Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Symmetry-Aware Bayesian Flow Networks for Crystal Generation

T0 review · 4 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 This paper claims that a Bayesian Flow Network operating on a symmetry-aware crystal representation generates stable crystal structures at least 50 times faster than the next-best generative method while keeping stability and novelty…

desk verdict A useful first BFN for crystals with real sampling efficiency, but both headline claims—space-group fidelity and the 50x speedup—need qualification before they stand. read the letter →

arxiv 2502.03146 v2 pith:OZWYUJCK submitted 2025-02-05 cs.LG cond-mat.mtrl-sci

Laura Ruple , Luca Torresi , Henrik Schopmans , Pascal Friederich This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sci
keywords BayesianFlowNetworkscrystalstructuregenerationspacegroupsymmetryasymmetricunitproperty-conditionedmaterialsdiscoverygenerativemodelingsamplingefficiency
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 introduces SymmBFN, a Bayesian Flow Network adapted to generate crystal structures from a symmetry-aware representation. The central claim is that by generating the asymmetric unit, its site symmetries, and the space-group-constrained lattice within one BFN framework, the model produces structures whose space-group distribution matches experimentally observed crystals while remaining competitive with diffusion baselines on stability and novelty. Its headline result is efficiency: stable structures are generated at 0.059 seconds each, at least 50 times faster than the next-best method, with a stability rate of 11.8 percent and a S.U.N. rate of 8.9 percent. The paper further claims that conditioning the network on a target property, such as formation energy per atom, shifts the generated distribution toward that target, including targets outside the training range. If these results hold, Bayesian Flow Networks become a practical alternative to diffusion for accelerating materials discovery.

What carries the argument

The central object is the Bayesian Flow Network as applied to a symmetry-aware crystal representation. A BFN does not denoise samples; it iteratively updates the parameters of a factored input distribution, using isotropic Gaussians for continuous variables and categorical distributions for discrete variables, with a graph neural network providing context-dependent estimates of the true values. SymmBFN couples this to the asymmetric-unit representation: the lattice is written as an orthogonal matrix times the exponential of a symmetric matrix, the symmetric part is expanded in six basis matrices and masked according to the space group, and fractional coordinates, atom types, and site-symmetry tags are generated for the asymmetric unit. The full cell is then reconstructed by mapping atoms to the nearest Wyckoff positions using the crystallographic post-processing library. The same BFN machinery handles all four variable types, which removes the need for separate diffusion formulations and enables the fast 100-step sampling.

What would settle it

Re-train SymmBFN and the next-best baseline on MP-20 using the reported hyperparameters, then relax every generated structure with density-functional theory instead of the CHGNet surrogate and compare the per-stable-structure times and stability rates.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a BFN can jointly model the continuous variables (fractional coordinates and lattice parameters) and categorical variables (atom types and site symmetry operations) that make up a crystal, and that doing so in the asymmetry-aware representation delivers both fidelity and speed. Concretely, SymmBFN samples a space group, then iteratively refines parameters for the lattice vector, the coordinates and atom types in the asymmetric unit, and the site-symmetry tags; the full unit cell is reconstructed afterward via Wyckoff positions. In the MP-20 benchmark the model reaches 94.27 percent structural validity, 83.93 percent compositional validity, 99.73 percent coverage recall, and a space-group Jensen-Shannon distance of 0.080, close to symmetry-aware diffusion competitors, while using only 100 sampling steps and 0.007 seconds per sample. The paper attributes the speedup to the smaller computational graph of generating the asymmetric unit and the simpler BFN update rule, and reports property-conditioned generation that tracks requested formation energies.

Load-bearing premise

The results assume that all baseline models were run with equally optimized implementations and that the learned surrogate potentials used to estimate stability rank thermodynamic stability correctly; if the timings are implementation-unfair or the surrogates are biased, the reported 50x speedup and competitive stability rates will not reproduce.

Editorial extensions

If this is right

  • SymmBFN can generate stable, unique, and novel crystals at 100 sampling steps, cutting the per-stable-structure cost to 0.059 seconds, about 50 times below the next-best method reported.
  • Because it jointly models site symmetries and lattice constraints, SymmBFN reproduces the space-group distribution of experimentally observed crystals with a Jensen-Shannon distance of 0.080, unlike models that generate only in the low-symmetry P1 space group.
  • Conditioning on a scalar property such as formation energy per atom shifts the generated distribution toward the requested target, even for target values outside the training range, while retaining stability rates above 7 percent.
  • The BFN framework unifies continuous and categorical variables in one training objective, avoiding the need for different diffusion formulations for coordinates, lattice parameters, atom types, and site symmetries.

Reading between the lines

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

  • If the 50x speedup persists on larger unit cells and more diverse datasets, which the paper does not test, BFN-based generators could make million-sample screening campaigns practical on a single GPU, something the diffusion baselines cannot do at comparable cost.
  • The symmetry-aware asymmetric-unit representation is not tied to BFN; the same representation could plausibly be dropped into flow-matching or autoregressive backbones, potentially inheriting part of the speed advantage. This is an extension, not a claim of the paper.
  • A natural next test is whether conditioning on multiple properties, such as band gap, density, and formation energy simultaneously, works as cleanly as the single-property conditioning shown here; the paper only demonstrates one scalar target.
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

4 major / 5 minor

Summary. The paper introduces SymmBFN, a Bayesian Flow Network for crystal generation that operates on a symmetry-aware representation: it generates the asymmetric unit (fractional coordinates, atom types, site symmetries) and the lattice parameters in a unified BFN framework, then reconstructs the full unit cell via PyXtal. The authors evaluate on the MP-20 dataset against DiffCSP, DiffCSP++, SymmCD, FlowMM, and Crystal-text-LLM, reporting competitive proxy metrics, a low Jensen-Shannon distance for space-group distributions, a stability rate of 11.8%, and a claimed speedup of at least 50x in the cost of generating a stable structure. They also present a property-conditioned variant for formation energy.

Significance. If the claims hold, SymmBFN would be a meaningful contribution: it is one of the first BFN applications to crystals, uses a unified treatment of continuous and categorical variables, and reports competitive stability with far fewer sampling steps than diffusion baselines. The paper is thorough in describing the model, the loss functions, and the experimental protocol, and it benchmarks against several strong recent baselines. However, the two headline contributions are undermined by the evaluation protocol: (1) the space-group marginal is sampled directly from the training distribution, so the low jsd(G) is not a test of the model's ability to infer or reproduce space-group statistics, and (2) the reported generation time excludes the PyXtal-based reconstruction step that is unique to SymmBFN, and the asserted negligibility of that step is unmeasured and load-bearing for the 50x claim. These issues are fixable with additional measurements and a reframing of claims, but they currently overstate what is demonstrated.

major comments (4)
  1. [Section 3, Sampling; Table 2] The claim that SymmBFN 'accurately reproduces the distribution of space groups found in experimentally observed crystals' is largely forced by the sampling protocol. The Sampling paragraph states that the space group G is first sampled from the dataset distribution, so the marginal distribution of G in generated samples is the training distribution by construction. The jsd(G) metric in Table 2 therefore measures only whether the reconstructed structures preserve the conditioned space group after PyXtal reconstruction, not whether the model has learned the space-group distribution. This weakens contribution (ii) and the corresponding abstract claim. The authors should either evaluate space-group reproduction by sampling G from the model's own prior (if the model is meant to learn it) or explicitly reframe the result as 'space-group consistency under conditioning on the dataset distribution.'
  2. [Section 4, Timing metrics; Section 3, Sampling; Table 3] The 50x speedup claim rests on an unmeasured quantity. The text states that 'the post-processing time is negligible compared to the network calls and is therefore not included,' but for SymmBFN the post-processing includes a non-trivial, method-specific PyXtal step: searching the closest Wyckoff positions and replicating atoms to reconstruct the full cell. With the reported numbers, FlowMM's S cost is 2.957 s and SymmBFN's is 0.059 s, a ratio of about 50.1; adding even 0.0002 s per sample to SymmBFN's post-processing drops the ratio below 50. The diffusion baselines generate the full cell directly and do not have an analogous reconstruction step, so excluding this step specifically benefits SymmBFN. The authors must measure and report the reconstruction time, or reduce the claim to a range that is robust to the measured overhead.
  3. [Section 3, Lattice (masking step)] The masking step used to enforce space-group constraints on the lattice parameters is not derived from the BFN objective. The text says that 'after the Bayesian updates and network calls, we introduce a masking step equivalent to Table 1,' but the training loss in Eq. (7) is derived for the unmasked Bayesian update. Zeroing components of the parameter vector θk changes the posterior parameters and the output distribution without a corresponding modification of the loss; this could bias the learned distributions or violate the variational bound. The authors should either provide a derivation showing that the masked process remains a valid BFN, or present an ablation or empirical justification that the masking step is sound.
  4. [Section 4.2, Figure 3 and Table 6] The property-conditioned generation evaluation is descriptive rather than quantitative. The paper states that 'the model consistently generates structures with the desired formation energy per atom,' but Figure 3 only shows histograms, and Table 6 reports stability and metastability rates without any error metric between the target and the generated distribution (e.g., mean absolute error or calibration). With three targets and no comparison to an unconditional baseline or other conditional generators, the claim is not fully supported. The authors should add a quantitative measure of conditioning accuracy and ideally a baseline comparison.
minor comments (5)
  1. [Abstract and Conclusion] The phrase 'at least 50 times faster' is ambiguous: Table 3 shows a per-sample time ratio of about 39x (0.007 s vs 0.275 s) and a stable-structure cost ratio of about 50x. State explicitly which quantity is being claimed.
  2. [Section 4, Metrics paragraph] The sentence 'For all models, the post-processing time is negligible compared to the network calls and is therefore not included' is an assertion without supporting measurements; even if true for some baselines, it should be qualified and, for SymmBFN, verified.
  3. [Appendix A.3] The hyperparameter list is given, but the values of the loss weights λx, λS, λa, λk are presented without a sensitivity analysis; an ablation on these weights would help assess robustness.
  4. [Section 3, Equation (15)] The notation L∞(x) is reused for the continuous-time loss of fractional coordinates and again in Eq. (24) for the lattice; consider adding subscripts to avoid confusion.
  5. [Software and Data] The paper states that code will be released 'soon'; for a reproducibility-oriented venue, providing the code at the time of review would strengthen the evaluation.

Circularity Check

1 steps flagged · score 6.0 of 10

Space-group JSD is set by the sampling protocol, not learned; core efficiency and stability benchmarks remain independent.

  1. self definitional [Section 3, 'Sampling' paragraph; Section 4, 'Metrics']
    "To generate new structures, we first sample the space group G and, conditioned on G, the number of atoms in the asymmetric unit from the dataset distribution. ... we also compute the Jensen-Shannon distance between the space group distribution of the generated structures that pass the validity checks and that of the test set, to determine whether the models accurately capture the real-world distribution of space groups."

    The claimed ability to 'accurately reproduce the distribution of space groups' (jsd(G)=0.080) is evaluated on a quantity that the sampling procedure injects directly: G is sampled from the dataset distribution. The output marginal distribution of space groups is therefore approximately the input distribution by construction, not a learned model output. Validity filtering and SpacegroupAnalyzer classification with tolerance can perturb the measured distribution, so the match is not perfectly forced, but the low JSD is largely an artifact of the sampling protocol rather than evidence that SymmBFN infers symmetry statistics.

full rationale

The main circular step is the space-group distribution claim. Because the sampler draws G directly from the dataset distribution, the jsd(G) metric compares the input prior to the test-set marginal, so the match is substantially predetermined. This affects one stated contribution, but the paper's other headline results are not circular: the stability, S.U.N. rate, and timing comparisons against DiffCSP, FlowMM, SymmCD, and Crystal-text-LLM are empirical benchmarks against external baselines, and the BFN loss functions and network architecture are standard constructions with no fitted parameter renamed as a prediction. The speed claim excludes post-processing, which is a measurement concern rather than a definitional equivalence. Overall, partial circularity from the constructed space-group prior warrants score 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard BFN mathematics, prior crystallographic symmetry encodings, and surrogate ML potentials. The paper adds loss weights, accuracy schedule hyperparameters, and a dataset-driven sampling protocol that largely determines the space-group metric.

free parameters (5)
  • sigma (input std for lattice and fractional coordinates) = 0.02
    Set for the continuous BFN input distributions in Appendix A.3; affects coordinate precision and loss scale.
  • beta_a(1) (accuracy schedule for atom types) = 0.75
    Hyperparameter for the categorical BFN accuracy schedule for atom types.
  • beta_S(1) (accuracy schedule for site symmetries) = 2.0
    Hyperparameter for the categorical BFN accuracy schedule for symmetry operations.
  • loss weights lambda_x, lambda_S, lambda_a, lambda_k = 1, 10, 3, 0.1
    Weights balancing the coordinate, symmetry, atom-type, and lattice losses, chosen by hand in Appendix A.3.
  • number of sampling steps n = 100
    Chosen for the main results; the speed claim depends on this value, as shown in the ablation tables.
assumptions (5)
  • domain assumption BFN loss functions and Bayesian update rules from Graves et al. (2023) remain valid for crystal representations.
    Section 2.2 imports the BFN framework; the paper does not re-derive the variational bound for wrapped fractional coordinates and masked lattice parameters.
  • domain assumption The 15-axis/13-operation site symmetry representation and the Table 1 lattice constraints from Levy et al. and Jiao et al. are complete and correct.
    Section 2.1 and Table 1 rely on prior crystallographic representations without independent verification.
  • domain assumption CHGNet and M3GNet surrogate potentials approximate DFT energies well enough for stability and hull comparisons.
    Section 4 uses CHGNet for relaxation and Ehull and M3GNet for formation energies, without error analysis.
  • ad hoc to paper Masking lattice parameters after Bayesian updates and network calls preserves the BFN objective.
    Section 3 (Lattice) introduces masking equivalent to Table 1 but does not analyze its effect on the variational loss.
  • ad hoc to paper Sampling the space group from the dataset distribution is an acceptable way to evaluate space-group reproduction.
    Section 3 (Sampling) draws G from the dataset; the JSD(G) metric is therefore a test of reconstruction and validity, not of learned symmetry statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetry-Aware Bayesian Flow Networks for Crystal Generation." pith.science (2026). https://pith.science/paper/OZWYUJCK

@misc{pith2026250203146,
  author       = {Pith},
  title        = {Pith review of: Symmetry-Aware Bayesian Flow Networks for Crystal Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZWYUJCK}},
  note         = {Machine review of arXiv:2502.03146}
}
read the original abstract

The discovery of new crystalline materials is essential to scientific and technological progress. However, traditional trial-and-error approaches are inefficient due to the vast search space. Recent advancements in machine learning have enabled generative models to predict new stable materials by incorporating structural symmetries and to condition the generation on desired properties. In this work, we introduce SymmBFN, a novel symmetry-aware Bayesian Flow Network (BFN) for crystalline material generation that accurately reproduces the distribution of space groups found in experimentally observed crystals. SymmBFN substantially improves efficiency, generating stable structures at least 50 times faster than the next-best method. Furthermore, we demonstrate its capability for property-conditioned generation, enabling the design of materials with tailored properties. Our findings establish BFNs as an effective tool for accelerating the discovery of crystalline materials.

Figures

Figures reproduced from arXiv: 2502.03146 by the authors.

Figure 1
Figure 1. Visualisation of the 2D crystallographic point group 4mm. The symbols indicate different Wyck￾off positions and the dashed lines show the symmetry axes. The light blue triangle is the asymmetric unit. points x and x ′ are part of the same crystallographic orbit if there is a transformation (O, t) ∈ G with (O, t) x = x ′ . If the orbits of points x and x ′ have conjugate site sym￾metry groups Sx and Sx′ , they belong… view at source ↗
Figure 2
Figure 2. Bayesian Flow Network. Illustration of BFN training procedure at the upper panel (A) and of the sam￾pling procedure at the lower panel (B). h : θ ′ ← h(θ, y, α). The Bayesian update distribution is thus obtained by marginalizing out y: pU (θ ′ | θ, x; α) = E pS (y|x;α) δ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Results for the property-conditioned generation for three different target values. The histograms show the distributions of the formation energy per atom of the generated structures in blue and of the training set in orange. The dashed red line represents the target of the generation while the blue line is the mean formation energy per atom of the generated structures. respectively, in Sections A.2 and A.3 of the ap… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Structures generated by SymmBFN with the corresponding space group. Visualisations created with Crystal Toolkit (Horton et al., 2023). 13 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MiAD: Mirage Atom Diffusion for De Novo Crystal Generation

    cs.LG 2025-11 conditional novelty 6.0 of 10

    Mirage infusion lets crystal diffusion models vary atom counts during generation and raises the S.U.N. rate on MP-20 to 8.2%.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    M., Butler, K

    Antunes, L. M., Butler, K. T., and Grau-Crespo, R. (2024). Crystal structure generation with autoregressive large language modeling. Nature Communications , 15(1)

    work page 2024

  2. [2]

    I., Perez-Mato, J

    Aroyo, M. I., Perez-Mato, J. M., Capillas, C., Kroumova, E., Ivantchev, S., Madariaga, G., Kirov, A., and Wondratschek, H. (2006). Bilbao crystallographic server: I. databases and crystallographic computing programs. Zeitschrift f\" u r Kristallographie - Crystalline Materials , 221(1):15–27

    work page 2006

  3. [3]

    Bartel, C. J. (2022). Review of computational approaches to predict the thermodynamic stability of inorganic solids. Journal of Materials Science , 57(23):10475--10498

    work page 2022

  4. [4]

    and Ong, S

    Chen, C. and Ong, S. P. (2022). A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science , 2(11):718–728

    work page 2022

  5. [5]

    W., Butler, K

    Davies, D. W., Butler, K. T., Jackson, A. J., Skelton, J. M., Morita, K., and Walsh, A. (2019). Smact: Semiconducting materials by analogy and chemical theory. Journal of Open Source Software , 4(38):1361

    work page 2019

  6. [6]

    J., and Ceder, G

    Deng, B., Zhong, P., Jun, K., Riebesell, J., Han, K., Bartel, C. J., and Ceder, G. (2023). Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence , 5(9):1031–1041

    work page 2023

  7. [7]

    Fredericks, S., Parrish, K., Sayre, D., and Zhu, Q. (2021). Pyxtal: A python library for crystal structure generation and symmetry analysis. Computer Physics Communications , 261:107810

    work page 2021

  8. [8]

    K., Atkinson, T., and Gomez, F

    Graves, A., Srivastava, R. K., Atkinson, T., and Gomez, F. (2023). Bayesian flow networks. arXiv preprint arXiv:2308.07037

    arXiv 2023

Show all 35 references
  1. [9]

    G., Zitnick, C

    Gruver, N., Sriram, A., Madotto, A., Wilson, A. G., Zitnick, C. L., and Ulissi, Z. W. (2024). Fine-tuned language models generate stable inorganic materials as text. In The Twelfth International Conference on Learning Representations

    2024

  2. [10]

    Hahn, T., Shmueli, U., Wilson, A., and of Crystallography, I. U. (1984). International Tables for Crystallography . Number Bd. 1 in International Tables for Crystallography. D. Reidel Publishing Company

    1984

  3. [11]

    M., Guha, R., Huck, P., Li, H

    Horton, M., Shen, J.-X., Burns, J., Cohen, O., Chabbey, F., Ganose, A. M., Guha, R., Huck, P., Li, H. H., McDermott, M., Montoya, J., Moore, G., Munro, J., O'Donnell, C., Ophus, C., Petretto, G., Riebesell, J., Wetizner, S., Wander, B., Winston, D., Yang, R., Zeltmann, S., Jai...

    2023

  4. [12]

    and Enss, C

    Hunklinger, S. and Enss, C. (2023). Festkörperphysik . De Gruyter Oldenbourg, Berlin, Boston

    2023

  5. [13]

    P., Hautier, G., Chen, W., Richards, W

    Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., et al. (2013). Commentary: The materials project: A materials genome approach to accelerating materials innovation. APL materials , 1(1):011002

    2013

  6. [14]

    Jiao, R., Huang, W., Lin, P., Han, J., Chen, P., Lu, Y., and Liu, Y. (2023). Crystal structure prediction by joint equivariant diffusion. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  7. [15]

    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

  8. [16]

    S., Kaba, S.-O., Zhu, Q., Galkin, M., Miret, S., and Ravanbakhsh, S

    Levy, D., Panigrahi, S. S., Kaba, S.-O., Zhu, Q., Galkin, M., Miret, S., and Ravanbakhsh, S. (2024). Symm CD : Symmetry-preserving crystal generation with diffusion models. In AI for Accelerated Materials Design - NeurIPS 2024

    2024

  9. [17]

    S., Aykol, M., Cheon, G., and Cubuk, E

    Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G., and Cubuk, E. D. (2023). Scaling deep learning for materials discovery. Nature , 624(7990):80--85

    2023

  10. [18]

    K., Chen, R

    Miller, B. K., Chen, R. T. Q., Sriram, A., and Wood, B. M. (2024). Flow MM : Generating materials with riemannian flow matching. In Forty-first International Conference on Machine Learning

    2024

  11. [19]

    H., Mimona, M

    Mobarak, M. H., Mimona, M. A., Islam, M. A., Hossain, N., Zohura, F. T., Imtiaz, I., and Rimon, M. I. H. (2023). Scope of machine learning in materials research—a review. Applied Surface Science Advances , 18:100523

    2023

  12. [20]

    Müller, U. (2013). Symmetry Relationships Between Crystal Structures - Applications of Crystallographic Group Theory in Crystal Chemistry . OUP Oxford, New York, London

    2013

  13. [21]

    P., Richards, W

    Ong, S. P., Richards, W. D., Jain, A., Hautier, G., Kocher, M., Cholia, S., Gunter, D., Chevrier, V. L., Persson, K. A., and Ceder, G. (2013). Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis . Computational Materials Science , ...

    2013

  14. [22]

    Park, H., Li, Z., and Walsh, A. (2024). Has generative artificial intelligence solved inverse materials design? Matter , 7(7):2355–2367

    2024

  15. [23]

    Riebesell, J. (2023). Matbench discovery v1.0.0

    2023

  16. [24]

    G., Hoogeboom, E., and Welling, M

    Satorras, V. G., Hoogeboom, E., and Welling, M. (2022). E(n) equivariant graph neural networks

    2022

  17. [25]

    Schmidt, J., Hoffmann, N., Wang, H.-C., Borlido, P., Carri c o, P. J. M. A., Cerqueira, T. F. T., Botti, S., and Marques, M. A. L. (2022). Large-scale machine-learning-assisted exploration of the whole materials space

    2022

  18. [26]

    and Ermon, S

    Song, Y. and Ermon, S. (2019). Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems , 32

    2019

  19. [27]

    Song, Y., Gong, J., Zhou, H., Zheng, M., Liu, J., and Ma, W.-Y. (2023). Unified generative modeling of 3d molecules with bayesian flow networks. In The Twelfth International Conference on Learning Representations

    2023

  20. [28]

    K., Chen, R

    Sriram, A., Miller, B. K., Chen, R. T. Q., and Wood, B. M. (2024). Flow LLM : Flow matching for material generation with large language models as base distributions. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  21. [29]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. (2017). Attention is all you need. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R., editors, Advances in Neural ...

    2017

  22. [30]

    Ward, L., Agrawal, A., Choudhary, A., and Wolverton, C. (2016). A general-purpose machine learning framework for predicting properties of inorganic materials. npj Computational Materials , 2(1)

    2016

  23. [31]

    J., Pritzel, A., and Blundell, C

    Wirnsberger, P., Papamakarios, G., Ibarz, B., Racaniere, S., Ballard, A. J., Pritzel, A., and Blundell, C. (2022). Normalizing flows for atomic solids. Machine Learning: Science and Technology , 3(2):025009

    2022

  24. [32]

    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

  25. [33]

    Yang, S., Cho, K., Merchant, A., Abbeel, P., Schuurmans, D., Mordatch, I., and Cubuk, E. D. (2024). Scalable Diffusion for Materials Generation

    2024

  26. [34]

    Zeni, C., Pinsler, R., Z \"u gner, D., Fowler, A., Horton, M., Fu, X., Shysheya, S., Crabb \'e , J., Sun, L., Smith, J., et al. (2023). Mattergen: a generative model for inorganic materials design. arXiv preprint arXiv:2312.03687

    2023 arXiv

  27. [35]

    Zimmermann, N. E. R. and Jain, A. (2020). Local structure order parameters and site fingerprints for quantification of coordination environment and crystal structure similarity. RSC Adv. , 10:6063--6081

    2020

Pith tools

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