Pith. sign in

REVIEW 3 major objections 4 minor 35 references

Property-conditioned diffusion can steer crystal generation toward target magnetic and hard materials while favoring higher-symmetry structures.

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 ↗

An adapter-based classifier-free guidance framework steers a pre-trained crystal diffusion model toward target properties and higher symmetry, with MLIP-based screening reporting modest success rates for stable magnetic and hard materials.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid empirical study of CFG in crystal diffusion; the symmetry-guidance effect is a real contribution, but the discovery rates are only as good as the surrogate predictors and need DFT sanity checks. the 3 major comments →

arxiv 2607.21849 v1 pith:RTIS6COV submitted 2026-07-23 cond-mat.mtrl-sci

Property-Guided Diffusion for Inverse Design of Crystalline Materials

Sourav Mal , Subhankar Mishra , Prasenjit Sen This is my paper
classification cond-mat.mtrl-sci
keywords inverse materials designdiffusion modelsclassifier-free guidancecrystal structure generationsaturation magnetizationVickers hardnessmachine learning interatomic potentialthermodynamic stability
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 reading

The paper aims to show that a lightweight, pre-trained diffusion model for crystal structures can be repurposed for inverse design by adding small property-specific adapters and sampling with classifier-free guidance. On two tasks—formation energy paired with saturation magnetization, and formation energy paired with Vickers hardness—stronger guidance moves the generated property distributions onto the requested targets, narrows their spread, and reduces the share of low-symmetry P1 structures. The authors further argue that these generated crystals are not just realistic-looking: after geometric screening and relaxation with a machine-learning interatomic potential, a substantial fraction pass thermodynamic and dynamical stability checks and retain the requested functionality. If correct, this makes high-throughput, controllable inverse design of functional crystals practical.

Core claim

The central discovery is that classifier-free guidance strength acts as a dial for inverse crystal design: as the guidance scale w rises from 0 to 8, the generated distributions of formation energy, saturation magnetization, and Vickers hardness converge toward the prescribed targets, while the fraction of P1 (lowest-symmetry) structures falls and the fraction of higher-symmetry space groups rises. Applying a validation pipeline based on a machine-learning interatomic potential—geometric prescreening, structural relaxation, phonon calculations, and hull-distance evaluation—the framework reports overall success rates of 12.3% for stable magnetic materials (M_s ≥ 1 T) and 3.9% for stable hard

What carries the argument

The load-bearing machinery is the combination of (1) zero-initialized adapter modules inserted into every U-Net layer of the DiffCrysGen score network, which let the model learn property-conditioned scores without forgetting its unconditional pre-training; (2) classifier-free guidance, which at sampling time extrapolates the conditional score away from the unconditional score by a factor w, giving inference-time control over target adherence; and (3) an MLIP-based validation workflow that relaxes generated structures, computes phonons and hull distances, and checks target properties, converting raw generation into a claimed discovery rate.

Load-bearing premise

The entire reported success rate rests on the assumption that the surrogate property predictors and the machine-learning interatomic potential remain accurate on out-of-distribution generated crystals, so that the computed stabilities and target-property values reflect real physics rather than model extrapolation.

What would settle it

Take a random sample of the generated crystals that passed the MLIP-based validation in either design task, recompute their relaxed structures, phonons, hull distances, and target properties using density functional theory, and check what fraction still satisfy the stated success criteria; a large drop would falsify the claim that the framework 'identifies' physically viable materials at the reported rates.

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

If this is right

  • If the 12.3% and 3.9% overall success rates hold up, the framework turns a single pre-trained diffusion model into a multi-property inverse-design engine that can screen millions of candidates at low computational cost.
  • The finding that guidance reduces P1 output suggests that conditioning on physically meaningful target properties implicitly encodes crystallographic priors, so explicit symmetry constraints are not strictly necessary for realistic generation.
  • Because the adapters are property-specific and modular, the same pre-trained backbone can be extended to arbitrary continuous properties without retraining the full model.
  • The empirical rule of thumb that intermediate guidance scales (w≈4–6) give most of the target convergence with the best structural quality gives practitioners a practical operating point for future design campaigns.

Where Pith is reading between the lines

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

  • A natural next test is first-principles (DFT) verification of a random sample of the MLIP-validated candidates; if the true success rate is materially lower, the reported rates would reflect surrogate-predictor agreement rather than physical stability.
  • The symmetry-shift observation is likely not specific to magnets or hard materials; it should reproduce for any target property whose high-quality examples occupy ordered crystal systems, which could be checked by conditioning on properties like band gap or dielectric constant.
  • The reported throughput (~115 structures per second) suggests the adapter+CFG recipe could be combined with active learning, where MLIP validation feedback is used to re-weight training data, to push success rates higher in a closed loop.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 4 minor

Summary. The paper proposes property-guided DiffCrysGen, an adapter-based fine-tuning of a pretrained unconditional diffusion model combined with classifier-free guidance (CFG), for inverse design of crystalline materials. Two multi-property tasks are studied: co-conditioning on formation energy and saturation magnetization, and co-conditioning on formation energy and Vickers hardness. The authors systematically vary the guidance scale w from 0 to 8 and report that increasing guidance steers property distributions toward targets while reducing the fraction of P1 structures and increasing higher-symmetry structures. They also introduce an MLIP-based (MatterSim) validation pipeline involving geometric prescreening, relaxation, thermodynamic/dynamical stability checks, and property evaluation, and report overall success rates of 12.3% for magnetic materials and 3.9% for hard materials.

Significance. If the quantitative claims were fully supported, the framework would be a practically useful inverse-design engine: adapter fine-tuning is parameter-efficient, the CFG sweep across guidance scales is systematic, the reported sampling speed of ~115 structures/s is attractive for high-throughput work, and the MLIP-based multi-stage screening is a thoughtful validation design. The central methodological insight—that stronger guidance reduces low-symmetry P1 output while improving target convergence—is interesting and potentially general. However, the paper's headline discovery rates are model-based estimates: target-property satisfaction is evaluated with the same surrogate predictors used for conditioning, no DFT validation is provided, and no comparison with existing conditional generators is made. With additional validation and reframing, the framework could be a solid contribution; as it stands, the quantitative discovery claims outrun the evidence.

major comments (3)
  1. [Evaluation metrics and success criteria / Table I] The overall success rates (12.3% magnetic, 3.9% hard) are computed with the target-property criterion (M_s >= 1 T, H >= 10 GPa) evaluated using the same surrogate predictors used for conditioning: the M_s predictor from Ref. [14] and the H predictor trained in this work. CFG drives generation toward regions where these predictors output high values, so evaluating success with the same predictors partly measures systematic predictor bias rather than physical viability. No DFT validation of any generated candidate is provided; Figs. 7 and 12 show predicted values only. The rates should be relabeled as surrogate-based estimates, and a DFT-validated subset (or a clear uncertainty statement) is needed to support the discovery claim.
  2. [Evaluation metrics and success criteria (dynamical stability)] The threshold omega_min >= -20 cm^-1 accepts structures with substantial imaginary phonon modes as dynamically stable. This permissive tolerance directly inflates the reported dynamical-stability fractions (65.5% and 78.1%) and the overall success rates. Please report the fractions for omega_min >= 0 and omega_min >= -5 cm^-1, and justify the -20 cm^-1 cutoff quantitatively, e.g., through convergence tests or comparison with a stricter criterion.
  3. [Results and Discussion (baseline comparisons)] The manuscript claims the framework is 'efficient' and implies superior inverse-design performance, but no comparison is made with existing conditional generators such as MatterGen [18] or RL-aligned diffusion models [21,22] on the same tasks and metrics. The internal comparison with w=0 shows that guidance helps, but it does not establish the framework's relative efficiency or success rate against state-of-the-art conditional generators. Please add baseline experiments on the same validation workflow, or restrict the efficiency/superiority claims accordingly.
minor comments (4)
  1. [Abstract / Table I] The overall success rates 12.3% and 3.9% are fractions of the geometrically prescreened candidates (1,709 and 1,318), not of all generated structures. This should be stated in the abstract to avoid overstatement.
  2. [Conditioning on formation energy and Vickers hardness / Fig. 11] The text reports a mean RMSD of 0.61 Å for the h_form-H task, while the Fig. 11 caption gives 0.57 Å. Reconcile the inconsistency.
  3. [Throughout] There are several typographical and notation issues: 'upto' in the Methods fine-tuning description, garbled accents in 'Fr´echet', inconsistent use of 'Ms' vs 'M_s' in figures and text, and awkward line breaks in Eqs. (9) and (10). Please clean these up.
  4. [Methods / Reproducibility] No code or data availability statement is provided. For a methods-focused paper, releasing the fine-tuning code, the H-predictor training details, and generation/validation scripts would substantially strengthen reproducibility.

Circularity Check

0 steps flagged

No circularity: property conditioning and MLIP validation are separate; surrogate-based property scoring is a validation limitation, not a by-construction equivalence.

full rationale

The paper's derivation chain is not circular. The conditional model is fine-tuned on labeled datasets (DFT-derived h_form and M_s from Alexandria, and B/G-derived Vickers hardness from Materials Project), and CFG is implemented with a standard score extrapolation (Eq. 7). The reported property distributions and success rates are empirical measurements made with the stated surrogate predictors (h_form, M_s from prior work [14]; H developed here), not identities forced by the conditioning equation. The physical validation uses MatterSim, an external MLIP, for relaxation, E_hull, and phonons, which is independent of the generative model. The target-property satisfaction is evaluated with fitted surrogates rather than DFT, so the absolute success rates carry an accuracy risk if those surrogates are biased on out-of-distribution structures; however, that is a validation/benchmarking limitation, not a definitional or construction-based circularity. The paper itself frames the candidates as 'a promising pool for subsequent first-principles validation,' acknowledging the absence of DFT confirmation. No step in the derivation reduces to its own inputs or to an unverified self-citation chain.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims depend on a chain of fitted models (surrogate property predictors, MatterSim MLIP) and hand-chosen thresholds. The framework contributes no new physical entities; the primary numerical outputs (success rates) are functions of these fitted components.

free parameters (6)
  • H surrogate predictor weights = fitted NN (not disclosed)
    Developed in this work; used to compute Vickers hardness for generated and relaxed structures, so the H≥10 GPa success rate depends on this fitted model.
  • M_s surrogate predictor weights = fitted NN (from prior work [14])
    Used to evaluate magnetization of generated structures; the M_s≥1 T success rate depends on this fitted model.
  • h_form surrogate predictor weights = fitted NN (from prior work [14])
    Used to measure generated formation energy distributions in the steering analysis (Figs 3, 8).
  • dynamical stability criterion (ω_min threshold) = -20 cm^-1
    Hand-chosen tolerance for MLIP phonons; directly affects the reported dynamical-stability percentage and overall success rate.
  • E_hull stability threshold = 0.1 eV/atom
    Hand-chosen; defines thermodynamic stability and success; the pre-training data are also filtered by E_hull ≤ 0.5 eV/atom.
  • minimum interatomic distance cutoff = 0.5 Å
    Hand-chosen geometric prescreening threshold; defines the candidate pool and thus the denominator of all success-rate percentages.
axioms (5)
  • domain assumption MatterSim MLIP predictions (total energies, forces, phonons) approximate DFT sufficiently for stability screening.
    The entire physical validation uses MatterSim rather than DFT; if MatterSim is inaccurate for novel structures, E_hull, ω_min, and relaxed structures are unreliable. Invoked in the validation workflow section.
  • domain assumption Surrogate property predictors generalize to out-of-distribution generated structures.
    The M_s, h_form, and H models are used to evaluate target-property satisfaction; if they do not generalize, the steering and success-rate conclusions are invalid. Invoked in the evaluation metrics section.
  • domain assumption The empirical hardness formula (Mazhnik-Oganov, Eq. 9) and the elastic-moduli relations (Eq. 10) are valid for the target chemical space.
    Used to define the H labels for fine-tuning and the H target; if the formula is invalid, the hardness objective is mis-specified.
  • domain assumption Materials Project reference energies for E_hull are compatible with MatterSim energies.
    Phase diagrams mix MP entries with MatterSim-predicted h_form; systematic offsets between the two energy scales would bias E_hull values.
  • standard math Standard score-matching and diffusion sampling assumptions (Karras et al.) hold for the crystal representation.
    The reverse diffusion and CFG extrapolation rely on standard theory; no new mathematical derivation is provided.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Property-Guided Diffusion for Inverse Design of Crystalline Materials." pith.science (2026). https://pith.science/paper/RTIS6COV

@misc{pith2026260721849,
  author       = {Pith},
  title        = {Pith review of: Property-Guided Diffusion for Inverse Design of Crystalline Materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTIS6COV}},
  note         = {Machine review of arXiv:2607.21849}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion-based generative models with property guidance have emerged as a promising paradigm for inverse materials design by enabling the generation of crystalline materials with user-specified target properties. However, despite recent advances, the effectiveness of property guidance, its influence on crystallographic symmetry, and the physical viability of generated materials remain poorly understood. To address these questions, we develop a property-guided framework based on the lightweight diffusion model DiffCrysGen using parameter-efficient adapter fine-tuning and classifier-free guidance (CFG). The resulting framework enables efficient multi-property crystal generation while preserving the knowledge learned during unconditional pre-training. Using formation energy together with saturation magnetization and Vickers hardness as representative inverse-design tasks, we systematically investigate the influence of CFG across a broad range of guidance strengths. Increasing the guidance scale progressively steers the generated property distributions toward the prescribed targets while reducing the fraction of lowest-symmetry ($P1$) structures and increasing the proportion of higher-symmetry structures. To evaluate physical viability, generated structures are geometrically prescreened and subsequently validated using a machine-learning interatomic potential (MLIP)-based workflow comprising structural relaxation and thermodynamic, dynamical, and property-specific analyses. The framework identifies thermodynamically and dynamically stable magnetic and mechanically hard materials with overall success rates of 12.3\% and 3.9\%, respectively. These results establish property-guided DiffCrysGen as an efficient framework for inverse materials design while providing new insights into the role of classifier-free guidance in crystal generation.

Figures

Figures reproduced from arXiv: 2607.21849 by Prasenjit Sen, Sourav Mal, Subhankar Mishra.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10 [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11 [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12 [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 2 canonical work pages

  1. [1]

    M. K. Horton, P. Huck, R. X. Yang, J. M. Munro, S. Dwarak- nath, A. M. Ganose, R. S. Kingsbury, M. Wen, J. X. Shen, T. S. Mathis, A. D. Kaplan, K. Berket, J. Riebesell, J. George, A. S. Rosen, E. W. C. Spotte-Smith, M. J. McDermott, O. A. Cohen, A. Dunn, M. C. Kuner, G.-M. Rignanese, G. Petretto, D. Waro- quiers, S. M. Griffin, J. B. Neaton, D. C. Chrzan,...

    2025

  2. [2]

    Curtarolo, W

    S. Curtarolo, W. Setyawan, G. L. Hart, M. Jahnatek, R. V . Chep- ulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy, M. J. 12 (a) 2 1 0 1 2 DFT hform (eV/atom) 0 200 400 600 800 1000 1200 1400Count = 0.75 = 0.89 Fine-tuning dataset 2 1 0 1 2 Pred. hform (eV/atom) 0 20 40 60 80 100 120 140 160 w=0 =-1.08 =0.84 2 1 0 1 2 Pred. hform (eV/atom) w=2 =-0.9...

    2000

  3. [3]

    Curtarolo, G

    S. Curtarolo, G. L. W. Hart, M. B. Nardelli, N. Mingo, S. San- vito, and O. Levy, The high-throughput highway to computa- tional materials design, Nature Materials12, 191 (2013)

    2013

  4. [4]

    J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolver- ton, Materials design and discovery with high-throughput den- sity functional theory: the open quantum materials database (oqmd), JOM65, 1501 (2013)

    2013

  5. [5]

    Zunger, Inverse design in search of materials with target functionalities, Nature Reviews Chemistry2, 0121 (2018)

    A. Zunger, Inverse design in search of materials with target functionalities, Nature Reviews Chemistry2, 0121 (2018)

    2018

  6. [6]

    Metni, L

    H. Metni, L. Ruple, L. N. Walters, L. Torresi, J. Teufel, H. Schopmans, J. streicher, Y . Zhang, M. Neubert, Y . Koide, K. Steiner, P. Link, L. Br, M. Petrova, G. Ceder, and P. Friederich, Generative models for crystalline materials, Ad- vanced Materials38, e23620 (2026)

    2026

  7. [7]

    Cheng, C.-L

    M. Cheng, C.-L. Fu, R. Okabe, A. Chotrattanapituk, A. Boonkird, N. T. Hung, and M. Li, Artificial intelligence- driven approaches for materials design and discovery, Nature Materials25, 174 (2026)

    2026

  8. [8]

    De Breuck, H.-C

    P.-P. De Breuck, H.-C. Wang, G.-M. Rignanese, S. Botti, and M. A. L. Marques, Generative ai for crystal structures: a review, npj Computational Materials11, 370 (2025)

    2025

  9. [9]

    Z. Chen, Z. Meng, T. He, H. Li, J. Cao, L. Xu, H. Xiao, Y . Zhang, X. He, and G. Fang, Crystal structure prediction meets artificial intelligence, The Journal of Physical Chemistry Letters16, 2581 (2025)

    2025

  10. [10]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion probabilistic models (2020), arXiv:2006.11239 [cs.LG]

    Pith/arXiv arXiv 2020

  11. [11]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Er- mon, and B. Poole, Score-based generative modeling through stochastic differential equations (2021), arXiv:2011.13456 [cs.LG]

    Pith/arXiv arXiv 2021

  12. [12]

    T. Xie, X. Fu, O.-E. Ganea, R. Barzilay, and T. Jaakkola, Crystal diffusion variational autoencoder for periodic material genera- tion (2022), arXiv:2110.06197 [cs.LG]

    Pith/arXiv arXiv 2022

  13. [13]

    R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y . Lu, and Y . Liu, Crystal structure prediction by joint equivariant diffu- sion (2024), arXiv:2309.04475 [cond-mat.mtrl-sci]

    Pith/arXiv arXiv 2024

  14. [14]

    S. Mal, N. Ahmed, J. Jami, S. Mishra, and P. Sen, Diffcrys- gen: a generative diffusion model for accelerated design of inorganic crystalline materials, npj Computational Materials 10.1038/s41524-026-02147-1 (2026)

    work page doi:10.1038/s41524-026-02147-1 2026

  15. [15]

    R. Jiao, W. Huang, Y . Liu, D. Zhao, and Y . Liu, Space group constrained crystal generation (2024), arXiv:2402.03992 [cs.LG]

    Pith/arXiv arXiv 2024

  16. [16]

    D. Levy, S. S. Panigrahi, S.-O. Kaba, Q. Zhu, K. L. K. Lee, M. Galkin, S. Miret, and S. Ravanbakhsh, Symmcd: Symmetry- 13 15 20 25 30 35 P1 Rate (%) (a) 44 46 48 50 52High SPG Rate (%) (b) 0 2 4 6 8 Guidance Scale 98.75 99.00 99.25 99.50 99.75 100.00 hform 0 Rate (%) (c) 0 2 4 6 8 Guidance Scale 10 20 30 40 50 H 10 GPa Rate (%) (d) FIG. 9.Effect of classi...

    Pith/arXiv arXiv 2025

  17. [17]

    F. E. Kelvinius, O. B. Andersson, A. S. Parackal, D. Qian, R. Armiento, and F. Lindsten, Wyckoffdiff – a generative dif- fusion model for crystal symmetry (2025), arXiv:2502.06485 [cond-mat.mtrl-sci]

    arXiv 2025

  18. [18]

    C. Zeni, R. Pinsler, D. Z ¨ugner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabb´e, S. Ueda, R. Sordillo, L. Sun, J. Smith, B. Nguyen, H. Schulz, S. Lewis, C.-W. Huang, Z. Lu, Y . Zhou, H. Yang, H. Hao, J. Li, C. Yang, W. Li, R. Tomioka, and T. Xie, A generative model for inorganic materials design, Nature639, 624 (2025)

    2025

  19. [19]

    H. Park, A. Onwuli, and A. Walsh, Exploration of crystal chem- ical space using text-guided generative artificial intelligence, Nature Communications16, 4379 (2025)

    2025

  20. [20]

    K. Das, S. Khastagir, P. Goyal, S.-C. Lee, S. Bhattacharjee, and N. Ganguly, Periodic materials generation using text-guided joint diffusion model (2025), arXiv:2503.00522 [cs.LG]

    Pith/arXiv arXiv 2025

  21. [21]

    Park and A

    H. Park and A. Walsh, Guiding generative models to uncover diverse and novel crystals via reinforcement learning, Nature Machine Intelligence 10.1038/s42256-026-01262-4 (2026)

    work page doi:10.1038/s42256-026-01262-4 2026

  22. [22]

    J. Chen, J. Guo, E. Fako, and P. Schwaller, Accelerating inverse materials design using generative diffusion models with rein- forcement learning (2025), arXiv:2511.03112 [physics.chem- ph]

    arXiv 2025

  23. [23]

    Ho and T

    J. Ho and T. Salimans, Classifier-free diffusion guidance (2022), arXiv:2207.12598 [cs.LG]

    Pith/arXiv arXiv 2022

  24. [24]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox, U-net: Convolu- tional networks for biomedical image segmentation (2015), arXiv:1505.04597 [cs.CV]

    Pith/arXiv arXiv 2015

  25. [25]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine, Elucidating the design space of diffusion-based generative models (2022), arXiv:2206.00364 [cs.CV]

    Pith/arXiv arXiv 2022

  26. [26]

    H. Yang, C. Hu, Y . Zhou, X. Liu, Y . Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni, M. Horton, R. Pinsler, A. Fowler, D. Zgner, T. Xie, J. Smith, L. Sun, Q. Wang, L. Kong, C. Liu, H. Hao, and Z. Lu, Mattersim: A deep learning atomistic model across el- ements, temperatures and pressures (2024), arXiv:2405.04967 [cond-mat.mtrl-sci]

    Pith/arXiv arXiv 2024

  27. [27]

    Hjorth Larsen, J

    A. Hjorth Larsen, J. Jrgen Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Duak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C. Jennings, P. Bjerre Jensen, J. Kermode, J. R. Kitchin, E. Leonhard Kolsb- jerg, J. Kubal, K. Kaasbjerg, S. Lysgaard, J. Bergmann Maron- sson, T. Maxson, T. Olsen, L. Pastewka, A. Peterson, C. Rost- ...

    2017

  28. [28]

    A. Togo, K. Shinohara, and I. T. and, Spglib: a software library for crystal symmetry search, Science and Technol- ogy of Advanced Materials: Methods4, 2384822 (2024), https://doi.org/10.1080/27660400.2024.2384822

    arXiv 2024

  29. [29]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V . L. Chevrier, K. A. Persson, and G. Ceder, Python materials genomics (pymatgen): a robust, open-source python library for materials analysis, Comput. Mater. Sci.68, 314 (2013)

    2013

  30. [30]

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

    2019

  31. [31]

    Schmidt, H.-C

    J. Schmidt, H.-C. Wang, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, A dataset of 175k stable and metastable ma- terials calculated with the pbesol and scan functionals, Scien- tific Data9, 64 (2022)

    2022

  32. [32]

    Schmidt, L

    J. Schmidt, L. Pettersson, C. Verdozzi, S. Botti, and M. A. L. Marques, Crystal graph attention networks for the prediction of 14 0 50 100 150 200 Initial space group 0 50 100 150 200 250 300Count (a) T M Ortho Tet Trig Hex Cubic 0 50 100 150 200 Final space group 0 50 100 150 200 (b) T M Ortho Tet Trig Hex Cubic 0.00 0.25 0.50 0.75 1.00 1.25 1.50 RMSD (Å...

    2021

  33. [33]

    Schmidt, N

    J. Schmidt, N. Hoffmann, H.-C. Wang, P. Borlido, P. J. M. A. Carrio, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, Machine-learning-assisted determination of the global zero- temperature phase diagram of materials, Advanced Materials 35, 2210788 (2023)

    2023

  34. [34]

    A. K. Cheetham and R. Seshadri, Artificial intelligence driv- ing materials discovery? perspective on the article: Scaling deep learning for materials discovery, Chemistry of Materials 36, 3490 (2024)

    2024

  35. [35]

    Mazhnik and A

    E. Mazhnik and A. R. Oganov, Application of machine learn- ing methods for predicting new superhard materials, Journal of Applied Physics128, 075102 (2020). 15 Mn4Si4 (194) hform=−0.30eV/atom Ehull=0.07eV/atom H=10.27GPa Mn3NbB2 (38) hform=−0.39eV/atom Ehull=0.04eV/atom H=15.80GPa Co4Ir8Ti4 (59) hform=−0.56eV/atom Ehull=0.01eV/atom H=12.58GPa Mn Nb B Co I...

    2020

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.