Pith. sign in

REVIEW 4 major objections 4 minor 3 cited by

Most AI-generated crystals fail a phonon stability test: only 25.83% pass, with MatterGen best at 41.0%.

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

2026-08-03 14:10 UTC pith:37FZDR3Y

load-bearing objection First large-scale phonon-stability benchmark for AI-generated crystals: direction is right, headline magnitudes need verification before they become quantitative claims. the 4 major comments →

arxiv 2512.21227 v3 pith:37FZDR3Y submitted 2025-12-24 cond-mat.mtrl-sci cs.AI

PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation

Xiao-Qi Han , Ze-Feng Gao , Wen-Kao Li , Peng-Jie Guo , Zhong-Yi Lu This is my paper
classification cond-mat.mtrl-sci cs.AI
keywords dynamical stabilityphononcrystal generationgenerative modelsbenchmarkMatterSimimaginary modesmaterials discovery
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.

This paper tries to establish that current crystal-generation models, despite scoring well on thermodynamic stability, mostly produce crystals that are dynamically unstable—they would not survive small atomic displacements. Running MatterSim-based phonon calculations on more than a hundred thousand relaxed AI-generated structures, the authors find the average dynamical-stability rate is about a quarter to a third (25.83% in the main body; the abstract reports 32.15% on a larger sample), with the best model, MatterGen, reaching 41.0% (45.05% in the abstract). They argue the standard S.U.N. framework misses this because it relies on convex-hull energy rather than on imaginary phonon modes. The practical upshot is a newly identified pool of 28,119 (or 32,995 per the abstract) fully phonon-stable structures, released as candidates for further materials exploration. If correct, dynamical stability should become a standard part of generative-model evaluation.

Core claim

The central claim is that dynamical stability—whether a crystal sits in a local potential well and withstands small displacements—is a bottleneck that current evaluation metrics conceal. The authors benchmarked six generative models, ran MatterSim-v1/Phonopy phonon calculations on 108,843 relaxed structures (2×2×2 supercells, −0.001 THz threshold), and report 28,119 stable (25.83%), with MatterGen best at 41.0% and CrystaLLM at 3.0%. Property-conditioned MatterGen gives 11.6–23.5% stability across band gaps; space-group-constrained CrystalFormer reaches 49.2% for cubic but 34.4% overall. The abstract reports different aggregate numbers (32.15%, 45.05%, 32,995 stable) that the body does not r

What carries the argument

The engine is MatterSim-v1, a universal machine-learning interatomic potential pretrained on 17 million first-principles data points, whose phonon accuracy was validated in a separate study on more than 10,000 materials with roughly 95% true-positive rate in dynamical-stability classification. Combined with Phonopy, the workflow builds 2×2×2 supercells, computes force constants from MatterSim forces, symmetrizes them, interpolates phonon frequencies by Fourier interpolation, and classifies a structure as dynamically stable if no imaginary mode appears below −0.001 THz. The load-bearing evaluation metric is the dynamical-stability ratio: phonon-stable structures divided by successfully relaxe

Load-bearing premise

The benchmark stands on the premise that the machine-learning potential used for the phonon calculations is as accurate on far-from-equilibrium AI-generated structures as it is on ordinary relaxed crystals; if not, all stability ratios and the list of stable structures shift.

What would settle it

Take roughly 200 structures the benchmark calls dynamically stable and 200 it calls unstable, recompute their phonon dispersions with density-functional-theory codes using larger supercells (e.g., 3×3×3) and denser q-meshes, and compare the imaginary-mode classifications. If the true-positive rate on the 'stable' set drops well below the ~95% reported for database materials, the benchmark's ratios—and the 28,119-stable pool—are not robust.

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

If this is right

  • The S.U.N. framework, which relies on thermodynamic stability (Ehull), undercounts instability; adding a phonon-based dynamical-stability check is necessary to judge whether a generated candidate can plausibly be synthesized and persist.
  • The 28,119 fully phonon-stable structures constitute a ready-made candidate pool that downstream DFT screening can focus on without redoing basic stability checks.
  • Pretraining on large, high-quality datasets (e.g., Alex20) appears to raise dynamical-stability rates, so dataset quality is a lever for improving generative models.
  • Generation speed and stability are decoupled: CrystalFlow is the fastest generator (333 crystals/min) but achieves only 16.7% dynamical stability, so throughput alone is not a proxy for quality.
  • In property- and symmetry-conditioned generation, stability rates remain low (best 23.5% at Eg = 0.5 eV, 49.2% for cubic systems), so explicit stability constraints or post-hoc phonon screening should be built into the generation pipeline.

Where Pith is reading between the lines

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

  • If MatterSim's phonon errors are larger on far-off-equilibrium structures than on the relaxed database materials where it was validated, then each reported ratio and even membership in the 28,119-stable pool could shift; a DFT spot-check on a random subset would quantify this.
  • The symmetry–stability correlation suggests that constraining generation to high-symmetry space groups is a cheap way to raise stability, but it may bias discovery away from low-symmetry but potentially stable materials.
  • The 2×2×2 supercell and Fourier-interpolation choice assumes no unstable modes are missed at un-sampled q-points; testing with larger supercells or denser q-meshes on a subset would bound this error.
  • Because this benchmark measures harmonic stability, it does not capture anharmonic or finite-temperature instability; extending the same pipeline to temperature-dependent free energies would be a natural next test for the candidate pool.

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 / 4 minor

Summary. The paper introduces PhononBench, a benchmark for dynamical stability of AI-generated crystals. Using MatterSim-v1 for relaxation and phonon force-constant calculations via Phonopy, the authors evaluate structures generated by several crystal generative models (CrystaLLM, MatterGen, DiffCSP, InvDesFlow-AL, CrystalFlow, CrystalFormer, plus variants and conditioning modes). The full-text abstract and Results report 108,843 successfully relaxed structures, an overall dynamical-stability rate of 25.83%, a best rate of 41.0% for MatterGen, and a pool of 28,119 phonon-stable structures. Additional case studies examine space-group-conditioned CrystalFormer and band-gap-conditioned MatterGen, finding generally low stability rates (e.g., 34.4% for space-group-controlled generation overall and 23.5% at the best band-gap condition of 0.5 eV). The paper argues that current thermodynamic-only stability metrics are insufficient and that a large-scale phonon-based benchmark is needed; it also promises open release of data and workflows.

Significance. If the numerical claims survive validation, PhononBench would be a valuable community resource: it is the first large-scale dynamical-stability assessment of generative crystal models, and its main direction—that current models mostly produce dynamically unstable structures—is plausible and consistent with the field's expectations. The workflow, the released structures, and the candidate pool of stable structures are potentially useful to practitioners. The use of an externally benchmarked universal interatomic potential is a sensible engineering choice, and the paper's convergence analysis for sample sizes is a useful safeguard for model ranking. However, the exact magnitudes of the headline rates and the integrity of the 28,119-structure candidate pool depend on transfer assumptions that are not yet verified, and the manuscript contains serious internal inconsistencies in the reported head numbers. For these reasons, the paper is not yet ready for acceptance.

major comments (4)
  1. [Methods (phonon workflow); Results, Figs. 2(f)–(h), 3(f)–(h)] The headline rates (25.83% overall, MatterGen 41.0%) and the 28,119-member stable pool are computed entirely with MatterSim-v1 for both relaxation and phonon force constants. MatterSim's phonon accuracy in Ref. [41] was established on a large set of mostly database-like, near-equilibrium structures. Here the generated structures are frequently far from that regime, as the paper's own Figs. 2(f)–(h) and 3(f)–(h) show. In this regime second-order force constants are very sensitive to potential-surface errors, and using the same potential for relaxation and for phonons can make errors correlated. A 95% true-positive classification rate on MP-like materials does not by itself determine the false-positive rate or the positive predictive value for this shifted distribution. I request a stratified DFT spot-check—for example, 50–100 structures spanning stable/unstable classifications, multiple m
  2. [Methods: '2x2x2 supercells... Fourier interpolation'; Results: '28,119... entire Brillouin zone'] The claim that the 28,119 structures are 'phonon-stable across the entire Brillouin zone' is stronger than the calculation supports. A 2x2x2 displacement supercell with finite-displacement Phonopy samples only the commensurate q points of that supercell. Unstable modes at other wavevectors, including long-wavelength or zone-center instabilities that require a larger supercell, can be missed. The paper reports no supercell-size or q-mesh convergence test, and no comparison of 2x2x2 with 3x3x3 or 4x4x4 supercells for a subsample. This limitation can only cause over-reporting of stability, so it directly affects the size and reliability of the recommended candidate pool. Either add such a convergence check or soften the claim to 'stable at the commensurate 2x2x2 phonon sampling.'
  3. [Abstract and front matter vs. full-text Results and Methods] The manuscript as presented contains two mutually inconsistent sets of central statistics. The standalone abstract reports 133,838 structures, 7 models, an average stability rate of 32.15%, MatterGen at 45.05%, and 32,995 stable structures at a strict threshold of -0.001 THz. The full-text abstract and Results report 108,843 structures, 6 models, 25.83%, MatterGen 41.0%, and 28,119 stable structures, while the Methods define the imaginary-mode stability threshold as < -1e-3 THz and make no mention of a -0.1 THz threshold. The full text also says 'six' models while Figure 1's caption says 'eight.' These are not minor wording issues: they change the quantitative conclusion of the paper. Please reconcile all headline numbers, clearly state which threshold applies to each reported ratio, and correct the model-count statements before publication.
  4. [Results and Table 3; InvDesFlow-AL-MP20 as a resampling] The total relaxed-structure count is confusing because Table 3 lists InvDesFlow-AL-MP20 as a separate row with 8,000 relaxed structures, when the text explains that these were obtained by random resampling from the 9,163 DiffCSP-MP20 samples. If a reader sums the Table 3 rows, the total is double-counted relative to the 108,843 stated in the text. Please add an explicit footnote to Table 3 explaining that this row is a subsample, not an independent generation run, and consider separating 'independent runs' from 'evaluation subsamples' in the table.
minor comments (4)
  1. [Figure 1 caption, Table 3, Introduction] The number of evaluated models is stated inconsistently: the Introduction says 'six commonly used generative models,' Figure 1 says 'eight generative models,' and Table 3 lists nine rows including variants. Please clarify whether variants are counted as separate models and ensure all statements agree.
  2. [Figures 2(d)–(e), 3(c)–(e)] Several chemical formulae appear corrupted or malformed, for example 'TmThpGeRhq2,' 'CapSm2Snq3,' and similar strings. These should be corrected to valid stoichiometric formulas, e.g., TmTh(GeRh)2 and Ca(Sm2Sn)3.
  3. [Results, 'total of 221,000'] The claimed total of 221,000 generated structures does not obviously match the sum of the model-specific totals in Table 3 plus the band-gap generation set (which sums to about 226,000). Please state explicitly whether '221,000' refers to pre-deduplication, post-deduplication, or something else.
  4. [Methods, stability threshold] The Methods state the threshold as 'threshold < -1e-3 THz,' but the front matter refers to a threshold of -0.1 THz and a 'strict' -0.001 THz threshold. Define the threshold operationally (e.g., 'structures with any phonon frequency below -0.001 THz are classified as unstable') and use one threshold consistently throughout.

Circularity Check

0 steps flagged

No significant circularity: the headline stability rates are measurements through an externally pretrained and independently benchmarked MLIP, not fits to the paper's own outputs.

full rationale

The derivation chain is: generate structures with six external models -> relax with MatterSim-v1 -> compute phonons with MatterSim/Phonopy -> count imaginary modes below threshold -> divide by successfully relaxed structures. None of these steps defines a target quantity in terms of the paper's own fitted parameters. The 25.83%/41.0% rates and the 28,119-stable pool are simply tallies of MatterSim phonon classifications; MatterSim was not retrained or fine-tuned on any of these generated structures or on the benchmark ratios. Its phonon accuracy is supported by Ref. [41], an independent benchmark (Loew/Sun/Wang/Botti/Marques) on >10,000 database materials; under the review rules, that is external evidence and does not become circular because the present authors cite it. The only self-model is InvDesFlow-AL, whose #2 rank is disclosed and whose removal would not change the aggregate conclusion; the paper even estimates the redundant InvDesFlow-AL-MP20 variant by resampling DiffCSP-MP20, an explicit statistical assumption rather than a fitted prediction. The main threats to the headline numbers—transfer of MatterSim accuracy to off-equilibrium generated structures and the 2x2x2 supercell/Fourier-interpolation completeness—are external-validity or correctness risks, not circularity; they are checkable by independent DFT spot checks. The abstract/full-text threshold discrepancy (-0.1 vs -0.001 THz) is an internal-consistency issue, not a circular one. Score 2 reflects only the minor non-load-bearing self-citation in the model comparison.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Everything in the Results reduces to: (i) the external MatterSim phonon benchmark [41], and (ii) the transfer of that benchmark to off-distribution AI-generated structures at 2×2×2 supercell resolution. The only hand-set numbers are the imaginary-mode cutoff and the convergence sample size. The paper is transparent that MatterSim, not DFT, is the instrument, but it does not quantify how classification error propagates into the 28,119-stable pool. No new physical entities are postulated.

free parameters (2)
  • imaginary-frequency stability threshold = −1×10⁻³ THz (Methods v2); −0.1 THz (abstract v1)
    Hand-set cutoff separating 'dynamically stable' from unstable; every headline count (28,119 or 32,995 stable; per-model ratios) is defined relative to it, and no sensitivity analysis is provided.
  • stability-ratio convergence cutoff = ≈4,000 evaluated samples
    Declared threshold above which model stability ratios are 'converged'; used to justify ranking CrystaLLM (1,951 samples) and other small-sample model comparisons. Chosen by inspection of Fig. 5, not by a formal error bound.
axioms (3)
  • domain assumption MatterSim-v1 yields DFT-level phonon frequencies and dynamical-stability classification on AI-generated structures, as validated in ref [41] on >10,000 database materials.
    Invoked in Methods ('we adopt this model to perform large-scale phonon and dynamical-stability evaluations'); generated structures near saddle points (paper's own Figs. 2f–h, 3f–h) are outside the validation distribution.
  • domain assumption A 2×2×2 supercell with 0.01 Å displacements, symmetrized force constants, and Fourier interpolation captures all imaginary modes in the full Brillouin zone.
    Methods: '2×2×2 supercells were generated... dynamical stability was assessed by checking for imaginary modes'; supports the 'stable across the entire Brillouin zone' wording with no convergence check reported.
  • standard math Harmonic phonons at 0 K on the relaxed geometry are the correct criterion for dynamical stability / synthesizability.
    Standard practice (refs [25, 42, 43]); finite-temperature and anharmonic effects are not addressed.

pith-pipeline@v1.3.0-alltime-deepseek · 13735 in / 23996 out tokens · 221604 ms · 2026-08-03T14:10:12.215581+00:00 · methodology

0 comments
read the original abstract

In recent years, generative artificial intelligence has made significant advances in the design of crystalline materials, giving rise to approaches based on graph neural networks, diffusion models, and large language models. Existing evaluations commonly follow the stability-uniqueness-novelty (S.U.N.) framework, where stability is primarily assessed using thermodynamic criteria, which do not fully capture the dynamical stability essential for a material's practical existence. Dynamical stability is a key determinant of whether a material can be synthesized and persist, with phonon spectrum calculations serving as the standard for its evaluation. However, the high computational cost of such calculations has prevented large-scale assessment of dynamical stability in generated crystals. In this work, we introduce PhononBench, the first large-scale benchmark for dynamical stability in AI-generated crystals. Leveraging the recently developed MatterSim interatomic potential, which achieves density-functional-theory (DFT)-level accuracy in phonon predictions across more than 10,000 materials, PhononBench enables efficient phonon calculations and dynamical-stability analysis for 133,838 crystal structures generated by 7 leading crystal generation models. PhononBench reveals a widespread limitation of current generative models: unless otherwise specified, all reported dynamical-stability metrics are evaluated at a phonon-frequency threshold of -0.1 THz, with the average dynamical-stability rate across all generated structures being only 32.15%, and the top-performing model, MatterGen, reaching just 45.05%.In addition, we identify 32,995 crystal structures that are phonon-stable across the entire Brillouin zone under a strict threshold of -0.001 THz. In addition, a web-based service is accessible at http://phononbench.cn/, enabling minute-level ultra-fast phonon predictions.

Figures

Figures reproduced from arXiv: 2512.21227 by Peng-Jie Guo, Wen-Kao Li, Xiao-Qi Han, Ze-Feng Gao, Zhong-Yi Lu.

Figure 1
Figure 1. Figure 1: Systematic Evaluation of Dynamical Stability in Crystal Generation Models. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dynamical Stability Analysis of Space-Group-Constrained Crystal Generation. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Dynamical Stability Analysis of Property-Constrained Crystal Generation. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Elemental distribution heatmap of dynamically stable crystals. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Convergence of the dynamical stability rate for various crystal generative models as a function of the number of tested samples. The figure demonstrates that stability estimates converge as sample size increases, with errors becoming negligible above 4,000 samples, ensuring robust and fair model ranking. 1 Convergence Analysis of Dynamical Stability Rate [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Elemental distribution of 28,119 dynamically stable crystals 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.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 3 Pith papers

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

  1. VibroML: an automated toolkit for high-throughput vibrational analysis and dynamic instability remediation of crystalline materials using machine-learned potentials

    cond-mat.mtrl-sci 2026-04 unverdicted novelty 6.0

    VibroML automates remediation of dynamic instabilities in crystalline materials by combining MLIPs with genetic algorithms for polymorph search, finite-temperature MD validation, and compositional alloying to yield st...

  2. 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.

  3. Inverse Design of Inorganic Compounds with Generative AI

    physics.chem-ph 2026-04 unverdicted novelty 2.0

    A review of generative AI for inverse design of inorganic compounds, analyzing adaptations for their complexity in composition, geometry, symmetry, and electronic structure, with discussion of future benchmarks and sy...

Reference graph

Works this paper leans on

45 extracted references · 1 linked inside Pith · cited by 3 Pith papers

  1. [1]

    A generative model for inorganic materials design.Nature, 639(8055):624–632, 2025

    Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Bichlien Nguyen, Hannes Schulz, Sarah Lewis, Chin-Wei Huang, Ziheng Lu, Yichi Zhou, Han Yang, Hongxia Hao, Jielan Li, Chunlei Yang, Wenjie Li, Ryota Tomioka, and Tian...

    2025

  2. [2]

    Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

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

    2023

  3. [3]

    In- vDesFlow: AnAI-drivenmaterialsinversedesignworkflowtoexplorepossiblehigh-temperature superconductors.Chin

    Xiao-Qi Han, Zhenfeng Ouyang, Peng-Jie Guo, Hao Sun, Ze-Feng Gao, and Zhong-Yi Lu. In- vDesFlow: AnAI-drivenmaterialsinversedesignworkflowtoexplorepossiblehigh-temperature superconductors.Chin. Phys. Lett., 42(4):047301, 2025. 12

    2025

  4. [4]

    AI-Driven Inverse Design of Materials: Past, Present, and Future

    Xiao-Qi Han, Xin-De Wang, Meng-Yuan Xu, Zhen Feng, Bo-Wen Yao, Peng-Jie Guo, Ze-Feng Gao, and Zhong-Yi Lu. AI-Driven Inverse Design of Materials: Past, Present, and Future. Chinese Physics Letters, 42(2):027403, 2025

    2025

  5. [5]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY, USA, 2020. Curran Associates Inc

    2020

  6. [6]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations.ICLR, 2021

    2021

  7. [7]

    Crystal Structure Prediction by Joint Equivariant Diffusion on Lattices and Fractional Coordinates

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal Structure Prediction by Joint Equivariant Diffusion on Lattices and Fractional Coordinates. In Workshop on ”Machine Learning for Materials” ICLR 2023, 2023

    2023

  8. [8]

    Crystalflow: a flow-based generative model for crystalline materials

    Xiaoshan Luo, Zhenyu Wang, Qingchang Wang, Xuechen Shao, Jian Lv, Lei Wang, Yanchao Wang, and Yanming Ma. Crystalflow: a flow-based generative model for crystalline materials. Nature Communications, 16(1):9267, 2025

    2025

  9. [9]

    Invdesflow-al: active learning-basedworkflowforinversedesignoffunctionalmaterials.npj Computational Materials, 11(1):364, 2025

    Xiao-Qi Han, Peng-Jie Guo, Ze-Feng Gao, Hao Sun, and Zhong-Yi Lu. Invdesflow-al: active learning-basedworkflowforinversedesignoffunctionalmaterials.npj Computational Materials, 11(1):364, 2025

    2025

  10. [10]

    Crystal Diffusion Variational Autoencoder for Periodic Material Generation

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi S Jaakkola. Crystal Diffusion Variational Autoencoder for Periodic Material Generation. InInternational Confer- ence on Learning Representations, 2021

    2021

  11. [11]

    Con-CDVAE: A method for the condi- tional generation of crystal structures.Computational Materials Today, 1:100003, May 2024

    Cai-Yuan Ye, Hong-Ming Weng, and Quan-Sheng Wu. Con-CDVAE: A method for the condi- tional generation of crystal structures.Computational Materials Today, 1:100003, May 2024

    2024

  12. [12]

    Deep learning generative model for crystal structure prediction.npj Computa- tional Materials, 10(1):254, 2024

    Xiaoshan Luo, Zhenyu Wang, Pengyue Gao, Jian Lv, Yanchao Wang, Changfeng Chen, and Yanming Ma. Deep learning generative model for crystal structure prediction.npj Computa- tional Materials, 10(1):254, 2024

    2024

  13. [13]

    Space group informed transformer for crystalline materials generation.Science Bulletin, 70(21):3522–3533, 2025

    Zhendong Cao, Xiaoshan Luo, Jian Lv, and Lei Wang. Space group informed transformer for crystalline materials generation.Science Bulletin, 70(21):3522–3533, 2025

    2025

  14. [14]

    CrystalFormer-RL: Reinforcement Fine-Tuning for Materials Design.arXiv preprint arXiv:2504.02367, 2025

    Zhendong Cao and Lei Wang. CrystalFormer-RL: Reinforcement Fine-Tuning for Materials Design.arXiv preprint arXiv:2504.02367, 2025

    arXiv 2025

  15. [15]

    Space group constrained crystal generation

    Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group constrained crystal generation. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun, editors, International Conference on Representation Learning, volume 2024, pages 6836–6853, 2024

    2024

  16. [16]

    GPT-4 Technical Report.arXiv preprint arXiv:2303.08774, 2024

    OpenAI, Josh Achiam, and Steven Adler. GPT-4 Technical Report.arXiv preprint arXiv:2303.08774, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Na- ture, 645(8081):633–638, 2025

    DeepSeek-AI. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Na- ture, 645(8081):633–638, 2025

    2025

  18. [18]

    Antunes, Keith T

    Luis M. Antunes, Keith T. Butler, and Ricardo Grau-Crespo. Crystal structure generation with autoregressive large language modeling.Nature Communications, 15(1):10570, 2024. 13

    2024

  19. [19]

    Anuroop Sriram, Benjamin Kurt Miller, Ricky T. Q. Chen, and Brandon M. Wood. Flowllm: Flow matching for material generation with large language models as base distributions, 2024

    2024

  20. [20]

    Janosh Riebesell, Rhys E. A. Goodall, Philipp Benner, Yuan Chiang, Bowen Deng, Gerbrand Ceder, Mark Asta, Alpha A. Lee, Anubhav Jain, and Kristin A. Persson. A framework to evaluate machine learning crystal stability predictions.Nature Machine Intelligence, 7(6):836– 847, 2025

    2025

  21. [21]

    DPA-2: a large atomic model as a multi-task learner.npj Computational Materials, 10(1):293, 2024

    Duo Zhang, Xinzijian Liu, Xiangyu Zhang, et al. DPA-2: a large atomic model as a multi-task learner.npj Computational Materials, 10(1):293, 2024

    2024

  22. [22]

    The physical significance of imaginary phonon modes in crystals.Electronic Structure, 4(3):033002, jul 2022

    Ioanna Pallikara, Prakriti Kayastha, Jonathan M Skelton, and Lucy D Whalley. The physical significance of imaginary phonon modes in crystals.Electronic Structure, 4(3):033002, jul 2022

    2022

  23. [23]

    High-temperature superconductivity inli 2auh6 mediated by strong electron-phonon coupling under ambient pressure.Phys

    Zhenfeng Ouyang, Bo-Wen Yao, Xiao-Qi Han, Peng-Jie Guo, Ze-Feng Gao, and Zhong-Yi Lu. High-temperature superconductivity inli 2auh6 mediated by strong electron-phonon coupling under ambient pressure.Phys. Rev. B, 111:L140501, Apr 2025

    2025

  24. [25]

    First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    Atsushi Togo and Isao Tanaka. First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    2015

  25. [26]

    Phonons and relatedcrystalpropertiesfromdensity-functionalperturbationtheory.Rev

    Stefano Baroni, Stefano de Gironcoli, Andrea Dal Corso, and Paolo Giannozzi. Phonons and relatedcrystalpropertiesfromdensity-functionalperturbationtheory.Rev. Mod. Phys., 73:515– 562, Jul 2001

    2001

  26. [27]

    Finding unprece- dentedly low-thermal-conductivity half-heusler semiconductors via high-throughput materials modeling.Phys

    Jesús Carrete, Wu Li, Natalio Mingo, Shidong Wang, and Stefano Curtarolo. Finding unprece- dentedly low-thermal-conductivity half-heusler semiconductors via high-throughput materials modeling.Phys. Rev. X, 4:011019, Feb 2014

    2014

  27. [28]

    The openlam challenges: Lam crystal philately competition.Machine Learning: Science and Technology, 6(2):020701, jun 2025

    Anyang Peng, Xinzijian Liu, Ming-Yu Guo, Linfeng Zhang, and Han Wang. The openlam challenges: Lam crystal philately competition.Machine Learning: Science and Technology, 6(2):020701, jun 2025

    2025

  28. [29]

    Mattersim: A deep learning atomistic model across elements, temperatures and pressures, 2024

    Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, ShuizhouChen, ClaudioZeni, MatthewHorton, RobertPinsler, AndrewFowler, DanielZügner, Tian Xie, Jake Smith, Lixin Sun, Qian Wang, Lingyu Kong, Chang Liu, Hongxia Hao, and Ziheng Lu. Mattersim: A deep learning atomistic model across elements, temperatures and pressures, 2024

    2024

  29. [30]

    Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.J

    Yutack Park, Jaesun Kim, Seungwoo Hwang, and Seungwu Han. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.J. Chem. Theory Comput., 20(11):4857–4868, 2024

    2024

  30. [31]

    Bartel, and Gerbrand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5(9):1031–1041, 2023. 14

    2023

  31. [32]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neu- ral networks for data-efficient and accurate interatomic potentials.Nature Communications, 13(1):2453, 2022

    2022

  32. [33]

    EquiformerV2: Improved Equiv- ariant Transformer for Scaling to Higher-Degree Representations

    Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. EquiformerV2: Improved Equiv- ariant Transformer for Scaling to Higher-Degree Representations. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  33. [34]

    Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs

    Saro Passaro and C Lawrence Zitnick. Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs. InInternational Conference on Machine Learning (ICML), 2023

    2023

  34. [35]

    Orb: A fast, scalable neural network potential, 2024

    Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb: A fast, scalable neural network potential, 2024

    2024

  35. [36]

    Orb-v3: atomistic simulation at scale, 2025

    Benjamin Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, and Mark Neumann. Orb-v3: atomistic simulation at scale, 2025

    2025

  36. [37]

    Improving the reliability of language model-predicted structures as docking targets through geometric graph learning.Journal of Medicinal Chem- istry, 68(2):1956–1969, 2025

    ChaoShen, XiaoqiHan, HengCai, TongChen, YuKang, PeichenPan, XiangyangJi, Chang-Yu Hsieh, Yafeng Deng, and Tingjun Hou. Improving the reliability of language model-predicted structures as docking targets through geometric graph learning.Journal of Medicinal Chem- istry, 68(2):1956–1969, 2025

    1956

  37. [38]

    Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    2019

  38. [39]

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

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

    2022

  39. [40]

    Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Ad- vances in Neural Information Processing Systems, 2022

    2022

  40. [41]

    Antoine Loew, Dewen Sun, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Univer- salmachinelearninginteratomicpotentialsarereadyforphonons.npj Computational Materials, 11(1):178, 2025

    2025

  41. [42]

    Implementation strategies in phonopy and phono3py.J

    Atsushi Togo, Laurent Chaput, Terumasa Tadano, and Isao Tanaka. Implementation strategies in phonopy and phono3py.J. Phys. Condens. Matter, 35(35):353001, 2023

    2023

  42. [43]

    First-principles phonon calculations with phonopy and phono3py.J

    Atsushi Togo. First-principles phonon calculations with phonopy and phono3py.J. Phys. Soc. Jpn., 92(1):012001, 2023

    2023

  43. [44]

    Gemnet: Universal directional graph neural networks for molecules

    Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, vol- ume 34, pages 6790–6802. Curran Associates, Inc., 2021

    2021

  44. [45]

    E(n) Equivariant Graph Neural Networks

    Víctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) Equivariant Graph Neural Networks. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 9323–9332. PMLR, 18–24 Jul 2021. 15

    2021

  45. [46]

    Total Generated

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. InProceedings of the 35th In- ternational Conference on Neural Information Processing Systems, NIPS ’21, Red Hook, NY, USA, 2021. Curran Associates Inc. Acknowledgement:The work is supported by Beijing Natur...

    2021