Pith. sign in

REVIEW 4 major objections 6 minor 55 references

Dis-GEN: Disordered crystal structure generation

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Dis-GEN is the first generative model for disordered inorganic crystals, representing partial occupancies and vacancies directly at Wyckoff sites.

desk verdict Novel Wyckoff partial-occupancy representation is learnable, but the generation claim rests on an unspecified validity filter and weak property evidence. read the letter →

arxiv 2507.18275 v1 pith:PSZKEYPO submitted 2025-07-24 cond-mat.mtrl-sci cond-mat.dis-nn

Martin Hoffmann Petersen , Ruiming Zhu , Haiwen Dai , Savyasanchi Aggarwal , Nong Wei , Andy Paul Chen , Arghya Bhowmik , Juan Maria Garcia Lastra
show 1 more author
Kedar Hippalgaonkar
This is my paper · ORCID
classification cond-mat.mtrl-scicond-mat.dis-nn
keywords disorderedcrystalspartialoccupancyWyckoffsitescrystalstructuregenerationvariationalautoencodercompositionaldisordervacanciessymmetryequivariance
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

This paper introduces Dis-GEN, a variational autoencoder that learns disordered inorganic crystal structures from experimental entries in the largest crystal-structure database, and claims it is the first generative model able to produce such materials. Disordered crystals, where two or more atomic species share a crystallographic site or a site is partly vacant, are common in doped oxides, battery electrodes, and thermoelectrics, yet existing generative models assume every site is fully occupied. Dis-GEN represents each crystal by its space group, lattice parameters, and a matrix of Wyckoff sites whose rows encode partial occupancy, multiplicity, disorder status, fractional coordinates, and site letter, so both compositional disorder and vacancies can be sampled from a single average unit cell. The model reconstructs lattice parameters to about 0.06 Å and reaches 98.4 percent symmetry-consistency on the test set, with generated structures passing a 96 percent no-overlap validity filter; the authors also state that some generated compositions are chemically unreasonable.

What carries the argument

The load-bearing object is the Wyckoff-site representation: each column of the atomic matrix corresponds to one Wyckoff site, a group of symmetry-equivalent positions in the unit cell, and stores the partial-occupancy distribution, multiplicity, disorder flag, fractional coordinates, and Wyckoff letter. Because space groups fix which Wyckoff letters and multiplicities are possible, the model can check its own reconstructions for internal symmetry consistency, and the paper uses this symmetry-matching accuracy as a filter that removes symmetry-violating structures before validity screening. The authors describe the representation as empirically equivariant: global rotations or translations that would change the encoding are reflected in the reconstruction behavior, while physical predictions are invariant. The VAE encoder processes the atomic matrix with a convolutional network and the crystal vector with a linear network, and the decoder emits seven outputs with separate loss terms; the partial-occupancy loss is a squared error between occupancy distributions, which is what lets the model learn site disorder and vacancies.

What would settle it

Generate, say, one thousand structures that pass the symmetry and no-overlap filters, compute their energy above the convex hull at the generated compositions with a converged electronic-structure method, and compare with the paper's Zn-V-O result; if the overwhelming majority are thermodynamically unstable, the validity metrics used here do not establish physical plausibility.

Watch

Extended reading notes

Core claim

The central claim is that disordered crystals can be generated without enumerating supercells. Dis-GEN encodes each structure as a crystal vector holding six lattice parameters and a one-hot space group, together with an atomic matrix whose columns are Wyckoff sites; each column carries one-hot encoded partial occupancies, Wyckoff multiplicity, a disorder indicator, fractional coordinates, and a Wyckoff letter. A variational autoencoder with convolutional and linear layers compresses this representation and reconstructs it through seven task-specific losses, so partial occupancy is treated as a distribution over species rather than a single label. On a held-out test set the model reconstructs lattice parameters to within a few hundredths of an angstrom, recovers the space group with 99.4 percent accuracy, and achieves 98.4 percent symmetry-matching accuracy, meaning the predicted Wyckoff letters and multiplicities agree with the predicted space group. Sampling from a kernel-density estimate of the latent space yields generated structures with roughly 99 percent symmetry-matching accuracy and 96 percent validity; the authors show example generations and a composition-conditioned search in the Zn-V-O system, while noting that the generated phases in that system are thermodynamically unstable.

Load-bearing premise

The load-bearing premise is that a disordered crystal is adequately described by an average unit cell in which each Wyckoff site carries independent partial occupancies, and that a generated structure counts as physically plausible if it is symmetry-consistent and has no pair of atoms closer than 0.5 Å.

Editorial extensions

If this is right

  • If Dis-GEN works as claimed, a disordered material can be sampled as one average unit cell rather than as hundreds of supercells, changing the cost of exploring doped and vacancy-containing phases.
  • The Wyckoff-site representation is not tied to the variational autoencoder, so the same encoding could be used by diffusion or transformer generators to handle partial occupancies.
  • Composition-conditioned sampling from the latent space gives a concrete way to propose dopant configurations in chemistries that are sparse in experimental databases, even before properties are computed.
  • The reported reconstruction errors suggest the model produces CIF-like structures that are suitable inputs for downstream relaxation or property screening pipelines.

Reading between the lines

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

  • A straightforward stress test would be to train the same VAE on a supercell-enumerated encoding of the same experimental entries; if the average-cell model matches the supercell model in downstream property averages, the independent-partial-occupancy assumption is validated.
  • Because only about half of generated structures pass a charge-neutrality filter, coupling the generator with an electronegativity- or charge-based chemical filter is likely to be needed before the outputs are useful for discovery; the paper notes such filters do not yet exist.
  • The representation's success at capturing partial occupancies suggests vacancies could be treated as a fourth 'species' with an occupancy value, but the paper does not evaluate vacancy energetics, so that extension remains untested.
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

4 major / 6 minor

Summary. This paper introduces Dis-GEN, a variational autoencoder for crystalline inorganic materials that include compositional disorder, i.e., sites with partial occupancies. The representation encodes each Wyckoff site by one-hot occupancy, multiplicity, Wyckoff letter, fractional coordinates, and a disorder indicator, together with lattice parameters and a one-hot space group. The model is trained on 138,692 structures filtered from ICSD, roughly half of which contain partial occupancies. The authors report low reconstruction errors on a held-out test set (lattice MAE ≈ 0.06–0.10 Å, space-group accuracy 99.4%, Wyckoff-letter/multiplicity accuracy 99.5–99.8%, partial-occupancy MAE 0.06). They then sample the latent space using a Gaussian, GMM, or KDE estimator and evaluate the sampled structures with symmetry-matching accuracy, a structural-validity cutoff, and a charge-neutrality filter. Conditional generation is demonstrated for a Zn-V-O composition-targeted search, but the generated phases are found to be thermodynamically unstable.

Significance. The core reconstruction experiments are clean and non-circular: a held-out ICSD split is encoded and decoded, and the errors on lattice parameters, space group, Wyckoff letters, multiplicities, and partial occupancies are small. These results support the less ambitious claim that a Wyckoff-site-level representation of disordered crystals with partial occupancies can be learned with a VAE. The use of experimental ICSD data is a genuine advantage over models trained only on DFT-relaxed ordered structures, and the geometric insight of grouping symmetry-equivalent sites is appropriate. If the evaluation issues were resolved, the representation could be a useful building block for generative modeling of disordered materials. However, the manuscript's central novelty claim—that Dis-GEN 'effectively generates' disordered inorganic crystals—is not yet supported: the generation-success metric is ambiguous, the validity metric is undefined for partial occupancies, charge neutrality passes only about 55%, and the one quantitative downstream test (Zn-V-O) yields unstable phases. No code or data are provided at submission, so the quantitative claims are not independently checkable.

major comments (4)
  1. [4.3, Table 3] The structural-validity metric is undefined for the partial-occupancy representation, so the reported validity percentages cannot be interpreted. The manuscript adopts Ref. 3's criterion that no two atoms be closer than 0.5 Å, but in Dis-GEN's representation a disordered Wyckoff site has several atomic species at identical fractional coordinates (e.g., La and Sr at (0,0,0.25) in Fig. 1). Under the literal rule, such a site has zero interatomic distance and would always fail the check. The paper does not state how partial occupancies are treated before the cutoff is applied—whether all species are retained, occupancies are collapsed to a majority species, or a random configurational supercell is constructed. Without this specification, the 96.44% KDE validity and 97.53% test-set validity provide no evidence about the physical plausibility of the generated structures. Since this is the main 'validity' column in Table 3, the generation claim rests on an unevaluated filter.
  2. [4.2–4.3, Table 3] The main generation metric is ambiguous. Section 4.2 defines the generation error as the percentage of sampled structures discarded during the reconstruction process, which gives 99.94% for KDE sampling. The Table 3 caption, however, says the generation error corresponds to 'the accuracy, defined as the fraction of structures that passed each respective filter,' which would mean 99.94% of KDE samples passed. These definitions point in opposite directions, and the manuscript never reports the joint success rate of reconstruction × SMA × validity × charge neutrality. If the discard interpretation is intended, only about 0.06% of KDE samples survive reconstruction, and combining with the 55.28% charge-neutrality rate gives an end-to-end success below 0.05%, so 'effective generation' is not demonstrated. If the pass interpretation is intended, the term 'generation error' and the Section 4.2 definition are misstated. The paper must correct this inconsistency and report the joint success rate explicitly.
  3. [4.3, Table 3; 4.4] Charge neutrality and chemical reasonableness are not established. The charge-neutrality filter passes only 50–56% of structures for all estimators and the test set, meaning nearly half of the surviving generated structures are not charge-balanced. The paper's explanation that this is due to multi-valence species partially explains the difficulty, but it does not change the fact that the generation pipeline fails a basic chemical-plausibility screen. This is compounded by the admission in Section 4.4 that some generated compositions (e.g., Na7.92Mg8H3.76F10.88) 'appear chemically unreasonable.' These facts, together with the weak SMA check, mean the evidence supports reconstruction accuracy but not physically plausible generation of disordered crystals.
  4. [4.5, Fig. 6, Section 5] The conditional-generation case study undercuts the discovery claim. The Zn-V-O conditioned sampling produces structures that the paper itself reports are thermodynamically unstable: 'Both the ordered and disordered phases generated by Dis-GEN are found to be thermodynamically unstable' (Section 4.5). The attribution of this failure to the absence of property conditioning does not repair the issue; it confirms that the current model lacks the physical constraints needed for the proposed applications. The final claims in Section 5 that Dis-GEN 'enables crystal structure prediction for disordered crystals' and initiates 'systematic exploration of disordered inorganic crystals' are therefore too strong. The conclusions should be limited to the demonstrated representational learning and symmetry-consistent sampling, with conditional generation presented as a proof of concept that needs additional physical/chemical validation.
minor comments (6)
  1. [4.1] There is a typo in 'determined by the sapce group'—it should read 'space group.'
  2. [3.1] The phrase 'rare atoms with a periodic number higher than 100' should be 'atomic number' rather than 'periodic number.'
  3. [Appendix A, Appendix D, Acknowledgements] There are several typos: 'Kullback-Leiber' should be 'Kullback-Leibler,' 'utalized' should be 'utilized,' and 'acknowlegde' should be 'acknowledge.'
  4. [Figure 7 caption] The caption states that the decoder 'yields two two outputs'; this should read 'yields two outputs.'
  5. [4.1] The SMA metric is an internal consistency check between predicted Wyckoff letter, multiplicity, and space group; it should not be described as evidence of symmetry preservation without also comparing the predicted symmetry attributes to the input structure's known values.
  6. [Data and code accessibility] The statement that data and code 'will be public upon release' gives no repository or timeline, which makes it impossible to reproduce the reported numbers independently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out ICSD test split and independent symmetry constraints anchor the evaluation; self-citations are not load-bearing.

full rationale

Dis-GEN's derivation chain is not circular. The VAE is trained on a filtered ICSD split, and reconstruction errors (Tables 1-2) are computed on a held-out 20% test set; no test-set quantity is used to fit any model parameter or loss coefficient, so the reconstruction accuracy is an external benchmark. The SMA metric compares decoded outputs (space group, Wyckoff letter, multiplicity) against the crystallographic relation among these quantities, which is an independent mathematical constraint rather than a target taken from the input; although this is an internal consistency check rather than a ground-truth comparison, it is not a fit renamed as a prediction. The KDE/GMM/Gaussian latent-space evaluations are conventional generative-model sampling and are scored with the same held-out metrics; no sampled structure is fed back into training. Self-citations (Refs 26, 29, 9, 47) are used for background representations and prior data usage, not as the proof of Dis-GEN's central capability. The 0.5 Å validity criterion is potentially undefined for partial occupancies and the charge-neutrality filter is weak, but these are evaluation gaps affecting correctness, not circular reductions. The paper even reports chemically unreasonable compositions and thermodynamic instability, which is inconsistent with a design that forces success by construction.

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

The central claim rests on standard crystallography, the domain assumption that partial occupancies in an average cell are a sufficient representation, and the adequacy of internal consistency metrics. No new physical entities are introduced. The main free choices are the loss coefficients and the dataset filtering thresholds.

free parameters (4)
  • Loss coefficients (lambda_occ=2000, lambda_spg=10, lambda_lattice=3, lambda_disorder=0.1, others 1.0) = Table 4: occ 2000, spg 10, lattice 3, mult 1, letter 1, disorder 0.1, coord 1, KL 1
    Chosen by validation loss; the extreme lambda_occ=2000 signals difficulty fitting partial occupancy and is not fully explained.
  • Dataset filter thresholds = P1 excluded; Z>100 excluded; >9 Wyckoff sites excluded; multiplicity >50 excluded; >6 disordered sites excluded; >6…
    Ad hoc reductions to keep the representation manageable; these thresholds bound the model's coverage of ICSD and are partly justified by data distribution.
  • Latent dimension = 256
    Architecture hyperparameter, no ablation or justification.
  • KDE bandwidth = not reported
    KDE sampling of the latent space drives generation, but the bandwidth (or other KDE hyperparameters) is not stated.
assumptions (4)
  • standard math Space groups and Wyckoff sites provide a complete description of crystal symmetry.
    Invoked throughout Section 2; assumes standard crystallographic definitions from Ref 25.
  • domain assumption Disorder can be represented by average partial occupancies at Wyckoff sites in the unit cell.
    Section 2.2; this is the premise of the entire representation, but real disorder may require supercell configurational sampling, as acknowledged in Sections 4.3 and C.
  • domain assumption Reconstruction accuracy and internal symmetry consistency imply useful generation.
    Sections 4.1-4.3; generation quality is judged by SMA, validity, and charge neutrality, not by external property validation.
  • domain assumption ICSD entries are reliable experimental structures.
    Section 3.1; training on curated experimental data underlies the claim of experimental relevance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dis-GEN: Disordered crystal structure generation." pith.science (2026). https://pith.science/paper/PSZKEYPO

@misc{pith2026250718275,
  author       = {Pith},
  title        = {Pith review of: Dis-GEN: Disordered crystal structure generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSZKEYPO}},
  note         = {Machine review of arXiv:2507.18275}
}
read the original abstract

A wide range of synthesized crystalline inorganic materials exhibit compositional disorder, where multiple atomic species partially occupy the same crystallographic site. As a result, the physical and chemical properties of such materials are dependent on how the atomic species are distributed among the corresponding symmetrical sites, making them exceptionally challenging to model using computational methods. For this reason, existing generative models cannot handle the complexities of disordered inorganic crystals. To address this gap, we introduce Dis-GEN, a generative model based on an empirical equivariant representation, derived from theoretical crystallography methodology. Dis-GEN is capable of generating symmetry-consistent structures that accommodate both compositional disorder and vacancies. The model is uniquely trained on experimental structures from the Inorganic Crystal Structure Database (ICSD) - the world's largest database of identified inorganic crystal structures. We demonstrate that Dis-GEN can effectively generate disordered inorganic materials while preserving crystallographic symmetry throughout the generation process. This approach provides a critical check point for the systematic exploration and discovery of disordered functional materials, expanding the scope of generative modeling in materials science.

Figures

Figures reproduced from arXiv: 2507.18275 by the authors.

Figure 1
Figure 1. Structural representation of crystals. A Crystallographic information file of [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The data distribution after filtering 138,692 structures from the Inorganic [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Schematic of the Variational auto-encoder (VAE) used in this work. All the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: The latent space of the test set compared to a Multivariate Gaussian sampling, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Four generated disordered inorganic crystal structures, using Dis-GEN. All [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Phase space of the Zn–V–O chemical system, combining structures from the [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The architecture of the VAE model used for Dis-GEN with a batch size of 64. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: The VAE training curves, with curves color-coded uniformly across the plots. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: The lattice parameter prediction of the test set along with the disordered [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: The ordered and disordered Wyckoff site predictions of the test set along [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: Comparison between the symmetry required and the reconstructed Wyckoff [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Statistical distributions of selected properties of 1000 virtual cells of site [PITH_FULL_IMAGE:figures/full_fig_p026_12.png]
Figure 13
Figure 13. Figure 13: Three generated disordered inorganic crystal structures with space group 62, [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 21 canonical work pages

  1. [1]

    06197 REFERENCES 17

    Desiraju G R 2002 Nature materials 1 77–79 URL https://arxiv.org/abs/2110. 06197 REFERENCES 17

    work page 2002

  2. [2]

    Oganov A R, Pickard C J, Zhu Q and Needs R J 2019 Nature Reviews Materials 4 331–348 ISSN 2058-8437 URL http://dx.doi.org/10.1038/s41578-019-0101-8

    doi:10.1038/s41578-019-0101-8 2019

  3. [3]

    Xie T, Fu X, Ganea O E, Barzilay R and Jaakkola T 2022 ICLR URL https: //arxiv.org/abs/2110.06197

    arXiv 2022

  4. [4]

    Jiao R, Huang W, Lin P, Han J, Chen P, Lu Y and Liu Y 2023 Advances in Neural Information Processing Systems 36 17464–17497 URL https://arxiv.org/abs/ 2309.04475

    arXiv 2023

  5. [5]

    org/abs/2402.03992

    Jiao R, Huang W, Liu Y, Zhao D and Liu Y 2024 ICLR URL https://arxiv. org/abs/2402.03992

    arXiv 2024

  6. [6]

    Zeni C, Pinsler R, Z¨ ugner D, Fowler A, Horton M, Fu X, Wang Z, Shysheya A, Crabb´ e J, Ueda S, Sordillo R, Sun L, Smith J, Nguyen B, Schulz H, Lewis S, Huang C W, Lu Z, Zhou Y, Yang H, Hao H, Li J, Yang C, Li W, Tomioka R and Xie T 2025 Nature 639 624–632 ISSN 1476-4687 URL http://dx.doi.org/10.1038/ s41586-025-08628-5

    work page 2025

  7. [7]

    Cornet F, Bergamin F, Bhowmik A, Garcia Lastra J M, Frellsen J and N Schmidt M 2025 Kinetic langevin diffusion for crystalline materials generation AI for Accelerated Materials Design - ICLR 2025URL https://openreview.net/forum? id=Mttf1RoKKM

    work page 2025

  8. [8]

    Antunes L M, Butler K T and Grau-Crespo R 2024 Nature Communications 15 ISSN 2041-1723 URL http://dx.doi.org/10.1038/s41467-024-54639-7

    doi:10.1038/s41467-024-54639-7 2024

Show all 55 references
  1. [9]

    Kazeev N, Nong W, Romanov I, Zhu R, Ustyuzhanin A, Yamazaki S and Hippalgaonkar K 2025 Wyckoff transformer: Generation of symmetric crystals (arXiv) URL https://arxiv.org/abs/2503.02407

    2025

  2. [10]

    Miller B K, Chen R T, Sriram A and Wood B M 2024 Flowmm: Generating materials with riemannian flow matching Forty-first International Conference on Machine Learning URL https://arxiv.org/abs/2406.04713

    2024 arXiv

  3. [11]

    Sriram A, Miller B, Chen R T and Wood B 2024 Advances in Neural Information Processing Systems 37 46025–46046 URL https://arxiv.org/abs/2410.23405

    2024 arXiv

  4. [12]

    Kim S, Noh J, Gu G H, Aspuru-Guzik A and Jung Y 2020 ACS Central Science 6 1412–1420 ISSN 2374-7951 URL http://dx.doi.org/10.1021/acscentsci. 0c00426

    2020 doi

  5. [13]

    Smith A E, Mizoguchi H, Delaney K, Spaldin N A, Sleight A W and Subramanian M A 2009 Journal of the American Chemical Society 131 17084–17086 ISSN 1520- 5126 URL http://dx.doi.org/10.1021/ja9080666

    2009 doi

  6. [14]

    jpclett.3c02043

    Ning C, Ji Q, Wu Y, Wang J and Ju M G 2023 The Journal of Physical Chemistry Letters 14 8034–8042 ISSN 1948-7185 URL http://dx.doi.org/10.1021/acs. jpclett.3c02043

    2023 doi

  7. [15]

    Chu K, Zong W, Xue G, Guo H, Qin J, Zhu H, Zhang N, Tian Z, Dong H, Miao Y E, Roeffaers M B J, Hofkens J, Lai F and Liu T 2023 Journal of REFERENCES 18 the American Chemical Society 145 21387–21396 ISSN 1520-5126 URL http: //dx.doi.org/10.1021/jacs.3c06402

    2023 doi

  8. [16]

    Funke K, Banhatti R, Grabowski P, Nowinski J, Wrobel W, Dinnebier R and Magdysyuk O 2015 Solid State Ionics 271 2–9 ISSN 0167-2738 URL http: //dx.doi.org/10.1016/j.ssi.2014.09.033

    2015 doi

  9. [17]

    Wang H, LaLonde A D, Pei Y and Snyder G J 2012 Advanced Functional Materials 23 1586–1596 ISSN 1616-3028 URL http://dx.doi.org/10.1002/ adfm.201201576

    2012

  10. [18]

    Zhong P, Gupta S, Deng B, Jun K and Ceder G 2024 ACS Energy Letters 9 2775–2781 ISSN 2380-8195 URL http://dx.doi.org/10.1021/acsenergylett. 4c00799

    2024 doi

  11. [19]

    Wang Z, Du Z, Wang L, He G, Parkin I P, Zhang Y and Yue Y 2024 Nano Energy 121 109250 ISSN 2211-2855 URL http://dx.doi.org/10.1016/j.nanoen.2023. 109250

    2024 doi

  12. [20]

    Chang J H, Kleiven D, Melander M, Akola J, Garcia-Lastra J M and Vegge T 2019 Journal of Physics: Condensed Matter 31 325901 ISSN 1361-648X URL http://dx.doi.org/10.1088/1361-648X/ab1bbc

    2019 doi

  13. [21]

    Su T, Blankenau B J, Kim N, Krogstad J A and Ertekin E 2024Acta Materialia 276 120088 ISSN 1359-6454 URL http://dx.doi.org/10.1016/j.actamat.2024. 120088

    2024 doi

  14. [22]

    Yong A X B, Su T and Ertekin E 2024 Digital Discovery 3 1889–1909 ISSN 2635- 098X URL http://dx.doi.org/10.1039/D4DD00100A

    2024 doi

  15. [23]

    AI4Science M, Hernandez-Garcia A, Duval A, Volokhova A, Bengio Y, Sharma D, Carrier P L, Benabed Y, Koziarski M and Schmidt V 2023 arXiv preprint arXiv:2310.04925 URL https://arxiv.org/abs/2310.04925

    2023

  16. [24]

    Simon S H 2013 The Oxford solid state basics (OUP Oxford)

    2013

  17. [25]

    Souvignier B, Wondratschek H, Aroyo M I, Chapuis G and Glazer A M 2016 Space groups and their descriptions (International Union of Crystallography) p 42–74 ISBN 9780470974230 URL http://dx.doi.org/10.1107/97809553602060000922

    2016 doi

  18. [26]

    Zhu R, Nong W, Yamazaki S and Hippalgaonkar K 2024 Matter 7 3469–3488 ISSN 2590-2385 URL http://dx.doi.org/10.1016/j.matt.2024.05.042

    2024 doi

  19. [27]

    Hellenbrandt M 2004 Crystallography Reviews 10 17–22 ISSN 1476-3508 URL http://dx.doi.org/10.1080/08893110410001664882

    2004 doi

  20. [28]

    corsci.2023.111505

    Peyret D, Kaczorowski D, Skocic M, Tribollet B and Vivier V 2023 Corrosion Science 224 111505 ISSN 0010-938X URL http://dx.doi.org/10.1016/j. corsci.2023.111505

    2023

  21. [29]

    Zhu R, Tian S I P, Ren Z, Li J, Buonassisi T and Hippalgaonkar K 2023 ACS Omega 8 8210–8218 ISSN 2470-1343 URL http://dx.doi.org/10.1021/ acsomega.2c04856 REFERENCES 19

    2023

  22. [30]

    Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder G and Persson K A 2013 APL Materials 1 ISSN 2166-532X URL http://dx.doi.org/10.1063/1.4812323

    2013 doi

  23. [31]

    jpowsour.2022.231633

    Ahaliabadeh Z, Kong X, Fedorovskaya E and Kallio T 2022 Journal of Power Sources 540 231633 ISSN 0378-7753 URL http://dx.doi.org/10.1016/j. jpowsour.2022.231633

    2022

  24. [32]

    3389/fchem.2019.00500

    Chen J, Yang H, Li T, Liu C, Tong H, Chen J, Liu Z, Xia L, Chen Z, Duan J and Li L 2019 Frontiers in Chemistry 7 ISSN 2296-2646 URL http://dx.doi.org/10. 3389/fchem.2019.00500

    2019

  25. [33]

    Chen Y C 2017 Biostatistics and Epidemiology 1 161–187 ISSN 2470-9379 URL http://dx.doi.org/10.1080/24709360.2017.1396742

    2017

  26. [34]

    Reynolds D 2009 Gaussian Mixture Models (Springer US) p 659–663 ISBN 9780387730035 URL http://dx.doi.org/10.1007/978-0-387-73003-5_196

    2009 doi

  27. [35]

    Fredericks S, Parrish K, Sayre D and Zhu Q 2021 Computer Physics Communications 261 107810 ISSN 0010-4655 URL http://dx.doi.org/10.1016/ j.cpc.2020.107810

    2021

  28. [36]

    Levy D, Panigrahi S S, Kaba S O, Zhu Q, Lee K L K, Galkin M, Miret S and Ravanbakhsh S 2025 arXiv preprint arXiv:2502.03638 URL https://arxiv.org/ abs/2502.03638

    2025 arXiv

  29. [37]

    van der Maaten L and Hinton G 2008 Journal of Machine Learning Research 9 2579–2605 URL http://jmlr.org/papers/v9/vandermaaten08a.html

    2008

  30. [38]

    21105/joss.01361

    Davies D, Butler K, Jackson A, Skelton J, Morita K and Walsh A 2019 Journal of Open Source Software 4 1361 ISSN 2475-9066 URL http://dx.doi.org/10. 21105/joss.01361

    2019

  31. [39]

    Cui J, Liu J, Wang H, Ouyang L, Sun D, Zhu M and Yao X 2014 J. Mater. Chem. A 2 9645–9655 ISSN 2050-7496 URL http://dx.doi.org/10.1039/C4TA00221K

    2014 doi

  32. [40]

    Li Z, Xu H, Dong A, Qiu J, He L, Zhang T, Du D, Xing H, Zhu G, Wang D and Sun B 2021 Materials Characterization 173 110953 ISSN 1044-5803 URL http://dx.doi.org/10.1016/j.matchar.2021.110953

    2021

  33. [41]

    Weng G M, Simon Tam L Y and Lu Y C 2017 Journal of Materials Chemistry A 5 11764–11771 ISSN 2050-7496 URL http://dx.doi.org/10.1039/C7TA00482F

    2017 doi

  34. [42]

    Batatia I, Benner P, Chiang Y, Elena A M, Kov´ acs D P, Riebesell J, Advincula X R, Asta M, Avaylon M, Baldwin W J, Berger F, Bernstein N, Bhowmik A, Blau S M, C˘ arare V, Darby J P, De S, Della Pia F, Deringer V L, Elijoˇ sius R, El-Machachi Z, Falcioni F, Fako E, Ferrari A C...

    arXiv

  35. [43]

    1038/s42256-023-00716-3

    Deng B, Zhong P, Jun K, Riebesell J, Han K, Bartel C J and Ceder G 2023 Nature Machine Intelligence 5 1031–1041 ISSN 2522-5839 URL http://dx.doi.org/10. 1038/s42256-023-00716-3

    2023

  36. [44]

    Chen C and Ong S P 2022 Nature Computational Science 2 718–728 ISSN 2662- 8457 URL http://dx.doi.org/10.1038/s43588-022-00349-3

    2022 doi

  37. [45]

    Xie J Z, Zhou X Y, Jin B and Jiang H 2024 Journal of Chemical Theory and Computation 20 6207–6217 ISSN 1549-9626 URL http://dx.doi.org/10.1021/ acs.jctc.4c00463

    2024

  38. [46]

    D¨ ohn J and Gross A 2023 URL http://dx.doi.org/10.26434/ chemrxiv-2023-vj973

    2023

  39. [47]

    org/abs/2506.18542

    Dai H, McDermott M J, Chen A P, Recatala-Gomez J, Nong W, Zhu R, Thway M, Morris S, Sch¨ urmann C, Pethe S D, Zhang C, Saw W G, Tran B N, Mishra P, Wei F, Handoko A D, Hachmioune S, Shao H, Lin M, Liew C W, Persson K A and Hippalgaonkar K 2025 Data-driven design-test-make-anal...

    2025 arXiv

  40. [48]

    Batatia I, Kovacs D P, Simm G, Ortner C and Cs´ anyi G 2022 Advances in Neural Information Processing Systems 35 11423–11436 URL https://arxiv.org/abs/ 2206.07697

    2022 arXiv

  41. [49]

    Barroso-Luque L, Shuaibi M, Fu X, Wood B M, Dzamba M, Gao M, Rizvi A, Zitnick C L and Ulissi Z W 2024 URL https://arxiv.org/abs/2410.12771

    2024 arXiv

  42. [50]

    Kingma D P and Welling M 2013 URL https://arxiv.org/abs/1312.6114

    2013 arXiv

  43. [51]

    Kullback S and Leibler R A 1951 The Annals of Mathematical Statistics 22 79–86 ISSN 0003-4851 URL http://dx.doi.org/10.1214/aoms/1177729694

    1951

  44. [52]

    org/abs/1412.6980

    Kingma D P and Ba J 2014 arXiv preprint arXiv:1412.6980 URL https://arxiv. org/abs/1412.6980

    2014 arXiv

  45. [53]

    1016/0254-0584(96)80069-1

    Abdelghany A, Elsayed S, Abdelwahab D, El Ela A and Mousa N 1996 Materials Chemistry and Physics 44 277–280 ISSN 0254-0584 URL http://dx.doi.org/10. 1016/0254-0584(96)80069-1

    1996

  46. [54]

    Wu H j, Chen S w, Ikeda T and Jeffrey Snyder G 2012Acta Materialia 60 6144–6151 ISSN 1359-6454 URL http://dx.doi.org/10.1016/j.actamat.2012.07.057

    2012 doi

  47. [55]

    9b01294 REFERENCES 21 A

    Chen C, Ye W, Zuo Y, Zheng C and Ong S P 2019 Chemistry of Materials 31 3564–3572 ISSN 1520-5002 URL http://dx.doi.org/10.1021/acs.chemmater. 9b01294 REFERENCES 21 A. VAE Model specifics The total reconstruction loss function consists of 7 loss functions Lrecon, along with a K...

    2019 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.