tilde) to denote the marginal site-symmetry priors of P2/m and P1, but the typography makes the distinction nearly invisible; please use a clearer notation such as g0 and g1.
  2. [Section 4.3 / Proposition 4.4] In the statement of Proposition 4.4, the definition of sigma_t contains mixed notation: it is written as sqrt((1-alpha_bar_{t-1})(1-alpha_t)/(1-alpha_bar_t)) with both alpha_t and alpha_bar_t in the same expression; please clarify the noise schedule notation consistently with the rest of the paper.
  3. [Algorithm 2, step 4] The sampling of the space-group trajectory in step 4 is ambiguous: 'G_t ∼ q_{t|0}(G_t|G_0) → G_{t+1} ∼ Π_t(G_t, G_{t+1})' suggests a two-stage draw, but it is not clear how this corresponds to the continuous-time jump-diffusion loss in Lemma 2.1. Please specify the exact distribution used for the pairs (G_t, G_{t+1}).
  4. [Table 1] The column header 'ASG' is not defined in the text; it appears to mean asymmetric-unit space-group representation, but this should be stated explicitly.
  5. [Appendix B.1] The list of oriented site-symmetry symbols is printed as a dense dictionary with integer values; presenting it as a table with the symbol names would aid readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SbCD learns the reverse of an explicitly constructed symmetry-breaking forward process, and its benchmarks are external; the flagged internal-consistency and factorization concerns are correctness gaps, not circular reductions.

full rationale

No load-bearing step in the derivation reduces to its own input. The forward processes (Propositions 4.2, 4.4, 4.5) are constructed corruption schedules, and the reverse rates and kernels are learned from data through the standard variational objectives of Campbell et al. and Austin et al.; the target metrics (structure/composition validity, coverage, S.U.N., CHGNet Ehull<0) are external benchmarks, and hyperparameters such as the transition rate lambda and loss weights are tuned, not fitted physical constants, so no fitted input is renamed as a prediction. Two caveats are correctness risks rather than circularity. First, the KL decomposition in Proposition 4.1 rests on an explicit product-of-marginals assumption stated in Appendix D.1, Eq. (10): 'We assume that the bridge measure factorizes according to the dependency structure of the crystal representation...' If the true posterior couples G, k, S, A, and F beyond this factorization, the bound is not a valid upper bound on NLL; this is an unverified modeling assumption, not an equivalence between input and output. Second, the central 'complete crystallographic specification' claim is weakened by the post-hoc Wyckoff retrieval in Appendix B.3 ('we compare the generated coordinate against all candidate Wyckoff positions... select the one whose symmetry-equivalent sites lie closest') and by the paper's own admission in Appendix E that 'enforcing such dynamically varying symmetry constraints on fractional coordinates remains a non-trivial open problem.' The evaluation does not check whether F0 lies on the declared Wyckoff manifold, so the headline capability may be vacuous or unsupported; but this is a soundness or validity gap, not a circular step. Because no equation or fitted parameter is equivalent to the claimed results by construction, the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central method adds one tuned hyperparameter (lambda) and several structural assumptions common to diffusion models; no new physical entity is introduced.

free parameters (1)
  • Space-group jump rate lambda = 1.5e-3 to 2.0e-3 depending on model (Tables 4, 5)
    Controls the timing of symmetry-breaking via the posterior holding time (Corollary 4.3); chosen by hyperparameter search (Appendix B.5), not derived from data.
assumptions (4)
  • domain assumption The exact reverse bridge measure factorizes according to the dependency structure of the crystal representation: Qhat_G, Qhat_{k|G}, Qhat_{S|G}, Qhat_A, Qhat_F.
    Used in Proposition 4.1 and Appendix D.1 Eq. (10) to decompose the KL; if untrue, the stated variational upper bound does not hold.
  • ad hoc to paper Every forward space-group jump goes directly to P1 with rate lambda(t), and P1 is absorbing.
    Proposition 4.2; the authors note in Section E that direct transitions are a simplification of real symmetry-breaking paths.
  • domain assumption Masking axioms: for higher-symmetry mask em and lower-symmetry mask m, em⊙m=em, m⊙mb=em⊙emb=0, and C mb=mb.
    Appendix D.4 Eq. (35); needed for the analytic lattice kernels. Properties hold for the six family masks but are not proven in general.
  • domain assumption The 81 oriented site-symmetry symbols inferred from all 230 space groups are sufficient for one-hot modeling.
    Appendix B.1; the encoding still leaves Wyckoff ambiguity, which is resolved by a nearest-site heuristic in Appendix B.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion." pith.science (2026). https://pith.science/paper/ZDMD2HY6

@misc{pith2026260813457,
  author       = {Pith},
  title        = {Pith review of: Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDMD2HY6}},
  note         = {Machine review of arXiv:2608.13457}
}
read the original abstract

Generating crystals has recently attracted significant interest due to their broad applications in materials science. However, existing generative models struggle to produce complete crystallographic specifications, limiting their ability to capture global symmetry and structural dependencies. In particular, current state-of-the-art approaches generate crystals only up to site symmetries and rely on sampling space groups from empirical distributions during generation. Inspired by \emph{spontaneous symmetry breaking} in physics, where crystals break symmetries under external conditions, we propose a novel diffusion-based framework that generates full structure specifications by reversing from the lowest-symmetry priors. Our method leverages a Markovian jump-diffusion process to model these symmetry-breaking dynamics, enabling it to traverse different space groups in a physically motivated manner. Our model, dubbed \emph{Symmetry-breaking Crystal Diffusion} (SbCD), introduces a principled approach to explicitly incorporate inter-space-group transitions into the generative process. In de novo generation experiments on MP20 and MPTS-52, SbCD outperforms its symmetry-preserving counterpart by a substantial margin, offering a promising perspective for generative modeling of crystalline materials.

Figures

Figures reproduced from arXiv: 2608.13457 by the authors.

Figure 1
Figure 1. SbCD Conceptual Framework. (Left, continuous state space) A symmetry-breaking diffusion process that undergoes a reduction in symmetry from a higher space group P2/m to a lower space group P1 at time step τ . The colored regions correspond to specific space-group constraints applied to the lattice component k. The upper panel depicts space-group transitions, while the lower panel visualizes lattice constraints for s… view at source ↗
Figure 2
Figure 2. Results on the Jensen–Shannon distance be [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 5
Figure 5. Stability (Ehull < 0) across dif￾ferent rate λ: SbCD versus SbCD♯ . 6 Conclusions We propose SbCD, a novel symmetry-breaking crystal diffusion model that enables sampling full crystallographic structure specifications from minimal symmetry assumptions. As such, SbCD marks an important step for de novo crystal generation. SbCD takes inspiration from spontaneous symmetry breaking and introduces novel adaptive space gr… view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Space group distributions of the MPTS-52 test set and crystals sampled by SbCD. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Space group distributions of the MP-20 test set and crystals sampled by SbCD. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Visualization of S.U.N. crystalline materials on MP-20, including their chemical formulas [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]
Figure 9
Figure 9. Figure 9: S.U.N. MP-20 materials discovered by SbCD. [PITH_FULL_IMAGE:figures/full_fig_p030_9.png]
Figure 10
Figure 10. Figure 10: S.U.N. MPTS-52 materials discovered by SbCD. [PITH_FULL_IMAGE:figures/full_fig_p031_10.png]
Figure 11
Figure 11. Figure 11: Comparison of the Ehull distributions on the MP-20 dataset. 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 Ehull per atom (eV) 0 100 200 300 400 Count Sta ble: Ehul <l 0 M eta sta ble: Ehul <l 0.1 SbCD SymmCD SGEquiDiff [PITH_FULL_IMAGE:figures/full_fig_p032_11.png]
Figure 12
Figure 12. Figure 12: Comparison of the Ehull distributions on the MPTS-52 dataset. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 17 canonical work pages

  1. [1]

    Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    Brian DO Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    1982

  2. [2]

    Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    2021

  3. [3]

    Sterling G Baird, Hasan M Sayeed, Joseph Montoya, and Taylor D Sparks. matbench-genmetrics: A python library for benchmarking crystal structure generative models using time-based splits of materials project structures.Journal of Open Source Software, 9(97):5618, 2024

    2024

  4. [4]

    Single crystal functional oxides on silicon.Nature communications, 7(1):10547, 2016

    Saidur Rahman Bakaul, Claudy Rayan Serrao, Michelle Lee, Chun Wing Yeung, Asis Sarker, Shang-Lin Hsu, Ajay Kumar Yadav, Liv Dedon, Long You, Asif Islam Khan, et al. Single crystal functional oxides on silicon.Nature communications, 7(1):10547, 2016

    work page 2016

  5. [5]

    A continuous time framework for discrete denoising models

    Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Informa- tion Processing Systems, 2022. URL https://openreview.net/forum?id=DmT862YAieY

    work page 2022

  6. [6]

    Space group equivariant crystal diffusion.arXiv preprint arXiv:2505.10994, 2025

    Rees Chang, Angela Pak, Alex Guerra, Ni Zhan, Nick Richardson, Elif Ertekin, and Ryan P Adams. Space group equivariant crystal diffusion.arXiv preprint arXiv:2505.10994, 2025

    arXiv 2025

  7. [7]

    Space group equivariant crystal diffusion.Advances in Neural Information Processing Systems, 38:72772–72805, 2026

    Rees Chang, Angela Pak, Alex Guerra, Ni Zhan, Nick Richardson, Elif Ertekin, and Ryan Adams. Space group equivariant crystal diffusion.Advances in Neural Information Processing Systems, 38:72772–72805, 2026

    work page 2026

  8. [8]

    Chemical synthesis and materials discovery.Nature Synthesis, 1(7):514–520, 2022

    Anthony K Cheetham, Ram Seshadri, and Fred Wudl. Chemical synthesis and materials discovery.Nature Synthesis, 1(7):514–520, 2022

    work page 2022

Show all 39 references
  1. [9]

    2d and 3d photonic crystal materials for photocatalysis and electrochemical energy storage and conversion.Science and Technology of advanced MaTerialS, 17(1):563–582, 2016

    Gillian Collins, Eileen Armstrong, David McNulty, Sally O’Hanlon, Hugh Geaney, and Colm O’Dwyer. 2d and 3d photonic crystal materials for photocatalysis and electrochemical energy storage and conversion.Science and Technology of advanced MaTerialS, 17(1):563–582, 2016

    2016

  2. [10]

    Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5(9):1031–1041, 2023

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J Bartel, and Gerbrand Ceder. Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5(9):1031–1041, 2023

    2023

  3. [11]

    Self-aggregating long-acting injectable microcrystals.Nature Chemical Engineering, 2(3):209–219, 2025

    Vivian R Feig, Sanghyun Park, Pier Giuseppe Rivano, Jinhee Kim, Benjamin Muller, Ashka Patel, Caroline Dial, Sofia Gonzalez, Hannah Carlisle, Flavia Codreanu, et al. Self-aggregating long-acting injectable microcrystals.Nature Chemical Engineering, 2(3):209–219, 2025

    2025

  4. [12]

    On the theory of stochastic processes, with particular reference to applications

    William Feller. On the theory of stochastic processes, with particular reference to applications

  5. [13]

    Pyxtal: A python library for crystal structure generation and symmetry analysis.Computer Physics Communications, 261: 107810, 2021

    Scott Fredericks, Kevin Parrish, Dean Sayre, and Qiang Zhu. Pyxtal: A python library for crystal structure generation and symmetry analysis.Computer Physics Communications, 261: 107810, 2021

    2021

  6. [14]

    Approximate accelerated stochastic simulation of chemically reacting systems.The Journal of chemical physics, 115(4):1716–1733, 2001

    Daniel T Gillespie. Approximate accelerated stochastic simulation of chemically reacting systems.The Journal of chemical physics, 115(4):1716–1733, 2001

    2001

  7. [15]

    Reidel Dordrecht, 1983

    Theo Hahn, Uri Shmueli, and JC Wilson Arthur.International tables for crystallography, volume 1. Reidel Dordrecht, 1983

    1983

  8. [16]

    Lie groups, lie algebras, and representations

    Brian C Hall. Lie groups, lie algebras, and representations. InQuantum Theory for Mathemati- cians, pages 333–366. Springer, 2013

    2013

  9. [17]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 10

    2020

  10. [18]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The materials project: A materials genome approach to accelerating materials innov...

    2013 doi

  11. [19]

    Crystal structure prediction by joint equivariant diffusion.Advances in Neural Information Processing Systems, 36:17464–17497, 2023

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal structure prediction by joint equivariant diffusion.Advances in Neural Information Processing Systems, 36:17464–17497, 2023

    2023

  12. [20]

    Space group constrained crystal generation.arXiv preprint arXiv:2402.03992, 2024

    Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group constrained crystal generation.arXiv preprint arXiv:2402.03992, 2024

    2024 arXiv

  13. [21]

    Wyckoff transformer: Generation of symmetric crystals

    Nikita Kazeev, Wei Nong, Ignat Romanov, Ruiming Zhu, Andrey E Ustyuzhanin, Shuya Yamazaki, and Kedar Hippalgaonkar. Wyckoff transformer: Generation of symmetric crystals. InForty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=eFHfRQRjJo

    2025

  14. [22]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  15. [23]

    SymmCD: Symmetry-preserving crystal generation with diffusion models

    Daniel Levy, Siba Smarak Panigrahi, Sékou-Oumar Kaba, Qiang Zhu, Kin Long Kelvin Lee, Mikhail Galkin, Santiago Miret, and Siamak Ravanbakhsh. SymmCD: Symmetry-preserving crystal generation with diffusion models. InThe Thirteenth International Conference on Learning Representat...

    2025

  16. [24]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  17. [25]

    Scaling deep learning for materials discovery.Nature, 624(7990):80–85, 2023

    Amil Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery.Nature, 624(7990):80–85, 2023

    2023

  18. [26]

    Flowmm: Generating materials with riemannian flow matching.arXiv preprint arXiv:2406.04713, 2024

    Benjamin Kurt Miller, Ricky TQ Chen, Anuroop Sriram, and Brandon M Wood. Flowmm: Generating materials with riemannian flow matching.arXiv preprint arXiv:2406.04713, 2024

    2024 arXiv

  19. [27]

    Symmetry relationships between related crystal structures

    Ulrich Müller. Symmetry relationships between related crystal structures. InSymmetry relations between related crystal structures, pages 156–178. Springer, 2011

    2011

  20. [28]

    BG Teubner, 1885

    Franz Neumann and Oskar Emil Meyer.Vorlesungen über die Theorie der Elasticität der festen Körper und des Lichtäthers, volume 5. BG Teubner, 1885

  21. [29]

    Miad: Mirage atom diffusion for de novo crystal generation.arXiv preprint arXiv:2511.14426, 2025

    Andrey Okhotin, Maksim Nakhodnov, Nikita Kazeev, Andrey E Ustyuzhanin, and Dmitry Vetrov. Miad: Mirage atom diffusion for de novo crystal generation.arXiv preprint arXiv:2511.14426, 2025

    2025 arXiv

  22. [30]

    Stochastic differential equations

    Bernt Øksendal. Stochastic differential equations. InStochastic differential equations: an introduction with applications, pages 38–50. Springer, 2003

    2003

  23. [31]

    Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.Computational Materials Science, 68:314–319, 2013

    Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L Chevrier, Kristin A Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.C...

    2013

  24. [32]

    Space group conditional flow matching

    Omri Puny, Yaron Lipman, and Benjamin Kurt Miller. Space group conditional flow matching. arXiv preprint arXiv:2509.23822, 2025

    2025

  25. [33]

    Matbench discovery–a framework to evaluate machine learning crystal stability predictions.arXiv preprint arXiv:2308.14920, 2023

    Janosh Riebesell, Rhys EA Goodall, Philipp Benner, Yuan Chiang, Bowen Deng, Alpha A Lee, Anubhav Jain, and Kristin A Persson. Matbench discovery–a framework to evaluate machine learning crystal stability predictions.arXiv preprint arXiv:2308.14920, 2023

    2023 arXiv

  26. [34]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015. 11

    2015

  27. [35]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  28. [36]

    An autonomous laboratory for the accelerated synthesis of inorganic materials.Nature, 624(7990):86, 2023

    Nathan J Szymanski, Bernardus Rendy, Yuxing Fei, Rishi E Kumar, Tanjin He, David Milsted, Matthew J McDermott, Max Gallant, Ekin Dogus Cubuk, Amil Merchant, et al. An autonomous laboratory for the accelerated synthesis of inorganic materials.Nature, 624(7990):86, 2023

    2023

  29. [37]

    Chapman and Hall/CRC, 2018

    Darren J Wilkinson.Stochastic modelling for systems biology. Chapman and Hall/CRC, 2018

    2018

  30. [38]

    Crys- tal diffusion variational autoencoder for periodic material generation.arXiv preprint arXiv:2110.06197, 2021

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crys- tal diffusion variational autoencoder for periodic material generation.arXiv preprint arXiv:2110.06197, 2021. 12 Algorithm 1Sampling of SbCD Require:Trained modelθ. 1:G T ←P1▷Lowest symmetry s...

    2021 arXiv

  31. [1949]

    URLhttps://api.semanticscholar.org/CorpusID:121027442

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.