Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

This paper argues that comparisons of generative crystal reconstruction models are confounded unless models receive equal information about the target, and provides a reusable benchmark showing that when they do not, measured rankings mirro

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

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

T0 review · deepseek-v4-flash

2026-08-04 09:17 UTC pith:ITHJNNWT

load-bearing objection The AtomBench harness and its info-access framing are genuinely useful, but the abstract and full text disagree on the headline result (MatterGen vs. AtomGPT) and the ranking is confounded by unequal information the authors largely admit — worth a serious referee, not acceptance as-is. the 4 major comments →

arxiv 2510.16165 v2 pith:ITHJNNWT submitted 2025-10-17 cs.LG cond-mat.supr-con

AtomBench: A Benchmarking Framework for Generative Crystal Reconstruction Models in Conventional Superconductors

Charles Rhys Campbell , Aldo H. Romero , Kamal Choudhary This is my paper
classification cs.LG cond-mat.supr-con
keywords generative modelscrystal structure reconstructionbenchmarkingsuperconductorsconditional generationinformation asymmetrydiffusion variational autoencoderflow matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper aims to establish a fair benchmarking protocol for generative crystal reconstruction models and to show that current architecture comparisons are confounded by unequal input information. It trains three models—a language model, a diffusion variational autoencoder, and a flow-matching model—on two DFT superconductivity datasets, each reconstructing held-out crystals from different amounts of information. Results show the diffusion model reconstructs lattices most accurately while the language model (and, per the abstract, a fourth model) reconstructs atomic coordinates best, and conditioning on critical temperature does not consistently help. The accuracy gap aligns with an information-theoretic ordering of how much each model sees about the target, and the authors release the harness for community use.

Core claim

The central discovery is that when generative models are asked to reconstruct crystals from a partially specified description, the amount of crystallographic information supplied at inference predicts reconstruction fidelity. The diffusion model receives the full structure through its encoder, the language model receives composition plus critical temperature, and the flow-matching model receives only composition; the conditional-entropy ordering H(M|full structure) < H(M|composition, Tc) < H(M|composition) matches the lattice-error ranking. The benchmark protocol, including Niggli-reduced cell comparison, KLD/MAE lattice metrics, and a new continuous per-structure coordinate metric, makes th

What carries the argument

The key mechanism is the explicit reconstruction-task definition coupled with the conditional-entropy inequality. By fixing the task as mapping a partial description (stoichiometry and optionally Tc) to a full crystal, and by measuring distribution-level divergence (KLD) and per-structure errors (MAE, RMSE, and the introduced ccRMSD) after Niggli reduction, the framework makes the information asymmetry visible and quantifiable.

Load-bearing premise

The ranking is treated as a comparison of architectures, but it assumes all models were trained and evaluated on the same held-out split of the same data with the same task; the paper only reports the seed and split for one model, and the models receive different input information by design.

What would settle it

Retrain the three models on identical train/test splits with matched conditioning (e.g., all given composition only, then all given composition plus Tc) and check whether the lattice and coordinate rankings change; if they do, the information-asymmetry explanation is confirmed; if they persist, the ranking reflects architecture rather than information.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Benchmarking studies should report and ideally match the conditioning information each generative model receives, otherwise architecture rankings are conflated with information advantage.
  • The released open-source package allows any inverse model that emits crystal reconstructions to be evaluated on standard superconductivity data, making results reproducible and directly comparable.
  • Conditioning on critical temperature, a natural property to steer generation, shows no consistent fidelity benefit, suggesting future models may not need it or need it implemented differently.
  • The continuous per-structure coordinate metric provides a finer-grained fidelity measure than distribution-level divergence alone, useful for spotting local geometric errors.

Where Pith is reading between the lines

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

  • If the information hierarchy explains the ranking, then the diffusion model's apparent lattice superiority is partly an artifact of receiving the full target structure; a truly fair architecture comparison would fix the input information across models.
  • A natural next experiment is to give all models identical conditioning (composition only, then composition plus Tc) and re-measure; the coordinate ranking may change, which would strengthen or weaken the paper's central claim.
  • The framework could be extended beyond superconductors to other property-conditioned structure problems, such as battery materials or thermoelectrics, where information asymmetry likely also confounds existing benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents AtomBench, a benchmarking framework for the crystal reconstruction task. It trains and evaluates AtomGPT, CDVAE, and FlowMM on the JARVIS Supercon-3D (1,058 structures) and Alexandria DS-A/B (8,253 entries; 825 held-out test structures) superconductivity datasets. Ten percent of each dataset is held out, and each model is asked to reconstruct lattice parameters and atomic coordinates from different conditioning information: CDVAE decodes from a full-structure latent, AtomGPT conditions on composition plus Tc, and FlowMM conditions on composition alone. Reconstruction fidelity is measured by KLD and MAE of six Niggli-reduced lattice parameters and by an average RMSE of atomic coordinates. The reported results are that CDVAE is most accurate on lattice KLD/MAE on both datasets, AtomGPT is most accurate on coordinate RMSE, and FlowMM is generally least accurate. The abstract additionally names MatterGen as the best coordinate reconstructor and introduces a ccRMSD metric, but the full text describes only the three models and does not define or report ccRMSD. The paper's Discussion acknowledges (Eq. 2) that the three models receive unequal amounts of target information and proposes matched-information experiments as future work.

Significance. If the empirical ranking were established, AtomBench would be a useful reusable evaluation harness for inverse crystal reconstruction. The paper ships a promising package with JARVIS-Leaderboard integration, uses Niggli reduction to canonicalize cells, and applies the benchmark to two curated superconductor datasets. However, the headline ranking is not currently established: the models are given different amounts of information about the target, and the paper's own Eq. (2) predicts the observed lattice ordering. In addition, the abstract and the full text report different best performers, and the ccRMSD metric from the abstract is absent from the body. These issues are fixable, and the framework's value would be considerably higher once the comparison is run under matched information and the abstract/body inconsistency is resolved.

major comments (4)
  1. [Results & Discussion, Eq. (2)] The central ranking is confounded by unequal information. CDVAE's encoder consumes the full structure (A,X,L), AtomGPT receives composition and Tc, and FlowMM receives composition only. The manuscript's Eq. (2) states H(M|f(A,X,L)) < H(M|A,Tc) < H(M|A) and notes that the observed lattice-error ordering matches this information ordering. CDVAE's lower lattice KLD/MAE can therefore be explained by its greater input information, not by architectural superiority. The Discussion acknowledges this ('Future work should supply each model with equivalent information prior to reconstruction'), yet the paper still frames the results as an architecture comparison. Please either re-frame the paper as an information-ablation study, or add experiments in which all models are conditioned on the same inputs (e.g., composition+Tc) and report the ranking under matched information.
  2. [Abstract vs. Results & Discussion] The abstract states that 'MatterGen achieves the best atomic-coordinate reconstruction' and introduces ccRMSD as a metric, but the full text reports experiments with only AtomGPT, CDVAE, and FlowMM; Figures 5-8 and the Results text do not mention MatterGen or ccRMSD. The Methods section defines only a normalized RMSE, not ccRMSD. As written, the paper's reported ranking and the abstract's ranking are inconsistent. Either add the MatterGen experiments and a precise definition of ccRMSD (including the matching procedure and normalization), or remove these claims from the abstract.
  3. [Methods, statistical metrics] The evaluation protocol is underspecified at load-bearing points. KLD is defined over 'histogrammed distributions' but the number of bins, bin width, histogram range, and smoothing are not given, so the KLD values in Figure 5 are not reproducible or comparable across models. The RMSE normalization 'by an appropriate structural length scale' is not specified, and the matching between predicted and reference atoms is not defined. In addition, the split procedure differs: Appendix Table 1 gives AtomGPT seed 3407 and test ratio 0.1, while Appendix Tables 2-3 give CDVAE/FlowMM splits of 0.8/0.1/0.1 with no seed or split-identity check. Without the same test split, the cross-model comparisons in Figures 5-8 may compare different held-out sets. Please specify all of these details and make the split generation reproducible.
  4. [Results & Discussion, Figures 5-8] No error bars, confidence intervals, or significance tests are reported for any KLD/MAE/RMSE value. Since each model/dataset combination is a single run, it is unclear whether the reported differences (e.g., 'AtomGPT and FlowMM produced comparable KLD values') are statistically meaningful. Report at least standard deviations over multiple seeds or bootstrap estimates, and state the test-set sizes explicitly in the Results text (the Alexandria count of 825 appears only in the Figure 3 caption).
minor comments (6)
  1. [Throughout] The dataset/infrastructure name is written inconsistently as 'JAR VIS' and 'JARVIS'; please pick one convention.
  2. [Introduction] The Introduction says 'we benchmark three inverse-design models' but the abstract says 'four models' (including MatterGen). Unify the count after resolving the abstract/body discrepancy.
  3. [Figure 1 caption] Typo: 'seperately' should be 'separately'.
  4. [Appendix, FlowMM] The heading 'FlowwMM' contains a typo.
  5. [Methods, Eq. (1)] The KLD formula uses a generic sum over x; please state that natural logarithms are used (units of nats) and specify how zero histogram bins are handled (e.g., smoothing or epsilon).
  6. [Data availability] The paper says the code 'will be made available' and gives a GitHub URL. Since reproducibility is a claimed contribution, please provide a working link and versioned release at the time of review.

Circularity Check

0 steps flagged

No significant circularity: the benchmark rankings rest on held-out test evaluation, with acknowledged information-confounding but no constructional reduction.

full rationale

The paper's central result is an empirical comparison of reconstruction models on held-out test splits (10% withheld) of two external DFT datasets, measured by KLD, MAE, and RMSE. No fitted parameter or training statistic is renamed as a prediction, and the models' outputs are evaluated against ground-truth structures that were not used for training. The main self-citations—AtomGPT, JARVIS datasets, and JARVIS-Leaderboard—reflect the authors' role as developers of some of the artifacts under comparison, but they do not supply the load-bearing argument: the ranking comes from the reported test-set metrics, not from a citation. Equation (2), the conditional-entropy inequality, is a standard information-theoretic statement and is used post hoc to explain why CDVAE's lattice errors are lowest given that it receives the full structure; the paper explicitly acknowledges this information imbalance as a confound and calls for matched-information experiments in future work. This is a validity limitation, not circularity. The abstract/body discrepancies (MatterGen named as best in the abstract but absent from the full-text experiments; ccRMSD mentioned in the abstract but not defined or used in the body) are reporting inconsistencies that undermine the stated conclusions, but they do not constitute a derivation that reduces to its own inputs. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The report depends on dataset labels, Niggli reduction, RMSE matching, and equivalent training/splits across models; none of these are specified precisely enough for an independent rerun at submission. The only new metric in the abstract, ccRMSD, is undefined in the full text, and code is promised but not shipped.

free parameters (4)
  • KLD histogram binning
    KLD values depend on the binning of lattice-parameter histograms; bin width, bin number, and treatment of empty bins are not specified.
  • RMSE normalization length scale
    The text states RMSE is normalized by a structural length scale but gives no formula; cross-model comparison requires identical normalization.
  • Train/test split seeds for CDVAE and FlowMM = not reported; only AtomGPT seed 3407
    If the models use different random 10% splits, they are not evaluated on the same structures, so rankings across models may reflect split identity rather than model quality.
  • Per-model training schedules = AtomGPT 2 epochs; CDVAE/FlowMM max 100 epochs, patience 5
    Unequal training budgets make the benchmark an evaluation of chosen configurations, not of architectures at comparable effort.
axioms (4)
  • domain assumption DFT-derived Tc labels in JARVIS Supercon-3D and Alexandria DS-A/B are reliable enough to serve as conditioning and evaluation targets.
    The benchmark treats Tc as a valid conditioning signal and a valid property; if the labels are systematically biased, the reconstruction results inherit that bias.
  • domain assumption Niggli reduction gives a canonical and unambiguous cell representation across models and ground truth.
    All lattice metrics are computed after Niggli reduction; non-unique or implementation-dependent reduction would make KLD/MAE comparisons unreliable.
  • domain assumption Atomic-coordinate RMSE is well-defined under some matching/alignment protocol.
    The text does not specify how predicted atoms are matched to reference atoms or how periodic images are handled; without this, RMSD values are ambiguous.
  • domain assumption All model checkpoints are converged representatives of their architectures.
    Fair comparison assumes the training runs are not under- or over-trained in a way that changes the ranking.
invented entities (1)
  • ccRMSD (continuous corrected RMSD) no independent evidence
    purpose: A continuous metric of local geometric fidelity defined for every test structure.
    The abstract introduces ccRMSD, but the full text neither defines nor reports it, so it cannot be computed or checked independently.

pith-pipeline@v1.3.0-alltime-deepseek · 12677 in / 19527 out tokens · 165391 ms · 2026-08-04T09:17:09.259414+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of AtomBench: A Benchmarking Framework for Generative Crystal Reconstruction Models in Conventional Superconductors." pith.science (2026). https://pith.science/paper/ITHJNNWT

@misc{pith2026251016165,
  author       = {Pith},
  title        = {Pith review of: AtomBench: A Benchmarking Framework for Generative Crystal Reconstruction Models in Conventional Superconductors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITHJNNWT}},
  note         = {Machine review of arXiv:2510.16165}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A key question in benchmarking generative crystal reconstruction models is how the amount and type of crystallographic information provided to a generative model affects its ability to reconstruct atomic structures. Yet such comparisons often overlook the fact that models receive unequal information about the target during reconstruction, thereby confounding architectural conclusions. We present AtomBench, an extensible, model-agnostic framework for comparing generative models on a well-defined crystal reconstruction task (rather than \textit{de novo} generation), which we here apply to conventional superconductors. We train and evaluate four models, AtomGPT, CDVAE, FlowMM, and MatterGen, on the JARVIS Supercon-3D and Alexandria DS-A/B datasets, grouping them by the information each accesses at inference. Reconstruction fidelity is measured by the Kullback-Leibler divergence (KLD) and mean absolute error (MAE) of lattice parameters and the root-mean-squared displacement (RMSD) of atomic coordinates. We further introduce the continuous corrected RMSD (ccRMSD), a continuous measure of local geometric fidelity defined for every structure in the test set. MatterGen achieves the best atomic-coordinate reconstruction, followed by AtomGPT, while CDVAE reconstructs lattices most accurately, and FlowMM is the least accurate but fastest overall. We find that conditioning on critical temperature T$_c$ does not consistently improve fidelity. We also release AtomBench as an open-source Python package that reproduces all reported reconstruction metrics, figures, and tables from one or more benchmark files and supports direct submission to the JARVIS-Leaderboard. Any inverse model emitting crystal reconstructions can be benchmarked with \texttt{atombench}, and we encourage community use. https://github.com/atomgptlab/atombench

Figures

Figures reproduced from arXiv: 2510.16165 by Aldo H. Romero, Charles Rhys Campbell, Kamal Choudhary.

Figure 1
Figure 1. Figure 1: Diagram showing the inverse model benchmarking study design. We compare [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Statistical comparison of the JARVIS Supercon-3D and Alexandria DS-A/B su [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reconstruction performance of AtomGPT, CDVAE, and FlowMM on the Alexan [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reconstruction performance of AtomGPT, CDVAE, and FlowMM on the JARVIS [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Kullback-Leibler Divergence in units of nats between the predicted and target [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean absolute error in units of angstroms between the predicted and target [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean absolute error in units of degrees between the predicted and target distribu [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Average root mean squared error in units of Angstroms between the predicted [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. VASP Plugins: Linking the Vienna ab-initio Simulation Package with Python

    cond-mat.mtrl-sci 2026-07 accept novelty 5.5

    A C++/pybind11 shared-memory plugin layer exposes VASP SCF and ionic data as NumPy arrays so Python can modify structure, forces, local potential, and occupancies in place.

Reference graph

Works this paper leans on

42 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Electron-phonon interactions from first principles

    Giustino, F. Electron-phonon interactions from first principles. Reviews of Modern Physics 2017, 89, 015003

    2017

  2. [2]

    Oliveira, L. N. d.; Gross, E.; Kohn, W. Density-functional theory for superconductors. Physical review letters 1988, 60, 2430

    1988

  3. [3]

    Ab initio theory of superconductivity

    L \"u ders, M.; Marques, M.; Lathiotakis, N.; Floris, A.; Profeta, G.; Fast, L.; Continenza, A.; Massidda, S.; Gross, E. Ab initio theory of superconductivity. I. Density functional formalism and approximate functionals. Physical Review B—Condensed Matter and Materials Physics 2005, 72, 024545

    2005

  4. [4]

    W.; Zhang, Y.; Lane, C.; Buda, I

    Furness, J. W.; Zhang, Y.; Lane, C.; Buda, I. G.; Barbiellini, B.; Markiewicz, R. S.; Bansil, A.; Sun, J. An accurate first-principles treatment of doping-dependent electronic structure of high-temperature cuprate superconductors. Communications Physics 2018, 1, 11

    2018

  5. [5]

    W.; Zhang, R.; Ning, J.; Barbiellini, B.; Markiewicz, R

    Pokharel, K.; Lane, C.; Furness, J. W.; Zhang, R.; Ning, J.; Barbiellini, B.; Markiewicz, R. S.; Zhang, Y.; Bansil, A.; Sun, J. Sensitivity of the electronic and magnetic structures of cuprate superconductors to density functional approximations. npj Computational Materials 2022, 8, 31

    2022

  6. [6]

    K.; Macridin, A.; Maier, T

    Kent, P.; Saha-Dasgupta, T.; Jepsen, O.; Andersen, O. K.; Macridin, A.; Maier, T. A.; Jarrell, M.; Schulthess, T. C. Combined density functional and dynamical cluster quantum Monte Carlo calculations of the three-band Hubbard model for hole-doped cuprate superconductors. Physical Review B—Condensed Matter and Materials Physics 2008, 78, 035132

    2008

  7. [7]

    M.; Ponc \'e , S

    Chen, S.; Wei, Y.; Monserrat, B.; Tomczak, J. M.; Ponc \'e , S. Impact of electronic correlations on the superconductivity of high-pressure CeH9. arXiv preprint arXiv:2507.12506 2025,

    arXiv 2025

  8. [8]

    Bandstructure meets many-body theory: theLDA+ DMFT method

    Held, K.; Andersen, O.; Feldbacher, M.; Yamasaki, A.; Yang, Y. Bandstructure meets many-body theory: theLDA+ DMFT method. Journal of Physics: Condensed Matter 2008, 20, 064202

    2008

  9. [9]

    U.; Tacchino, F.; Tavernelli, I.; Eckl, T

    Selisko, J.; Amsler, M.; Wever, C.; Kawashima, Y.; Samsonidze, G.; Haq, R. U.; Tacchino, F.; Tavernelli, I.; Eckl, T. Dynamical mean field theory for real materials on a quantum computer. arXiv preprint arXiv:2404.09527 2024,

    Pith/arXiv arXiv 2024

  10. [10]

    W.; Choudhary, A.; Agrawal, A.; Billinge, S

    Choudhary, K.; DeCost, B.; Chen, C.; Jain, A.; Tavazza, F.; Cohn, R.; Park, C. W.; Choudhary, A.; Agrawal, A.; Billinge, S. J.; others Recent advances and applications of deep learning methods in materials science. npj Computational Materials 2022, 8, 59

    2022

  11. [11]

    F.; Reid, A

    Choudhary, K.; Garrity, K. F.; Reid, A. C.; DeCost, B.; Biacchi, A. J.; Hight Walker, A. R.; Trautt, Z.; Hattrick-Simpers, J.; Kusne, A. G.; Centrone, A.; others The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design. npj computational materials 2020, 6, 173

    2020

  12. [12]

    The JARVIS infrastructure is all you need for materials design

    Choudhary, K. The JARVIS infrastructure is all you need for materials design. Computational Materials Science 2025, 259, 114063

    2025

  13. [13]

    Designing high-TC superconductors with BCS-inspired screening, density functional theory, and deep-learning

    Choudhary, K.; Garrity, K. Designing high-TC superconductors with BCS-inspired screening, density functional theory, and deep-learning. npj Computational Materials 2022, 8, 244

    2022

  14. [14]

    F.; Gupta, V.; Romero, A

    Choudhary, K.; Wines, D.; Li, K.; Garrity, K. F.; Gupta, V.; Romero, A. H.; Krogel, J. T.; Saritas, K.; Fuhr, A.; Ganesh, P.; others JARVIS-Leaderboard: a large scale benchmark of materials design methods. npj Computational Materials 2024, 10, 93

    2024

  15. [15]

    F.; DeCost, B.; Biacchi, A

    Wines, D.; Gurunathan, R.; Garrity, K. F.; DeCost, B.; Biacchi, A. J.; Tavazza, F.; Choudhary, K. Recent progress in the JARVIS infrastructure for next-generation data-driven materials design. Applied Physics Reviews 2023, 10

    2023

  16. [16]

    J.; Garrity, K

    Wines, D.; Choudhary, K.; Biacchi, A. J.; Garrity, K. F.; Tavazza, F. High-throughput DFT-based discovery of next generation two-dimensional (2D) superconductors. Nano letters 2023, 23, 969--978

    2023

  17. [17]

    Data-driven design of high pressure hydride superconductors using DFT and deep learning

    Wines, D.; Choudhary, K. Data-driven design of high pressure hydride superconductors using DFT and deep learning. Materials futures 2024, 3, 025602

    2024

  18. [18]

    F.; Romero, A

    Schmidt, J.; Cerqueira, T. F.; Romero, A. H.; Loew, A.; J \"a ger, F.; Wang, H.-C.; Botti, S.; Marques, M. A. Improving machine-learning models in materials science through large datasets. Materials Today Physics 2024, 48, 101560

    2024

  19. [19]

    F.; Fang, Y.-W.; Errea, I.; Ludwig, A.; Marques, M

    Sanna, A.; Cerqueira, T. F.; Fang, Y.-W.; Errea, I.; Ludwig, A.; Marques, M. A. Prediction of ambient pressure conventional superconductivity above 80 K in hydride compounds. npj Computational Materials 2024, 10, 44

    2024

  20. [20]

    F.; Wang, H.-C.; Botti, S.; Marques, M

    Gao, K.; Cui, W.; Cerqueira, T. F.; Wang, H.-C.; Botti, S.; Marques, M. A. Enhanced Superconductivity in X4H15 Compounds via Hole-Doping at Ambient Pressure. Advanced Science 2025, e08419

    2025

  21. [21]

    F.; Sanna, A.; Marques, M

    Cerqueira, T. F.; Sanna, A.; Marques, M. A. Sampling the materials space for conventional superconducting compounds. Advanced Materials 2024, 36, 2307085

    2024

  22. [22]

    K.; Dreizler, R

    Gross, E. K.; Dreizler, R. M. Density functional theory; Springer Science & Business Media, 2013; Vol. 337

    2013

  23. [23]

    Atomistic line graph neural network for improved materials property predictions

    Choudhary, K.; DeCost, B. Atomistic line graph neural network for improved materials property predictions. npj Computational Materials 2021, 7, 185

    2021

  24. [24]

    De Breuck, P.-P.; Wang, H.-C.; Rignanese, G.-M.; Botti, S.; Marques, M. A. Generative AI for Crystal Structures: A Review. arXiv preprint arXiv:2509.02723 2025,

    Pith/arXiv arXiv 2025

  25. [25]

    M.; Butler, K

    Antunes, L. M.; Butler, K. T.; Grau-Crespo, R. Crystal structure generation with autoregressive large language modeling. Nature Communications 2024, 15, 10570

    2024

  26. [26]

    Space group informed transformer for crystalline materials generation

    Cao, Z.; Luo, X.; Lv, J.; Wang, L. Space group informed transformer for crystalline materials generation. Science Bulletin 2025,

    2025

  27. [27]

    Crystal diffusion variational autoencoder for periodic material generation

    Xie, T.; Fu, X.; Ganea, O.-E.; Barzilay, R.; Jaakkola, T. Crystal diffusion variational autoencoder for periodic material generation. arXiv preprint arXiv:2110.06197 2021,

    Pith/arXiv arXiv 2021

  28. [28]

    Crystal structure prediction by joint equivariant diffusion

    Jiao, R.; Huang, W.; Lin, P.; Han, J.; Chen, P.; Lu, Y.; Liu, Y. Crystal structure prediction by joint equivariant diffusion. Advances in Neural Information Processing Systems 2023, 36, 17464--17497

    2023

  29. [29]

    K.; Hsu, J.; Macke, S.; Li, S.; Ham, J.; Liu, Z

    Miller, B. K.; Hsu, J.; Macke, S.; Li, S.; Ham, J.; Liu, Z. FlowMM: Generating Crystal Structures with Riemannian Flow Matching. arXiv\,2402.12345 2024,

    Pith/arXiv arXiv 2024

  30. [30]

    R.; Overly, L.; Kumar, A

    Choudhary, K.; Campbell, C. R.; Overly, L.; Kumar, A. AtomGPT: Generative Transformer Models for Atomic Structure Discovery. arXiv\,2309.12345 2023,

    Pith/arXiv arXiv 2023

  31. [31]

    Crystal Diffusion Variational Autoencoder

    Xie, T.; Geiger, M.; Friederich, P.; Batzner, S.; Kozinsky, B. Crystal Diffusion Variational Autoencoder. Proceedings of the International Conference on Learning Representations (ICLR). 2022

    2022

  32. [32]

    Designing high-TC superconductors with BCS-inspired screening, density functional theory, and deep-learning

    Choudhary, K.; Garrity, K. Designing high-TC superconductors with BCS-inspired screening, density functional theory, and deep-learning. npj Computational Materials 2022, 8

    2022

  33. [33]

    Cerqueira, T. F. T.; Fang, Y.-W.; Errea, I.; Sanna, A.; Marques, M. A. L. Searching Materials Space for Hydride Superconductors at Ambient Pressure. Advanced Functional Materials 2024, 34, 2404043

    2024

  34. [34]

    Choudhary, K. et al. JARVIS-Leaderboard: a large scale benchmark of materials design methods. npj Computational Materials 2024, 10

    2024

  35. [35]

    Niggli reduction and Bravais lattice determination

    Shi, H.-L.; Li, Z.-A. Niggli reduction and Bravais lattice determination. Journal of Applied Crystallography 2022, 55, 204–210

    2022

  36. [36]

    Kullback, S.; Leibler, R. A. On Information and Sufficiency. The Annals of Mathematical Statistics 1951, 22, 79--86

    1951

  37. [37]

    L.; Cococcioni, M.; Dabo, I.; others QUANTUM ESPRESSO: a modular and open-source software project for quantumsimulations of materials

    Giannozzi, P.; Baroni, S.; Bonini, N.; Calandra, M.; Car, R.; Cavazzoni, C.; Ceresoli, D.; Chiarotti, G. L.; Cococcioni, M.; Dabo, I.; others QUANTUM ESPRESSO: a modular and open-source software project for quantumsimulations of materials. Journal of physics: Condensed matter 2009, 21, 395502

    2009

  38. [38]

    P.; Ruzsinszky, A.; Csonka, G

    Perdew, J. P.; Ruzsinszky, A.; Csonka, G. I.; Vydrov, O. A.; Scuseria, G. E.; Constantin, L. A.; Zhou, X.; Burke, K. Restoring the density-gradient expansion for exchange in solids and surfaces. Physical Review Letters 2008, 100, 136406

    2008

  39. [39]

    Morel, P.; Anderson, P. W. Calculation of the Superconducting State Parameters with Retarded Electron-Phonon Interaction. Physical Review 1962, 125, 1263–1271

    1962

  40. [40]

    Jiang, A. Q. et al. Mistral 7B. 2023; https://arxiv.org/abs/2310.06825

    Pith/arXiv arXiv 2023

  41. [41]

    G.; Zitnick, C

    Gruver, N.; Sriram, A.; Madotto, A.; Wilson, A. G.; Zitnick, C. L.; Ulissi, Z. Fine-Tuned Language Models Generate Stable Inorganic Materials as Text. 2025; https://arxiv.org/abs/2402.04379

    Pith/arXiv arXiv 2025

  42. [42]

    Chen, R. T. Q.; Lipman, Y. Flow Matching on General Geometries. 2024; https://arxiv.org/abs/2302.03660 mcitethebibliography main.tex0000664000000000000000000026453115074511124011237 0ustar rootroot [journal=jacsat,manuscript=article] achemso [version=3] mhchem caption placeins xurl hyperref [usenames,dvipsnames] xcolor float gensymb amsmath multirow amssy...

    Pith/arXiv arXiv 2024