Pith. sign in

REVIEW 4 major objections 6 minor 47 references

Stress-Testing Multimodal Foundation Models for Crystallographic Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that nine state-of-the-art vision-language foundation models cannot generalize crystallographic structure prediction to unseen supercell sizes or unseen chemical compositions, with compositional transfer…

desk verdict The dataset and SE/CE holdout protocols are a genuinely useful new benchmark for VLM crystallographic reasoning, but Table 2's transfer ratios are arithmetically irreconcilable with the paper's own error tables, so the headline 'orders-of-magnitude collapse' claim is currently unsupported. read the letter →

arxiv 2506.13051 v1 pith:7PVZZBKF submitted 2025-06-16 cs.CV cond-mat.mtrl-scics.CLcs.LG

Can Polat , Hasan Kurban , Erchin Serpedin , Mustafa Kurban This is my paper
classification cs.CVcond-mat.mtrl-scics.CLcs.LG
keywords crystallographicreasoningvision-languagefoundationmodelsmultimodalbenchmarksspatialexclusioncompositionallatticeparameterpredictionhallucinationevaluationmaterialsinformatics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces a stress-testing benchmark for multimodal foundation models in crystallography: a dataset of ten crystalline materials rendered as images at four supercell radii, together with two withholding protocols. Spatial-exclusion hides one supercell size, and compositional-exclusion hides an entire chemical composition, forcing models to extrapolate rather than retrieve. Run across nine vision-language models, the errors on lattice parameters, density, and atom count stay high, and the transfer ratio $T = \mathrm{CE}/\mathrm{SE}$ climbs from $2.2\times 10^3$ to $2.3\times 10^4$, with one model diverging. The authors conclude that these models memorize numeric templates instead of internalizing geometric and physical constraints, meaning that high in-distribution accuracy is not evidence of genuine crystallographic reasoning and that zero-shot use of these models for materials discovery would be unreliable.

What carries the argument

The load-bearing mechanism is the pair of exclusion protocols together with the transfer ratio that compares them. Spatial-exclusion (SE) holds out one supercell radius for a material while giving the model other radii as context; compositional-exclusion (CE) holds out one entire chemistry. Each test instance supplies only the Cartesian coordinates of the target structure, and the model must produce lattice parameters, density, atom count, and primitive-cell geometry without seeing that radius or that composition. The dataset itself is generated from experimental unit cells (e.g., gold with lattice constant $a_{\mathrm{Au}} = 4.0782$ Å) by forming spherical supercells through an integer multiplicity matrix $S \in \mathbb{Z}^{3\times3}_{>0}$ with $\det(S)\le 8$, relaxing with SCC-DFTB, and rendering nine Fibonacci-sphere rotation views plus a native view as Gaussian-blurred disks. The comparison of SE and CE errors (the ratio $T$) is what turns the two benchmarks into a diagnostic of template reliance versus genuine reasoning, and the correlation-shift analysis over property-pair errors adds a second diagnostic of which physical couplings survive distribution shift.

What would settle it

Re-run the SE and CE probes with the rendered image removed and the same atomic coordinates supplied as a compact numeric table; if the high errors collapse, the models' failure is in numerical extraction rather than crystallographic reasoning, which would undermine the paper's central conclusion, whereas unchanged errors would support it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that current vision-language foundation models fail to internalize the geometric and physical constraints governing crystal structures. When all supercells of a given radius are withheld (SE), models already show sizable percent errors in atom count, volume, lattice parameters, and density; when an entire chemistry is withheld (CE), errors jump by three to four orders of magnitude, with transfer ratios $T = \mathrm{CE}/\mathrm{SE}$ between $2.2\times 10^3$ and $2.3\times 10^4$, and one model diverging entirely. The authors attribute this to reliance on memorized numeric templates for familiar materials, and support the interpretation with physical-law compliance scores below acceptable thresholds (worst for TiO$_2$) and hallucination scores exceeding 40% for several materials. Their conclusion is that in-distribution performance does not imply genuine crystallographic reasoning, and that explicit physical priors, symmetry constraints, and uncertainty-aware mechanisms are needed in future architectures.

Load-bearing premise

The benchmark assumes that giving a model only a long list of Cartesian coordinates still tests crystallographic reasoning from images, yet every target quantity (density, volume, lattice parameters) is an exact function of those coordinates, so the observed failures could stem from numeric processing or structured-output errors rather than from a missing grasp of crystal physics.

Editorial extensions

If this is right

  • High accuracy on familiar sizes and chemistries cannot be taken as evidence that a multimodal model can reason about crystals; exclusion-based evaluation is needed to detect extrapolation failure.
  • Zero-shot vision-language models are currently not reliable tools for high-throughput screening of novel compounds, because compositional transfer errors can be orders of magnitude higher than in-distribution errors.
  • The transfer ratio $T = \mathrm{CE}/\mathrm{SE}$ gives a simple, quantitative signal for memorization versus generalization: very large $T$ indicates the model is reproducing memorized templates rather than applying physical rules.
  • The observed instability of error correlations between properties (for example, the volume–energy coupling dropping sharply under CE) implies that even internally consistent-looking predictions are not guaranteed to remain physically coherent under distribution shift.

Reading between the lines

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

  • The paper's protocol conflates two skills: reading a long list of Cartesian coordinates and reasoning about crystal structure. A follow-up that provides the same coordinates in a compact structured table, or that removes the image, could separate the contribution of numerical processing from genuine geometric understanding; the authors themselves note the evaluation is zero-shot and synthetic, so
  • A testable extension is to give models the explicit functional relationships (density as mass divided by volume, volume from the supplied cell vectors) and compare errors with and without that hint; this would indicate whether failures are arithmetic execution errors or missing physical concepts.
  • The SE/CE transfer ratio could serve as a general-purpose probe for template reliance in other scientific multimodal benchmarks, beyond crystals, wherever a numeric target is a deterministic function of the prompt's coordinates.
  • The correlation-shift analysis suggests a possible predictive use: pairwise error correlations under one exclusion regime may forecast which physical relationships will break under another, turning the benchmark into a diagnostic of the specific priors a model has learned.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper introduces a dataset of ten crystalline materials with spherical supercells at four radii (R7–R10), each rendered in multiple orientations, together with two evaluation protocols. The Spatial-Exclusion (SE) protocol withholds one supercell radius per material and asks models to predict structural properties of a held-out radius from the remaining radii; the Compositional-Exclusion (CE) protocol withholds a whole composition and asks models to predict its properties from the other nine materials. Nine vision-language foundation models are prompted with structural context and evaluated via percent errors in lattice parameters, density, atom counts, and angles, plus physics-consistency and hallucination scores. The paper reports high errors overall, an "orders-of-magnitude" degradation in transfer from SE to CE, and shifts in inter-property error correlations, concluding that current multimodal foundation models fail to internalize crystallographic constraints.

Significance. The dataset construction is transparent and the code is publicly available, and the evaluation uses no fitted parameters or circular derivations, so the benchmark is a potentially useful resource. If the reported transfer collapse were correct, it would be an important cautionary result for the use of vision-language models in materials science. However, the central quantitative evidence in Table 2 is internally inconsistent, the test input is not actually multimodal as claimed, and the correlation analysis relies on an undefined formation-energy property. These issues currently undermine the headline claims, though they may be addressable in a major revision.

major comments (4)
  1. [Section 5, Table 2] The transfer ratios reported in Table 2 are arithmetically inconsistent with the SE and CE columns in the same table. For Claude Opus 4, CE/SE = 0.91/0.06 ≈ 15, not 2.17×10^3; for no model does CE/SE approach 2.2×10^3 to 2.3×10^4, and no T = ∞ appears. The text states that "the transfer ratio T = CE/SE surges from 2.2×10^3 to 2.3×10^4, with one model diverging entirely", which is not supported by the displayed data. Moreover, the SE column values (0.04–0.18%) are impossible means of the per-material errors in Table 1(a), where values such as Au R10 %ΔV = 583.53 appear. If Table 2 aggregates a different quantity, such as normalized or log-scaled errors, that metric must be defined; otherwise the paper's central "orders of magnitude" transfer-degradation claim is unsupported.
  2. [Section 4, Spatial-Exclusion and Compositional-Exclusion Protocols] The protocols state that "Each test instance uses only the Cartesian coordinates of (mi, R*, k)", so the test query contains no crystallographic image, despite the paper's framing as a multimodal vision-language benchmark. The rendered images play no explicit role in the test query, meaning the experiments cannot distinguish between failures of image-based crystallographic reasoning and failures to perform arithmetic on long coordinate lists or to emit structured output. Because lattice parameters and density are exact functions of the Cartesian coordinates, the reported errors may reflect output-formatting or calculation limitations rather than absence of physical understanding. The authors should either include the rendered image in the test input and describe its role, or reframe the claims as evaluating text-based reasoning from coordinates and remove "vision-language" from the title and abstract.
  3. [Section 5, Table 3] The correlation-shift analysis includes formation energy ¯ε as a property, but this quantity is not defined in the evaluation metrics of Section 4, in the expert annotation schema of Section 3.3, or in the dataset description. Without a definition of ¯ε and how it is computed from model outputs or reference structures, the reported shifts (e.g., V↔¯ε decreasing by 0.64 between SE and CE) cannot be reproduced or interpreted. Please add the definition and output schema for ¯ε, or remove the pairs involving it from Table 3.
  4. [Section 5, Knowledge Transfer] The comparison T = CE/SE is not well posed because the two protocols evaluate different property sets: Table 1(a) reports SE errors for a, b, c, and ρ, while Table 1(b) reports CE errors for ap, bp, cp, and angular deviations |Δαp|, |Δβp|, |Δγp|. A ratio of errors on incommensurate physical quantities conflates the protocol effect with the property-set change, so the "several orders of magnitude" degradation does not isolate cross-composition transfer. A common property set should be used, or the validity of the ratio should be explicitly argued.
minor comments (6)
  1. [Section 5, Knowledge Transfer] The sentence "all eight models achieve low mean percent errors" is inconsistent with Table 2, which lists nine models; please correct the count.
  2. [Table 2, column header] The column header "T ×10^3" is ambiguous; the displayed values 2.17–22.54 do not correspond to the ratios computed from the SE and CE columns, nor to the text's claim that T ranges from 2.2×10^3 to 2.3×10^4. Define the notation unambiguously.
  3. [Section 4] The protocols use five rotations for the context and five for the test, while Section 3.2 describes ten orientations per supercell. Explain how these five are selected and why the remaining five orientations are unused.
  4. [Table 1 caption] The caption refers to green and red coloring, but the table as typeset is monochrome; provide a legend or remove the color references.
  5. [Section 3.2] The phrase "nine evenly spaced rotation axes" is imprecise for a Fibonacci sphere with N=9; such points are approximately uniform but not exactly evenly spaced. Consider rewording.
  6. [Section 2.2] The sentence "HoneyComb [35] introduces an LLM-driven frameworks" contains a grammatical error: "introduces" should agree with the plural subject, and "frameworks" should be singular.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmarks compare model outputs against independently defined reference structures, and self-citations appear only in non-load-bearing related-work passages.

full rationale

The paper's central derivation chain is an empirical evaluation: models are prompted with crystallographic images and text, their structured outputs are compared with reference structures via explicitly defined percent-error, compliance, and hallucination metrics, and aggregate statistics are reported. No parameter is fitted to the target data and then renamed a prediction; no quantity used as an input is defined in terms of the claimed output; and no uniqueness or theoretical result is imported from the authors' own prior work to force a conclusion. The self-citations present (e.g., refs. 20, 26, 31, 34) occur in the background discussion of related benchmarks and modeling approaches, and none is load-bearing for the SE/CE evaluation design or for the conclusion that models fail to internalize crystallographic constraints. The notable numerical inconsistency between Table 1's per-material errors and Table 2's very low SE means is a serious internal-consistency and correctness concern, and the limitations section does not address it, but it is not a circularity of the kind defined here: there is no reduction of a claimed prediction to a fitted input or to a self-citation chain. The benchmark is self-contained against externally defined reference structures, so the honest finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central numerical results depend on the generative supercell construction (Equations 1-2), the rendering pipeline, the DFTB-based reference structures, and a text parser that is not described. The benchmark thresholds of 10% and 25% are hand-chosen. No fitted parameters are required to derive the headline percent errors.

free parameters (1)
  • Physics-consistency and hallucination thresholds = 0.10 and 0.25 relative error
    Hand-chosen thresholds in Section 4 define sp and hp; they shape the reported compliance and hallucination scores but do not affect the underlying percent-error measurements.
assumptions (4)
  • domain assumption Spherical supercells with det(S) <= 8 and cutoff radius R capture physically meaningful clusters
    Used in Section 3.1, Equations (1)-(2); all reference structures derive from this construction, so errors are measured against this model rather than an independent experimental dataset.
  • domain assumption Rotated 2D projections with Gaussian-blurred disks retain the crystallographic information needed to infer lattice parameters and density
    Section 3.2 rendering pipeline; if this fails, high model errors reflect an information bottleneck rather than poor reasoning.
  • domain assumption Reference structures relaxed with SCC-DFTB and experimental unit cells are correct ground truth
    Section 4 Dataset; no independent validation of these references against experiment is provided.
  • domain assumption Model outputs can be faithfully parsed into the MATERIAL PROPERTIES schema
    Section 3.3 and Section 4; the parser and failure handling are not described, so parse failures could masquerade as prediction errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stress-Testing Multimodal Foundation Models for Crystallographic Reasoning." pith.science (2026). https://pith.science/paper/7PVZZBKF

@misc{pith2026250613051,
  author       = {Pith},
  title        = {Pith review of: Stress-Testing Multimodal Foundation Models for Crystallographic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PVZZBKF}},
  note         = {Machine review of arXiv:2506.13051}
}
read the original abstract

Evaluating foundation models for crystallographic reasoning requires benchmarks that isolate generalization behavior while enforcing physical constraints. This work introduces a multiscale multicrystal dataset with two physically grounded evaluation protocols to stress-test multimodal generative models. The Spatial-Exclusion benchmark withholds all supercells of a given radius from a diverse dataset, enabling controlled assessments of spatial interpolation and extrapolation. The Compositional-Exclusion benchmark omits all samples of a specific chemical composition, probing generalization across stoichiometries. Nine vision--language foundation models are prompted with crystallographic images and textual context to generate structural annotations. Responses are evaluated via (i) relative errors in lattice parameters and density, (ii) a physics-consistency index penalizing volumetric violations, and (iii) a hallucination score capturing geometric outliers and invalid space-group predictions. These benchmarks establish a reproducible, physically informed framework for assessing generalization, consistency, and reliability in large-scale multimodal models. Dataset and code are available at https://github.com/KurbanIntelligenceLab/StressTestingMMFMinCR.

Figures

Figures reproduced from arXiv: 2506.13051 by the authors.

Figure 1
Figure 1. Gallery of atomic structures for each material. The first column shows the primitive unit cell for each material, while the subsequent [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

  1. [1]

    R. W. G. Wyckoff, Crystal Structures Volume 1. Inter- science Publishers, 1963. 1, 10

    work page 1963

  2. [2]

    Bhadeshia, Geometry of crystals, vol

    H. Bhadeshia, Geometry of crystals, vol. 8. Institute of Materials London, 2001. 1

    work page 2001

  3. [3]

    Theory and simulation of crystal growth,

    A. C. Levi and M. Kotrla, “Theory and simulation of crystal growth,”Journal of Physics: Condensed Matter, vol. 9, no. 2, p. 299, 1997. 1

    work page 1997

  4. [4]

    On the effective lattice parameter of bi- nary alloys,

    V . Lubarda, “On the effective lattice parameter of bi- nary alloys,” Mechanics of materials, vol. 35, no. 1-2, pp. 53–68, 2003. 1

    work page 2003

  5. [5]

    Self-consistent equations including exchange and correlation effects,

    W. Kohn and L. J. Sham, “Self-consistent equations including exchange and correlation effects,”Physical Review, vol. 140, no. 4A, p. A1133, 1965. 1

    work page 1965

  6. [6]

    Hydrothermal growth of zno nanostructures,

    S. Baruah and J. Dutta, “Hydrothermal growth of zno nanostructures,” Science and technology of advanced materials, vol. 10, no. 1, p. 013001, 2009. 1

    work page 2009

  7. [7]

    Chemical vapor de- position,

    J.-O. Carlsson and P. M. Martin, “Chemical vapor de- position,” in Handbook of Deposition Technologies for films and coatings, pp. 314–363, Elsevier, 2010. 1

    work page 2010

  8. [8]

    Bertucco and G

    A. Bertucco and G. Vetter,High pressure process tech- nology: fundamentals and applications, vol. 9. Else- vier, 2001. 1

    work page 2001

Show all 47 references
  1. [9]

    The crystal- lographic information file (cif): a new standard archive file for crystallography,

    S. R. Hall, F. H. Allen, and I. D. Brown, “The crystal- lographic information file (cif): a new standard archive file for crystallography,”Foundations of Crystallogra- phy, vol. 47, no. 6, pp. 655–685, 1991. 1

    1991

  2. [10]

    Aflow: An automatic frame- work for high-throughput materials discovery,

    S. Curtarolo, W. Setyawan, G. L. Hart, M. Jahnatek, R. V . Chepulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy,et al., “Aflow: An automatic frame- work for high-throughput materials discovery,”Compu- tational Materials Science, vol. 58, pp. 218–226, 2012. 1

    2012

  3. [11]

    Commentary: The materials project: A materi- als genome approach to accelerating materials innova- tion,

    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 materi- als genome approach to accelerating materials innova- tion,” APL Materials, vol. 1, no. 1, 2013. 1

    2013

  4. [12]

    Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd),

    J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolverton, “Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd),” Jom, vol. 65, pp. 1501–1509, 2013. 1

    2013

  5. [13]

    Numerical methods for the inverse problem of density functional theory,

    D. S. Jensen and A. Wasserman, “Numerical methods for the inverse problem of density functional theory,” International Journal of Quantum Chemistry, vol. 118, no. 1, p. e25425, 2018. 2

    2018

  6. [14]

    Self-interaction and strong correlation in dftb,

    B. Hourahine, S. Sanna, B. Aradi, C. K ¨ohler, T. Niehaus, and T. Frauenheim, “Self-interaction and strong correlation in dftb,” The Journal of Physical Chemistry A, vol. 111, no. 26, pp. 5671–5677, 2007. 2

    2007

  7. [15]

    Dftb3: Extension of the self-consistent-charge density-functional tight- binding method (scc-dftb),

    M. Gaus, Q. Cui, and M. Elstner, “Dftb3: Extension of the self-consistent-charge density-functional tight- binding method (scc-dftb),” Journal of Chemical The- ory and Computation, vol. 7, no. 4, pp. 931–948, 2011. 2

    2011

  8. [16]

    The slater– koster tight-binding method: a computationally effi- cient and accurateapproach,

    D. Papaconstantopoulos and M. Mehl, “The slater– koster tight-binding method: a computationally effi- cient and accurateapproach,” Journal of Physics: Con- densed Matter, vol. 15, no. 10, p. R413, 2003. 2

    2003

  9. [17]

    Machine learn- ing material properties from the periodic table us- ing convolutional neural networks,

    X. Zheng, P. Zheng, and R.-Z. Zhang, “Machine learn- ing material properties from the periodic table us- ing convolutional neural networks,”Chemical Science, vol. 9, no. 44, pp. 8426–8432, 2018. 2

    2018

  10. [18]

    Transformers in material science: roles, chal- lenges, and future scope,

    N. Rane, “Transformers in material science: roles, chal- lenges, and future scope,”Challenges and Future Scope (March 26, 2023), 2023

    2023

  11. [19]

    Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations,

    Y .-L. Liao, B. Wood, A. Das, and T. Smidt, “Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations,” arXiv preprint arXiv:2306.12059, 2023

    2023 arXiv

  12. [20]

    En- hancing the electronic properties of tio2 nanoparticles through carbon doping: An integrated dftb and com- puter vision approach,

    M. Kurban, C. Polat, E. Serpedin, and H. Kurban, “En- hancing the electronic properties of tio2 nanoparticles through carbon doping: An integrated dftb and com- puter vision approach,” Computational Materials Sci- ence, vol. 244, p. 113248, 2024. 2

    2024

  13. [21]

    Schnet: A continuous-filter convolutional neural network for modeling quantum interactions,

    K. Sch ¨utt, P.-J. Kindermans, H. E. Sauceda Felix, S. Chmiela, A. Tkatchenko, and K.-R. M¨uller, “Schnet: A continuous-filter convolutional neural network for modeling quantum interactions,” Advances in Neural Information Processing Systems, vol. 30, 2017. 2

    2017

  14. [22]

    Directional message passing for molecular graphs,

    J. Gasteiger, J. Groß, and S. G¨unnemann, “Directional message passing for molecular graphs,” arXiv preprint arXiv:2003.03123, 2020

    2003 arXiv

  15. [23]

    Densegnn: universal and scalable deeper graph neural networks for high-performance property prediction in crystals and molecules,

    H. Du, J. Wang, J. Hui, L. Zhang, and H. Wang, “Densegnn: universal and scalable deeper graph neural networks for high-performance property prediction in crystals and molecules,” npj Computational Materials, vol. 10, no. 1, p. 292, 2024

    2024

  16. [24]

    Faenet: Frame averaging equivariant gnn for materials modeling,

    A. A. Duval, V . Schmidt, A. Hern ´andez-Garcıa, S. Miret, F. D. Malliaros, Y . Bengio, and D. Rolnick, “Faenet: Frame averaging equivariant gnn for materials modeling,” in International Conference on Machine Learning, pp. 9013–9033, PMLR, 2023. 2

    2023

  17. [25]

    Multimodal machine learning for materials science: composition-structure bimodal learning for experimentally measured properties,

    S. Gong, S. Wang, T. Zhu, Y . Shao-Horn, and J. C. Grossman, “Multimodal machine learning for materials science: composition-structure bimodal learning for experimentally measured properties,” arXiv preprint arXiv:2309.04478, 2023. 2

    2023 arXiv

  18. [26]

    Multimodal neu- ral network-based predictive modeling of nanoparticle properties from pure compounds,

    C. Polat, M. Kurban, and H. Kurban, “Multimodal neu- ral network-based predictive modeling of nanoparticle properties from pure compounds,” Machine Learning: Science and Technology, vol. 5, no. 4, p. 045062, 2024. 8

    2024

  19. [27]

    Chemvlm: Exploring the power of multimodal large language models in chemistry area,

    J. Li, D. Zhang, X. Wang, Z. Hao, J. Lei, Q. Tan, C. Zhou, W. Liu, Y . Yang, X. Xiong,et al., “Chemvlm: Exploring the power of multimodal large language models in chemistry area,” in Proceedings of the AAAI Conference on Artificial Intelligence, no. 1, pp. 415– 423, 2025. 2

    2025

  20. [28]

    Learn to explain: Multimodal reasoning via thought chains for science question answering,

    P. Lu, S. Mishra, T. Xia, L. Qiu, K.-W. Chang, S.- C. Zhu, O. Tafjord, P. Clark, and A. Kalyan, “Learn to explain: Multimodal reasoning via thought chains for science question answering,” Advances in Neural Information Processing Systems , vol. 35, pp. 2507– 2521, 2022. 2

    2022

  21. [29]

    Moleculenet: a benchmark for molecular machine learning,

    Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A. S. Pappu, K. Leswing, and V . Pande, “Moleculenet: a benchmark for molecular machine learning,” Chemical Science, vol. 9, no. 2, pp. 513– 530, 2018. 2

    2018

  22. [30]

    Lab-bench: Measuring capabilities of language models for biology research,

    J. M. Laurent, J. D. Janizek, M. Ruzo, M. M. Hinks, M. J. Hammerling, S. Narayanan, M. Ponnapati, A. D. White, and S. G. Rodriques, “Lab-bench: Measuring capabilities of language models for biology research,” arXiv preprint arXiv:2407.10362, 2024. 2

    2024 arXiv

  23. [31]

    Tdcm25: A multi-modal multi-task benchmark for temperature-dependent crystalline materials,

    C. Polat, H. Kurban, E. Serpedin, and M. Kurban, “Tdcm25: A multi-modal multi-task benchmark for temperature-dependent crystalline materials,” in AI for Accelerated Materials Design-ICLR 2025, 2025. 2

    2025

  24. [32]

    Chemlit-qa: A human evaluated dataset for chemistry rag tasks,

    G. Wellawatte, H. Guo, M. Lederbauer, A. Borisova, M. Hart, M. Brucka, and P. Schwaller, “Chemlit-qa: A human evaluated dataset for chemistry rag tasks,” inAI for Accelerated Materials Design-NeurIPS 2024, 2024. 2

    2024

  25. [33]

    Matterchat: A multi-modal llm for material science,

    Y . Tang, W. Xu, J. Cao, W. Gao, S. Farrell, B. Erichson, M. W. Mahoney, A. Nonaka, and Z. Yao, “Matterchat: A multi-modal llm for material science,”arXiv preprint arXiv:2502.13107, 2025. 2

    2025 arXiv

  26. [34]

    xchemagents: Agentic ai for explainable quantum chemistry,

    C. Polat, M. Tuncel, H. Kurban, E. Serpedin, and M. Kurban, “xchemagents: Agentic ai for explainable quantum chemistry,” arXiv preprint arXiv:2505.20574,

    arXiv

  27. [35]

    Hon- eycomb: A flexible llm-based agent system for materi- als science,

    H. Zhang, Y . Song, Z. Hou, S. Miret, and B. Liu, “Hon- eycomb: A flexible llm-based agent system for materi- als science,” arXiv preprint arXiv:2409.00135, 2024. 2

    2024 arXiv

  28. [36]

    Are llms ready for real-world materials discovery?,

    S. Miret and N. M. Krishnan, “Are llms ready for real-world materials discovery?,” arXiv preprint arXiv:2402.05200, 2024. 2

    2024 arXiv

  29. [37]

    H. W. King, CRC Handbook of Chemistry and Physics. CRC Press, 83 ed., 2002. Standard phase data for gold (Au). 2, 10

    2002

  30. [38]

    The fibonacci lattice,

    R. P. Stanley, “The fibonacci lattice,” The Fibonacci Quarterly, vol. 13, no. 3, pp. 215–232, 1975. 3

    1975

  31. [39]

    H. W. King, CRC Handbook of Chemistry and Physics. CRC Press, 83 ed., 2002. Standard phase data for silver (Ag). 10

    2002

  32. [40]

    Wmd-group/hybrid-perovskites: Collection 1 (v1.0)

    A. Walsh, elds22, F. Brivio, and J. M. Frost, “Wmd-group/hybrid-perovskites: Collection 1 (v1.0).” https://doi.org/10.5281/zenodo. 2641358, 2019. Hybrid perovskite CH 3NH3PbI3 structural data. 10

    2019 doi

  33. [41]

    Crystal structure and isothermal compression of fe2o3, cr2o3, and v2o3 to 50 kbars,

    L. W. Finger and R. M. Hazen, “Crystal structure and isothermal compression of fe2o3, cr2o3, and v2o3 to 50 kbars,” Journal of Applied Physics, vol. 51, pp. 5362– 5367, 1980. 10

    1980

  34. [43]

    Mos2 struc- tural properties,

    R. Grau-Crespo and R. Lopez-Cordero, “Mos2 struc- tural properties,” Phys. Chem. Chem. Phys. , vol. 4, p. 4078, 2002. 10

    2002

  35. [44]

    R. W. G. Wyckoff, Crystal Structures Volume 1. Inter- science Publishers, 1963. 10

    1963

  36. [45]

    The rutile structure of sno2,

    W. H. Baur, R. A. Sass, et al., “The rutile structure of sno2,” Acta Crystallographica Section B , vol. 27, p. 2133, 1971. 10

    1971

  37. [46]

    Physics and chemistry of srtio 3 perovskites,

    R. H. Mitchell and M. A. Carpenter, “Physics and chemistry of srtio 3 perovskites,” Physics and Chem- istry of Minerals, vol. 27, p. 583, 2000. 10

    2000

  38. [47]

    Structure of anatase tio2,

    M. Horn, C. R. Meagher, et al., “Structure of anatase tio2,” Zeitschrift f¨ur Kristallographie, vol. 136, p. 273,

  39. [1972]

    Appendix A.1

    10 9 A. Appendix A.1. Crystal Parameters Silver (Ag). Silver adopts an FCC lattice with lattice con- stant a = 4 .0857 ˚A. The cubic crystal belongs to space group Fm3m (No. 225), Pearson symbol cF4, and Schoen- flies notation O5 h. A single Ag atom occupies the origin of the ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.