Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Crystal structure prediction can be recast as direct symmetry inference from composition, with diffusion steered by LLM-inferred Wyckoff sites to yield more stable, novel, and unique candidates.

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 ↗

Language models predict space-group and Wyckoff-site patterns from composition, a beam search enforces multiplicity constraints, and a diffusion model generates the 3D structure, improving SUN and matching benchmarks.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely different way to inject symmetry into diffusion CSP, with a plausible but not fully controlled benchmark. the 4 major comments →

arxiv 2602.17176 v4 pith:LM5TA23G submitted 2026-02-19 cond-mat.mtrl-sci cs.AIphysics.comp-ph

NextCrystal: a Symmetry-Driven Generative Framework for Crystal Structure Prediction

Jinming Mu , Lixin He , Xudong Zhu , Shi Yin This is my paper
classification cond-mat.mtrl-sci cs.AIphysics.comp-ph
keywords crystal structure predictionWyckoff positionsspace group inferencelarge language modelsdiffusion modelsbeam searchmaterial discoverysymmetry constraints
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

Crystal structure prediction traditionally requires either expensive combinatorial symmetry searching or retrieval of known prototype templates. This paper claims that a crystal's space group and its fine-grained Wyckoff site assignments can be inferred directly from atomic stoichiometry by two language models, with a linear-complexity beam search enforcing the hard algebraic constraint that site multiplicities sum exactly to the atom counts. These inferred templates then constrain a diffusion generator, forcing the denoising trajectory onto a symmetry-valid geometric manifold. If correct, the result is a generator that proposes new polymorphs without database lookup, winning on stability, uniqueness, and novelty benchmarks, and surfacing a previously unreported low-energy HfO2 phase.

Core claim

On the paper's own terms: the central discovery is that symmetry need not be retrieved from known structures; it can be generated. Two transformer-based language models take an expanded atom sequence and predict a probability distribution over 230 space groups and over Wyckoff letters for each atom. A constrained beam search then solves the assignment problem in time linear in the number of atoms, enforcing that for every element the assigned count on each Wyckoff letter is an exact multiple of that letter's multiplicity and that the letters collectively exhaust the cell's stoichiometry. The predicted space group and Wyckoff template are applied at every denoising step of a diffusion model—m

What carries the argument

The Wyckoff template is the load-bearing object: a site-symmetry assignment that fixes each atom's permitted coordinate degrees of freedom in a given space group. The key identity is the algebraic consistency condition—for every element, the count assigned to a Wyckoff letter must be an exact multiple of that letter's multiplicity, and the letters must sum to the cell's atom counts—which turns assignment into a constrained optimization that is NP-hard in general. The paper's linear-complexity beam search makes this optimization tractable, and the same template acts as a geometric projector in the diffusion dynamics, so every denoising step ends on the symmetry manifold.

Load-bearing premise

The load-bearing premise is that the heuristic beam search, with its fixed beam width and early pruning, will find a feasible, high-scoring Wyckoff assignment whenever one exists for the true space group; if that fails, the entire diffusion run is locked onto a wrong symmetry template.

What would settle it

Run an exhaustive or complete search for valid Wyckoff assignments on a sample of compositions, compare with the beam search's outputs, and identify any composition where a feasible assignment exists but the beam search returns none or a different, lower-scoring one; on those cases, check whether the generated structures collapse in the SUN and matching metrics. Separately, recompute the claimed HfO2 Pnma phase with a plane-wave DFT code and a different exchange-correlation functional plus a full phonon calculation; if the 0.056 eV/atom energy difference disappears or the structure relaxes to

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

If this is right

  • Generators can propose genuinely new space-group and Wyckoff combinations for a composition, not just recombine archived structures.
  • The claimed performance gains hold for small 5-atom cells and for larger cells up to 52 atoms, showing the symmetry-constrained trajectory scales beyond toy systems.
  • The simultaneous stability-uniqueness-novelty metric and the structural matching rate both improve, indicating the constrained manifold does not exclude real ground-truth structures.
  • First-principles screening of generated candidates can find new polymorphs; the reported HfO2 Pnma phase is 0.056 eV/atom lower in energy than the known high-pressure Pnma phase and dynamically stable.

Where Pith is reading between the lines

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

  • The 'no template retrieval' claim applies at inference time; the language models themselves were trained on existing crystal databases, so the space-group distribution they propose may still reflect common prototypes even when individual candidates are new.
  • The HfO2 result is a single data point; whether this pipeline routinely finds low-energy polymorphs across a range of chemistries would require a larger first-principles screening campaign.
  • The beam search carries no completeness guarantee, so a composition whose only valid assignment falls outside the beam will silently produce a wrong template; pairing the search with an infeasibility detector or an anytime exact search would make such failures observable.
  • The full-benchmark stability assessment relies on a learned force-field surrogate rather than density-functional theory; spot-checking a subset with full DFT would clarify whether the SUN gains survive first-principles relaxation.
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

4 major / 5 minor

Summary. NextCrystal is a two-stage generative framework for composition-conditioned crystal structure prediction. An LLM first predicts a distribution over the 230 space groups from an expanded atomic sequence; a second, FiLM-conditioned LLM predicts Wyckoff-letter probabilities for each atom. A beam search (Algorithm 1, Appendix A.3) converts these probabilities into a stoichiometrically consistent Wyckoff assignment by enforcing multiplicity constraints. This symmetry template is then fed into a DiffCSP++ diffusion backbone with lattice-mask and coordinate-projection rectification (Appendix A.4). The paper reports Top-20 SUN and matching-rate results on Perov-5, MP-20, and MPTS-52, claiming large improvements over the DiffCSP++/CSPML retrieval baseline (Tables I and II). It also claims, in the abstract, that a first-principles screening of HfO2 identifies a dynamically stable Pnma phase 0.056 eV/atom below the conventional high-pressure Pnma phase.

Significance. If the reported results are correct, the paper makes a useful contribution to template-free CSP: instead of retrieving a known prototype, it uses learned LLM distributions plus a constrained search to propose symmetry templates, and then hard-constrains a diffusion model to those templates. The methodological transparency is a strength: Appendix A gives a full algorithm, Appendix B gives hyperparameters, and the authors state that the DiffCSP++ baseline is run from its official code with the same splits. The paper also explicitly discloses that stability is evaluated with MatterSim rather than DFT for the SUN benchmarks. However, the significance is currently undercut by two external-control issues: the baseline SUN comparison does not demonstrate that all methods were scored with the same MatterSim protocol, and the HfO2 discovery claim appears only in the abstract with no computational details. These are repairable within the manuscript's scope.

major comments (4)
  1. [Appendix C.1, Table I] The stability component of SUN is defined by MatterSim relaxation and a 0.1 eV/atom convex-hull threshold. Table I credits DiffCSP++ results to 'official source code implementation [27]' but never states that the DiffCSP++ candidate structures were relaxed and scored with the same MatterSim pipeline as the NextCrystal candidates. If the baseline numbers were taken from a publication that used DFT or a different relaxation engine, the reported 124%/255% relative stability gains are not controlled. Because MatterSim is a learned surrogate, it may systematically favor the symmetry-constrained candidates. Please report the exact scoring pipeline for every entry in Table I and, if necessary, recompute the DiffCSP++ SUN scores under the identical protocol.
  2. [Abstract; no supporting section] The abstract claims that first-principles screening of HfO2 identifies a previously unreported dynamically stable Pnma phase 0.056 eV/atom lower in energy than the conventional high-pressure Pnma phase. No DFT functional, pseudopotential, k-point sampling, cutoff, plane-wave settings, structural parameters, phonon supercell, or raw energy data appear anywhere in the main text or appendices. This is a load-bearing discovery claim and is currently uncheckable. Add a dedicated Supporting Information section with full computational details and data, or remove the claim from the abstract.
  3. [Algorithm 1, Appendix A.3; Eq. (3)] The beam search is described as 'rigorously enforcing algebraic consistency' and as having 'linear complexity,' but it is a heuristic with K=100 and early pruning. No completeness or optimality guarantee is provided for finding a feasible, high-scoring Wyckoff assignment whenever one exists. The manuscript also does not report how often the algorithm returns no feasible assignment or fewer than the Top-4 templates assumed by the hierarchical sampling, nor the sensitivity of the SUN/matching results to K. This gap weakens the claim that Eq. (3) is solved in a rigorous sense. Please provide beam-search success statistics and a discussion of when pruning is exact versus approximate.
  4. [Table II and 'Matching Rate' definition] The manuscript defines Top-20 Matching Rate using pymatgen StructureMatcher with stol=0.5, angle_tol=10°, ltol=0.3. However, the CDVAE, DiffCSP, and CrystaLLM entries in Table II are 'taken from Antunes et al.' with no verification that those values were obtained with the same Top-20 candidate protocol and the same tolerances. If those published numbers were computed at Top-1 or with a different number of candidates, the comparison in Table II is not apples-to-apples. Please recompute all baselines under the identical Top-20 protocol or clearly state the original protocol and justify the comparison.
minor comments (5)
  1. [Appendix B] The text 'Gemini said' appears immediately before 'This section details the implementation specifications.' This is a drafting artifact and should be removed.
  2. [Throughout] The paper repeatedly describes the LLM-inferred Wyckoff assignments as 'ab initio.' Since both the LLMs and the diffusion backbone are trained on data, 'template-free' or 'without database retrieval' is more accurate; the phrase 'first-principles' should be reserved for the HfO2 validation once DFT details are supplied.
  3. [Section II, after Eq. (3)] Typo: 'we employ propose a beam search algorithm' should be 'we employ a beam search algorithm'.
  4. [Eq. (3)] The constraint Count_e(l_n) mod mult(l_n) = 0 is written per atom n, but the multiplicity condition is a property of the final assignment over all atoms of each element. Rephrase as a constraint on the final counts for all letters to avoid confusion.
  5. [Fig. 2 caption] The figure caption does not identify the axes. Please state that the x-axis is the space-group index and the y-axis is the unique count per composition under the unique-counting principle.

Circularity Check

0 steps flagged

No significant circularity: the claimed SUN and matching results are empirical measurements against external oracles, not self-referential derivations.

full rationale

NextCrystal's pipeline is a supervised generative model: LLMs predict space-group and Wyckoff-letter distributions from composition, a beam search assigns letters under multiplicity constraints, and a DiffCSP++-based diffusion backbone is rectified onto the inferred symmetry template. The reported SUN and Matching Rate numbers are empirical measurements against external criteria: stability via MatterSim relaxation relative to the convex hull, uniqueness via StructureMatcher within the generated batch, novelty via StructureMatcher against a large merged reference set, and matching via pymatgen StructureMatcher against ground-truth structures. None of these metrics is defined using NextCrystal's predicted templates or fitted parameters, so the scores are not forced by construction. The baseline DiffCSP++ numbers are stated to be reproduced from the official code with the CSPML retrieval setting; even if there were ambiguity about whether the SUN oracle was identical for the baseline, that would be an evaluation-protocol concern, not logical circularity. The 'linear-complexity' beam search is a heuristic without optimality guarantees, and the HfO2 claim lacks DFT/phonon details, but these are rigor and verifiability issues. No equation in the paper defines the target result in terms of its own outputs, and no load-bearing component is justified only by self-citation; the cited diffusion backbone, MatterSim, and benchmark protocols are external to this work's derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

NextCrystal introduces no new physical entity, symmetry operator, or conservation law. Its load-bearing assumptions are all about representational adequacy (Wyckoff/space-group description), the transferability of learned LLM priors, the accuracy of a surrogate force field, and the completeness of a heuristic search. The paper contributes a method, not a new physical postulate.

free parameters (4)
  • Beam width K = 100
    Chosen by hand for the constrained beam search (Table IV); controls search completeness/quality, with no sensitivity analysis reported.
  • Stability threshold = 0.1 eV/atom
    A structure is labeled stable if its formation energy is within 0.1 eV/atom of the convex hull (Appendix C.1); this threshold defines the headline SUN metric.
  • Novelty elemental-ratio tolerance = 10%
    The novelty matcher permits up to 10% deviation in elemental ratios (Appendix C.1); this hand-set tolerance directly affects the reported novelty scores.
  • Matching tolerances = stol=0.5, angle_tol=10°, ltol=0.3
    Standard StructureMatcher tolerances inherited from DiffCSP; not fitted here but hand-chosen and load-bearing for the matching-rate numbers.
axioms (6)
  • domain assumption Every crystal structure is adequately representable by a space group plus Wyckoff positions.
    The entire framework, including the diffusion rectification machinery, assumes the target structure lies on the manifold defined by one of 230 space groups and its Wyckoff orbits. This is standard crystallography, though it excludes incommensurate or otherwise non-Wyckoff-representable structures.
  • domain assumption LLMs trained on structural databases generalize symmetry and Wyckoff priors to unseen compositions.
    The 'ab initio' claim depends on the LLM's learned chemical rules being transferable to new compositions; no composition-level generalization analysis is provided beyond benchmark test splits.
  • domain assumption MatterSim is a high-fidelity surrogate for DFT stability.
    Stability labels for the SUN metric are computed with MatterSim, not DFT (Appendix C.1). The paper cites prior validation but gives no error analysis on the generated candidates.
  • ad hoc to paper Beam search with pruning and K=100 finds a feasible, high-scoring Wyckoff assignment whenever one exists.
    Algorithm 1 is described as 'rigorous' and linear-time, but no completeness or optimality proof is supplied. This is a heuristic assumption specific to the paper's pipeline.
  • domain assumption The DiffCSP++ diffusion backbone and its symmetry rectification are correct as reused.
    The framework adopts DiffCSP++'s GNN backbone and training hyperparameters; errors or artifacts in that backbone would propagate into NextCrystal's results.
  • ad hoc to paper The correct space group is among the Top-5 LLM predictions and the correct Wyckoff assignment among the Top-4 beam-search results.
    The 20-candidate generation strategy (Top-5 space groups x Top-4 Wyckoff templates) assumes the true symmetry is reachable within these truncation thresholds; otherwise no correct structure can be generated.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of NextCrystal: a Symmetry-Driven Generative Framework for Crystal Structure Prediction." pith.science (2026). https://pith.science/paper/LM5TA23G

@misc{pith2026260217176,
  author       = {Pith},
  title        = {Pith review of: NextCrystal: a Symmetry-Driven Generative Framework for Crystal Structure Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LM5TA23G}},
  note         = {Machine review of arXiv:2602.17176}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Crystal structure prediction (CSP), which aims to predict the 3D atomic arrangement of a crystal from its composition, is central to materials discovery and mechanistic understanding. Crystal symmetry plays a crucial role in CSP, but given the composition in a unit cell, existing methods either struggle with the NP-hard combinatorial challenge of enforcing symmetry rigorously or rely on retrieving known templates, inherently limiting both physical fidelity and the discovery of genuinely new materials. To address this challenge, we introduce NextCrystal, a symmetry-driven generative framework that employs large language models to encode chemical semantics and directly generate fine-grained Wyckoff site patterns from atomic stoichiometry, eliminating reliance on database lookups. To overcome the combinatorial complexity of site assignments, we incorporate domain knowledge via an efficient, linear-complexity heuristic beam search, rigorously enforcing algebraic consistency between site multiplicities and atomic stoichiometry. By integrating this symmetry-consistent template into a diffusion backbone, the framework constrains the stochastic generative trajectory to a physically plausible geometric manifold. NextCrystal achieves state-of-the-art performance on stability, uniqueness, and novelty (SUN) benchmarks, as well as superior structural matching, establishing a rigorous paradigm for exploring previously unexplored crystallographic space without relying on prior structural templates. As a representative application, first-principles screening of HfO2 candidates generated by NextCrystal identifies a previously unreported dynamically stable Pnma phase, 0.056~eV/atom lower in energy than the conventional high-pressure Pnma phase.

Figures

Figures reproduced from arXiv: 2602.17176 by Jinming Mu, Lixin He, Shi Yin, Xudong Zhu.

Figure 1
Figure 1. Figure 1: FIG. 1: Schematic of the proposed symmetry-driven generative framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Absolute frequency distribution (raw counts) of space groups among the candidates across the testing set of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Visual comparison of prediction results from DiffCSP++ and our method for a testing sample, i.e., [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Visual comparison of prediction results from DiffCSP++ and our method for a testing sample, i.e., [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

33 extracted references · 2 linked inside Pith

  1. [1]

    M. A. Green, A. Ho-Baillie, and H. J. Snaith, The emer- gence of perovskite solar cells, Nature photonics8, 506 (2014)

    2014

  2. [2]

    L. I. Berger,Semiconductor materials(CRC press, 2020)

    2020

  3. [3]

    Jiang and K

    Q. Jiang and K. Zhu, Rapid advances enabling high- performance inverted perovskite solar cells, Nature Re- views Materials9, 399 (2024)

    2024

  4. [4]

    J. Ding, D. Ji, Y. Yue, and M. M. Smedskjaer, Amor- phous materials for lithium-ion and post-lithium-ion bat- teries, Small20, 2304270 (2024)

    2024

  5. [5]

    Langer and D

    R. Langer and D. A. Tirrell, Designing materials for bi- ology and medicine, Nature428, 487 (2004)

    2004

  6. [6]

    Z. W. Seh, J. Kibsgaard, C. F. Dickens, I. Chorkendorff, J. K. Nørskov, and T. F. Jaramillo, Combining theory and experiment in electrocatalysis: Insights into materi- als design, Science355, eaad4998 (2017)

    2017

  7. [7]

    Marzari, A

    N. Marzari, A. Ferretti, and C. Wolverton, Electronic- structure methods for materials design, Nature materials 20, 736 (2021)

    2021

  8. [8]

    R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu, and Y. Liu, Crystal structure prediction by joint equiv- ariant diffusion on lattices and fractional coordinates, in Workshops of the Eleventh International Conference on Learning Representations(2023)

    2023

  9. [9]

    L. M. Antunes, K. T. Butler, and R. Grau-Crespo, Crys- tal structure generation with autoregressive large lan- guage modeling, Nature Communications (2024)

    2024

  10. [10]

    are reproduced from the official source code [27] using the CSPML [21] retrieval-based template method as described in the original study. Dataset Method Matching Rate MP-20 CDV AE 66.95 DiffCSP 77.93 CrystaLLM 73.97 DiffCSP++ 74.65 Ours 81.70 Perov-5 CDV AE 88.51 DiffCSP 98.60 CrystaLLM 97.60 DiffCSP++ 99.84 Ours 99.95 MPTS-52 CDV AE 20.79 DiffCSP 34.02 ...

  11. [11]

    S. Yang, S. Batzner, R. Gao, M. Aykol, A. Gaunt, B. Mc- Morrow, D. Rezende, D. Schuurmans, I. Mordatch, and E. D. Cubuk, Generative hierarchical materials search, inAdvances in Neural Information Processing Systems (2024) pp. 38799–38819

    2024

  12. [12]

    R. Jiao, W. Huang, Y. Liu, D. Zhao, and Y. Liu, Space group constrained crystal generation, inThe Twelfth International Conference on Learning Representations (2024)

    2024

  13. [13]

    Y. Han, C. Ding, J. Wang, H. Gao, J. Shi, S. Yu, Q. Jia, S. Pan, and J. Sun, Efficient crystal structure prediction based on the symmetry principle, Nature Computational Science (2025)

    2025

  14. [14]

    They employ a joint generation paradigm that co-samples chemical composition and structural coordi- nates

    and CrystalFormer [20] are also leading methods in materials generation, they are excluded from direct com- parison due to a fundamental discrepancy in task formu- lation. They employ a joint generation paradigm that co-samples chemical composition and structural coordi- nates. While this enables broad exploration of the joint space of compositions and st...

  15. [15]

    R. Zhu, W. Nong, S. Yamazaki, and K. Hippalgaonkar, Wycryst: Wyckoff inorganic crystal generator frame- work, Matter7, 3469 (2024)

    2024

  16. [16]

    C. Zeni, R. Pinsler, D. Z¨ ugner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabb´ e, S. Ueda,et al., A generative model for inorganic materials design, Na- ture (2025)

    2025

  17. [17]

    J. Xu, X. Wang, W. Cheng, Y.-P. Cao, Y. Shan, X. Qie, and S. Gao, Dream3d: Zero-shot text-to-3d synthesis us- ing 3d shape prior and text-to-image diffusion models, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2023) pp. 20908–20918

    2023

  18. [18]

    Wang, H.-Y

    C. Wang, H.-Y. Peng, Y.-T. Liu, J. Gu, and S.-M. Hu, Diffusion models for 3d generation: A survey, Computa- tional Visual Media11, 1 (2025)

    2025

  19. [19]

    Y. Xin, Z. Xie, Y. Chen, P. Gao, H. J. Zhao, J. Lv, Y. Wang, and Y. Ma, Symmetry-oriented enumeration method for molecular crystal structure prediction and its application, Physical Review B111, 134112 (2025)

    2025

  20. [20]

    T. Xie, X. Fu, O. Ganea, R. Barzilay, and T. S. Jaakkola, Crystal diffusion variational autoencoder for periodic ma- terial generation, inThe Tenth International Conference on Learning Representations(2022)

    2022

  21. [21]

    R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu, and Y. Liu, Crystal structure prediction by joint equivariant diffusion, inAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023(2023)

    2023

  22. [22]

    Z. Cao, X. Luo, J. Lv, and L. Wang, Space group in- formed transformer for crystalline materials generation, Science Bulletin (2025)

    2025

  23. [23]

    Kusaba, C

    M. Kusaba, C. Liu, and R. Yoshida, Crystal structure prediction with machine learning-based element substi- tution, Computational Materials Science211, 111496 (2022)

    2022

  24. [24]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need, inConference on Neural Infor- mation Processing Systems(2017) pp. 5998–6008

    2017

  25. [25]

    Puigcerver, C

    J. Puigcerver, C. R. Ruiz, B. Mustafa, and N. Houlsby, From sparse to soft mixtures of experts, inThe Twelfth International Conference on Learning Representations (2024)

    2024

  26. [26]

    Perez, F

    E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville, Film: Visual reasoning with a general con- ditioning layer, inProceedings of the AAAI conference on artificial intelligence(2018)

    2018

  27. [27]

    I. E. Castelli, D. D. Landis, K. S. Thygesen, S. Dahl, I. Chorkendorff, T. F. Jaramillo, and K. W. Jacobsen, New cubic perovskites for one-and two-photon water 10 splitting using the computational materials repository, Energy & Environmental Science5, 9034 (2012)

    2012

  28. [28]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., Commentary: The materials project: A materials genome approach to accelerating materials innovation, APL materials1, 011002 (2013). [27]https://github.com/jiaor17/DiffCSP-PP

    2013

  29. [29]

    J. Su, M. H. M. Ahmed, Y. Lu, S. Pan, W. Bo, and Y. Liu, Roformer: Enhanced transformer with rotary po- sition embedding, Neurocomputing568, 127063 (2024)

    2024

  30. [30]

    H. Yang, C. Hu, Y. Zhou, X. Liu, Y. Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni,et al., Mattersim: A deep learning atomistic model across elements, temperatures and pressures, arXiv preprint arXiv:2405.04967 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    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 (py- matgen): A robust, open-source python library for mate- rials analysis, Computational Materials Science68, 314 (2013)

    2013

  32. [32]

    Schmidt, N

    J. Schmidt, N. Hoffmann, H.-C. Wang, P. Borlido, P. J. Carri¸ co, T. F. Cerqueira, S. Botti, and M. A. Marques, Large-scale machine-learning-assisted exploration of the whole materials space, arXiv preprint arXiv:2210.00579 (2022). Appendix A: Detailed Design on Model Architectures and Algorithms A.1. Input F eaturization and Embeddings As defined in the ...

    Pith/arXiv arXiv 2022

  33. [33]

    employed density functional theory (DFT) for the final stability verification of 1,024 samples, per- forming DFT relaxations across the entire test set is computationally prohibitive under our extensive Top-20 evaluation setting. To ensure computa- tional feasibility, we utilize MatterSim [29], an ac- curate machine learning force field method vali- dated...

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.