Pith. sign in

REVIEW 4 major objections 5 minor 150 references

Generative AI can propose new crystals but cannot yet prove they are novel materials.

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

2026-08-01 09:23 UTC pith:Y4AIUO5D

load-bearing objection Useful novelty taxonomy; the evidence-concentration claim needs better empirical support than the Scopus keyword counts provide. the 4 major comments →

arxiv 2607.21660 v1 pith:Y4AIUO5D submitted 2026-07-22 cond-mat.mtrl-sci cs.AIcs.LG

Generative and multimodal AI for materials prediction and design: Progress, challenges, and perspectives

Xianyuan Liu , Charles Anjah , Benjamin E. Jolly , Jonathon F. S. Markanday , Joshua Berry , Haolin Wang , Nicola A. Morley , Robert D. J. Oliver
show 4 more authors
Alexandra J. Ramadan Delvin Ce Zhang Katerina A. Christofidou Haiping Lu
This is my paper
classification cond-mat.mtrl-sci cs.AIcs.LG
keywords multimodal learninggenerative AImaterials designnovelty taxonomystructure-property-performancematerials databenchmarkingdeployment 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.

This Perspective argues that generative and multimodal AI have succeeded at structural novelty—proposing new compositions and crystal structures—but have not established physical or deployment novelty, because the evidence needed to support those higher forms of novelty is missing. It introduces a materials property hierarchy separating intrinsic, composition-determined properties from extrinsic, processing-dependent performance, and a three-level novelty taxonomy. Through that lens, current models are trained on composition and idealized structure, while processing history, microstructure, negative results, and long-duration testing remain scarce and poorly integrated. The paper's central claim is that closing this evidence gap, not improving model capacity alone, is what would let AI design experimentally realisable materials with defensible novelty.

Core claim

The paper's central claim is that novelty in materials discovery cannot be reduced to statistical distance in composition or latent space. It distinguishes three levels: structural novelty (new composition/crystallography), physical novelty (behaviour outside the known distribution under comparable conditions), and deployment novelty (performance advantage under realistic synthesis and operating conditions, including service-life retention). Reviewing multimodal data categories—composition, microstructure, processing, testing/characterisation—the paper finds current evidence is concentrated in composition and idealised structure, while the processing, microstructural, and experimental observ

What carries the argument

The carrying framework is a two-part conceptual structure. The materials property hierarchy ranks information from atomic-scale composition and intrinsic properties up through microstructure, processing history, and engineering performance, explaining why path-dependent extrinsic properties resist prediction from idealised structure alone. Built on it, the three-level novelty taxonomy (structural, physical, deployment) assigns each level the type of evidence required to establish novelty and the type of multimodal data needed. This hierarchy does the argument's work: it converts the abstract question 'can AI design novel materials?' into the concrete question 'does the evidence base reach th

Load-bearing premise

The paper's empirical claim that evidence is concentrated in composition and idealised structure rests on a keyword-based classification of published papers; if that classification misjudges how often processing and microstructure data actually appear in model inputs, the central diagnosis loses its foundation.

What would settle it

Take a random sample of, say, 200 recent AI-for-materials papers and record the data modalities actually used as model inputs, bypassing keyword categories. If processing and microstructural data appear as frequently as composition data, or if testing-and-characterisation data turn out to be used as prediction targets rather than evidence, the paper's central evidence-concentration claim would be contradicted. A narrower check: compare keyword-based counts of 'processing' papers with a manual reading of the same sample to measure classification bias.

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

If this is right

  • If the evidence-concentration claim is right, generative models should be credited with structural novelty only; claims of physical or deployment novelty need independent experimental demonstration.
  • Current benchmarks built on computational labels and proxy novelty metrics will overstate deployability, so evaluation must shift to out-of-distribution splits, synthesis success rates, and experimental confirmation.
  • Building datasets that link each composition to processing route, microstructure, measurement conditions, and negative results becomes a prerequisite for progress, not an optional extra.
  • Process-aware models that treat processing as a primary modality, and generators that return design tuples (composition, route, microstructure, properties, uncertainty), are the concrete direction for moving up the hierarchy.
  • Community-wide metadata standards and deposition of failed/null trials are direct corollaries: without them, higher-level novelty claims remain ungrounded regardless of model advances.

Where Pith is reading between the lines

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

  • One consequence the paper leaves implicit: the taxonomy gives a practical screening rule reviewers could apply—any novelty claim for a processed material should be backed by a processed sample, not a simulated structure, which would immediately separate structural from physical novelty claims.
  • The emphasis on 'compound vs material' suggests that idealized-structure databases may systematically mislead generative models, and that provenance records deserve the same status as composition in benchmark design—a step the paper motivates but does not itself implement.
  • If evidence scarcity is the bottleneck, then automated laboratories that record failed trials are not merely validation tools but data-generation infrastructure; their output should be treated as first-class training data for feasibility-first generation.
  • A testable extension: design a benchmark that holds out entire processing routes for a fixed composition and requires models to predict measured properties; the performance drop relative to random splits would quantify how much processing evidence current models actually use.

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. This Perspective argues that generative and multimodal AI for materials design has made strong progress on 'structural novelty' but lags on 'physical novelty' and 'deployment novelty'. The authors introduce a materials property hierarchy (intrinsic to extrinsic) and a three-level novelty taxonomy, then review multimodal data, representation learning, knowledge integration, generative models, and benchmarks. Their central claim, stated in Sec. 6 and the abstract, is that current evidence available to AI models is concentrated in composition and idealised structure, while processing, microstructure, and experimental characterisation data are under-represented and weakly integrated. This evidence-concentration claim is supported by a Scopus bibliometric analysis (SI S1, Fig. 3), by qualitative discussion of databases and benchmarks (Secs. 3.3 and 5), and by examples from steels, metastable phases, and creep. The paper concludes with four opportunities (multimodal data construction, process-aware modelling, feasibility-first generation, deployment-aware benchmarking) and community recommendations.

Significance. If the central claim is correct, the paper provides a useful framework for assessing AI-driven materials discovery and a clear agenda: shift the field from structure-centric generative models toward process-aware, experimentally grounded, deployment-aware systems. The three-level novelty taxonomy and the emphasis on 'dark data' (missing negative results) are valuable contributions for the community. The paper is a Perspective rather than a methods paper, and it ships reproducible bibliometric data on GitHub, which is a strength. However, the load-bearing empirical assertion about evidence concentration is not adequately established by the presented analysis, and the paper's own Figure 3b appears to contradict it. Since this claim motivates the entire recommendations section, the manuscript needs revision to either provide direct evidence about model training data or carefully reframe the claim as a qualitative observation about data modality usage in benchmarks and databases.

major comments (4)
  1. [Abstract and Sec. 6 vs Fig. 3b] The abstract and Sec. 6 state that current evidence is 'concentrated in composition and idealised structure', but Fig. 3b shows that, in the authors' own bibliometric analysis, composition is the smallest of the four data categories in every year from 2020 to 2025 (around 10–20% of the categorized publications), while testing and characterisation dominates. The paper attempts to reconcile this in Sec. 3.1 by saying composition 'is often used as baseline information ... rather than analysed as a primary data input.' This is a plausible interpretation, but it is an untested assumption: no evidence is given that the 'baseline' usage in the counted publications corresponds to 'evidence available to models.' The reader's and the paper's own numbers do not support the unqualified wording in the abstract. The claim must be rephrased to distinguish 'frequently used as a conditioning or auxiliary
  2. [SI S1, Supplementary Table 3] The bibliometric analysis that underlies the evidence-concentration claim uses Scopus TITLE-ABS-KEY queries with keyword groups that are not comparable in breadth. The 'Composition' group contains relatively narrow terms (e.g., 'stoichiometr*', 'chemical formula'), while the 'Microstructure' group bundles many terms including crystal structure, CIF, space group, defects, and graph representations, and 'Testing and characterisation' lists many named experimental techniques. Consequently, the relative proportions in Fig. 3b could be an artifact of query construction rather than a measure of evidence availability. The authors acknowledge in SI S1 that the analysis is 'inherently approximate,' but no sensitivity analysis is provided (e.g., varying keyword groups, using alternative databases, or checking whether the results are robust to adding/removing synonyms). More fundamentally, publicat
  3. [Sec. 3.3 and Sec. 5 (qualitative support)] The qualitative support for the evidence-concentration claim is selective. For instance, Sec. 3.3 lists Materials Project, OQMD, and JARVIS as composition/structure-centric resources, and Sec. 5.1's Table 2 of benchmarks shows that property-prediction benchmarks indeed rely primarily on composition and structure with DFT labels. However, these examples do not quantify 'concentration' or 'under-representation' across the field. A systematic comparison of, say, the number of entries in open materials databases by modality, or the modality coverage of a representative set of recent generative-model papers, would make the claim more robust. Without such data, the argument in Sec. 6 that 'the evidence available to models is concentrated at the level of composition and idealised structure' remains a plausible but under-determined assertion.
  4. [Sec. 6.1 (deployment-aware benchmarking)] The paper correctly notes the difficulty of distinguishing a genuinely new candidate from a candidate that is simply absent from a limited dataset (Sec. 6.1). However, this caveat also applies to the paper's own bibliometric evidence: the keyword counts cannot distinguish between 'under-represented in the literature' and 'under-represented in model training data.' This reinforces the need for the direct evidence requested above. The point is not to reject the paper's thesis, but to ensure that the framing matches the evidence level.
minor comments (5)
  1. [Throughout] The manuscript header contains placeholder text 'IOP Publishing J. Phys. Mater vv(yyyy) aaaaaa Authoret al', which should be replaced with proper journal style. Also, 'Authoret al' in the running head is a typo.
  2. [SI S1, Fig. 3a caption] The caption says 'The annual number of publications that match the multimodal AI for materials query and involve at least one of four materials data categories.' This is fine, but the SI text says 'This data was obtained by adding the OR union...' — use 'These data were obtained'.
  3. [Sec. 3.1, Table 1] Table 1 is helpful, but the row for 'Microstructure' lists both atomic-scale and larger-length-scale information under one category, which is very broad. Consider separating 'crystal structure' from 'microstructure' in later discussions, since the central claim treats 'idealised structure' separately from 'microstructure'.
  4. [Sec. 5.1, Table 2] In Table 2, the column header 'Comp.' is used but the table notes say 'Comp., Microstruct., Proc. and Test. & character.' — ensure abbreviations are defined in the caption for readers.
  5. [Sec. 6.2] The recommendation about metadata standards is reasonable, but it could briefly mention that such standards already exist in parts of the community (e.g., via NOMAD, OPTIMADE) to avoid appearing to ignore prior infrastructure work. This is a presentation issue, not a substantive flaw.

Circularity Check

0 steps flagged

No significant circularity: the paper is a perspective whose central claims rest on bibliometric data and literature review, not on fitted inputs or self-citation chains.

full rationale

The paper does not claim a derivation in the sense of equations fitted to data. Its central Sec. 6 assertion—that progress toward physical and deployment novelty is limited by evidence concentrated in composition and idealised structure while higher-level evidence is under-represented—is supported by the Scopus keyword analysis (SI S1, Table 3, Fig. 3) and by qualitative reviews of databases and benchmarks (Secs. 3.3, 5). These are external empirical inputs, not outcomes of fitting the conclusion. The taxonomy in Sec. 2 is a conceptual framework, not estimated from data, and it is used to organize observations rather than to predict them. The self-citations (refs 21, 23) are background examples of multimodal learning and are not load-bearing for the article's main claim. The SI explicitly acknowledges the bibliometric analysis is approximate and intended for broad trends, and Sec. 3.1's explanation that composition appears small because it is often used as baseline information is an interpretive assumption about the data. That is a potential validity concern about the evidence base, not a circular reduction of a prediction to its input. No fitted parameter is renamed as a prediction, and no claim reduces to a self-citation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

This Perspective introduces no free parameters or invented entities. Its quantitative claims are based on a keyword-search bibliometric classification and qualitative examples; those assumptions are recorded here.

axioms (4)
  • ad hoc to paper The three-level novelty taxonomy (structural, physical, deployment) is a useful and sufficient organizing scheme for evaluating AI-proposed materials.
    Introduced by this paper (Sec. 2.3); the central framework is not independently established, but is a reasonable conceptual proposal.
  • domain assumption Intrinsic vs extrinsic property hierarchy captures the key distinction between composition/structure-derived and processing-dependent behavior.
    Standard materials-science framing (processing-structure-property-performance), invoked throughout Sec. 2; its applicability in functional materials is acknowledged as blurred.
  • domain assumption Scopus keyword queries in Supplementary Table 3 are a valid proxy for the modality distribution of AI-for-materials research.
    The paper's quantitative claims about under-represented modalities rest on this; it is acknowledged as approximate but not validated.
  • domain assumption Computational reference labels (e.g., DFT) and experimental measurements are not on a single fidelity continuum.
    Load-bearing for the benchmarking critique in Sec. 5.2; argued from examples but not proven in a formal sense.

pith-pipeline@v1.3.0-alltime-deepseek · 20451 in / 14312 out tokens · 131196 ms · 2026-08-01T09:23:03.391138+00:00 · methodology

0 comments
read the original abstract

Artificial intelligence (AI) is accelerating materials prediction and design by enabling efficient exploration of chemical and structural spaces, with particular promise for novel materials discovery. However, novelty in materials discovery encompasses chemical plausibility, structural distinctiveness, property relevance and experimental realisability, making AI-driven novelty claims difficult to substantiate. We introduce a materials property hierarchy, from intrinsic, composition-determined properties to extrinsic, processing-dependent performance, to clarify deployment constraints and distinguish structural, physical and deployment novelty. This framework motivates an evidence-based view of multimodal materials data spanning chemical composition, microstructure, processing, and testing and characterisation, showing that current evidence remains concentrated in composition and idealised structure while heterogeneous, under-represented and weakly integrated modalities limit support for physical and deployment novelty. It also highlights the limitations of benchmarks based mainly on computational labels and proxy novelty criteria. Community-wide standards for data collection, modality alignment and evidence synthesis are needed to support multimodal data construction, process-aware multimodal modelling, feasibility-first generative modelling and deployment-aware benchmarking, so that generative and multimodal AI can design experimentally realisable materials with defensible scientific and practical novelty.

Figures

Figures reproduced from arXiv: 2607.21660 by Alexandra J. Ramadan, Benjamin E. Jolly, Charles Anjah, Delvin Ce Zhang, Haiping Lu, Haolin Wang, Jonathon F. S. Markanday, Joshua Berry, Katerina A. Christofidou, Nicola A. Morley, Robert D. J. Oliver, Xianyuan Liu.

Figure 1
Figure 1. Figure 1: Growth trends of AI for materials research from 2020 to 2025. (a) The annual proportion of multimodal AI, generative AI, and generative and multimodal AI publications relative to all AI for materials publications. The “generative and multimodal AI” term refers to publications that matched both the “multi￾modal AI” and “generative AI” groups. (b) The annual number of AI for materials publications in two key… view at source ↗
Figure 2
Figure 2. Figure 2: Hierarchical pyramid of materials property, from composition-determined intrinsic properties to processing-dependent extrinsic performance. Intrinsic properties are primarily determined by the ground-state chemistry and atomic arrangement. When composition and phase are fixed, these properties are often relatively insensitive to processing history or microstructural variation. This is why crystal structure… view at source ↗
Figure 3
Figure 3. Figure 3: Multimodal AI for materials publications by data category from 2020 to 2025. (a) The annual number of publications that match the multimodal AI for materials query and involve at least one of four materials data categories. (b) Relative proportion of the data categories each year, where publications spanning multiple categories are counted in each. Publication data were retrieved from the Elsevier Scopus d… view at source ↗
Figure 4
Figure 4. Figure 4: Generative and multimodal AI for materials discovery, including adaptive prediction-design loop [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

150 extracted references · 4 linked inside Pith

  1. [1]

    Merchant A, Batzner S, Schoenholz S S, Aykol M, Cheon G and Cubuk E D 2023Nature62480–85

  2. [2]

    2025Nature639624–632

    Zeni C, Pinsler R, Z¨ ugner D, Fowler A, Horton M, Fu X, Wang Z, Shysheya A, Crabb´e J, Ueda Set al. 2025Nature639624–632

  3. [3]

    Hashemi S M, Parvizi S, Baghbanijavid H, Tan A T, Nematollahi M, Ramazani A, Fang N X and Elahinia M 2022International Materials Reviews671–46

  4. [4]

    Kang Y, Park H, Smit B and Kim J 2023Nature Machine Intelligence5309–318

  5. [5]

    Moro V, Loh C, Dangovski R, Ghorashi A, Ma A, Chen Z, Kim S, Lu P Y, Christensen T and Soljaˇci´c M 2025Newton1

  6. [6]

    Wu Y, Ding M, He H, Wu Q, Jiang S, Zhang P and Ji J 2025npj Computational Materials11

  7. [7]

    Joshi C K, Fu X, Liao Y L, Gharakhanyan V, Miller B K, Sriram A and Ulissi Z W 2025 All-atom diffusion transformers: Unified generative modelling of molecules and materialsInternational Conference on Machine Learning

    2025

  8. [8]

    Ye C, Wang Y, Xie X, Zhu T, Liu J, He Y, Zhang L, Zhang J, Fang Z, Wang Let al.2025npj Computational Materials12

  9. [9]

    Chenebuah E T, Nganbe M and Tchagang A B 2024npj Computational Materials10

  10. [10]

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

    2024

  11. [11]

    Xie T and Grossman J C 2018Physical Review Letters120145301

  12. [12]

    Choudhary K and DeCost B 2021npj Computational Materials7

  13. [13]

    Deng B, Zhong P, Jun K, Riebesell J, Han K, Bartel C J and Ceder G 2023Nature Machine Intelligence51031–1041

  14. [14]

    Ye C Y, Weng H M and Wu Q S 2024Computational Materials Today1100003

  15. [15]

    Luo X, Wang Z, Gao P, Lv J, Wang Y, Chen C and Ma Y 2024npj Computational Materials10

  16. [16]

    Khajeh A, Lei X, Ye W, Yang Z, Hung L, Schweigert D and Kwon H K 2025Digital Discovery4 11–20

  17. [17]

    Lambard G, Yamazaki K and Demura M 2023Scientific Reports13566

  18. [18]

    Altoyuri A H, Sarmah A and Jain M K 2024Acta Materialia281120431

  19. [19]

    Jiang X, Wang W, Tian S, Wang H, Lookman T and Su Y 2025npj Computational Materials11

  20. [20]

    Schilling-Wilhelmi M, R ´ıos-Garc´ıa M, Shabih S, Gil M V, Miret S, Koch C T, M´arquez J A and Jablonka K M 2025Chemical Society Reviews541125–1150

  21. [21]

    Liu X, Ouyang B and Zeng Y 2025Nature Computational Science592–94

  22. [22]

    Zhang L, Bing Q, Qin H, Yu L, Li H and Deng D 2025Matter8

  23. [23]

    Liu X, Zhang J, Zhou S, van der Plas T L, Vijayaraghavan A, Grishina A, Zhuang M, Schofield D, Tomlinson C, Wang Yet al.2025Nature Machine Intelligence1–13 14 IOP PublishingJ. Phys. Matervv(yyyy) aaaaaa Authoret al

  24. [24]

    Jablonka K M, Rosen A S, Krishnapriyan A S and Smit B 2023ACS Central Science9563–581

  25. [25]

    Omee S S, Fu N, Dong R, Hu M and Hu J 2024npj Computational Materials10

  26. [26]

    Li K, Rubungo A N, Lei X, Persaud D, Choudhary K, DeCost B, Dieng A B and Hattrick-Simpers J 2025Communications Materials69

  27. [27]

    Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder Get al.2013APL Materials1

  28. [28]

    Kirklin S, Saal J E, Meredig B, Thompson A, Doak J W, Aykol M, R¨ uhl S and Wolverton C 2015npj Computational Materials1

  29. [29]

    Saal J E, Kirklin S, Aykol M, Meredig B and Wolverton C 2013Jom651501–1509

  30. [30]

    Choudhary K, Garrity K F, Reid A C, DeCost B, Biacchi A J, Hight Walker A R, Trautt Z, Hattrick-Simpers J, Kusne A G, Centrone Aet al.2020npj Computational Materials6

  31. [31]

    Antoniuk E R, Cheon G, Wang Get al.2023npj Computational Materials155

  32. [32]

    Walsh A and Zunger A 2017Nature Materials16964–967

  33. [33]

    Xie J, Zhou Y, Faizan M, Li Z, Li T, Fu Y, Wang X and Zhang L 2024Nature Computational Science 4322–333

  34. [34]

    Zheng Y, Xu H, Li Z, Li L, Yu Y, Jiang P, Shi Y, Zhang J, Huang Y, Luo Q, Lou Z and Wang L 2025 Advanced Materials372504378

    2025

  35. [35]

    Yang H, Hu C, Zhou Y, Liu X, Shi Y, Li J, Li G, Chen Z, Chen S, Zeni Cet al.2024arXiv preprint arXiv:2405.04967

    Pith/arXiv arXiv

  36. [36]

    Hayes S M and McCullough E A 2018Resources Policy59192–199

  37. [37]

    Mannan S, Myers R J, Batra R, Mercado R, Wondraczek L and Krishnan N 2026arXiv preprint arXiv:2601.21527

    arXiv

  38. [38]

    Betala S, Gleason S P, Ramlaoui A, Xu A, Channing G, Levy D, Fourrier C, Kazeev N, Joshi C K, Kaba S Oet al.2025arXiv preprint arXiv:2512.04562

    arXiv

  39. [39]

    Baird S G, Sayeed H M, Montoya J and Sparks T D 2024Journal of Open Source Software95618

  40. [40]

    Gruver N, Sriram A, Madotto A, Wilson A G, Zitnick L and Ulissi Z 2024 Fine-tuned language models generate stable inorganic materials as textInternational Conference on Learning Representations

    2024

  41. [41]

    Ong S P, Richards W D, Jain A, Hautier G, Kocher M, Cholia S, Gunter D, Chevrier V L, Persson K A and Ceder G 2013Computational Materials Science68314–319

  42. [42]

    Widdowson D E and Kurlin V A 2026SIAM Journal on Applied Mathematics86898–918

  43. [43]

    Juelsholt M 2026Materials Horizons13(11) 5672–5679

  44. [44]

    Cheetham A K and Seshadri R 2024Chemistry of Materials363490–3495

  45. [45]

    Yu D, Griesemer S, Liu T c, Wolverton C and Zhu Y 2025npj Computational Materials11

  46. [46]

    Meredig B, Agrawal A, Kirklin S, Saal J E, Doak J W, Thompson A, Zhang K, Choudhary A and Wolverton C 2014Physical Review B89094104

  47. [47]

    Ward L, Agrawal A, Choudhary A and Wolverton C 2016npj Computational Materials2

  48. [48]

    Butler K T, Davies D W, Cartwright H, Isayev O and Walsh A 2018Nature559547–555

  49. [49]

    Wang Z, Wang L, Zhang H, Xu H and He X 2024Nano Convergence118

  50. [50]

    Kononova O, Huo H, He T, Rong Z, Botari T, Sun W, Tshitoyan V and Ceder G 2019Scientific Data6 203

  51. [51]

    Lee S, Cruse K, Baibakova V, Ceder G and Jain A 2025Scientific Data

  52. [52]

    Teng Y, Tan H, Huang W and Shan G 2025Communications Physics8 15 IOP PublishingJ. Phys. Matervv(yyyy) aaaaaa Authoret al

  53. [53]

    Wissel S, Scheunert J, Dextre A, Ahmed S, Beyer A, Volz K and Xu B X 2025Materials & Design 115069

  54. [54]

    Wanni J, Bronkhorst C A and J T D 2024npj Computational Materials10

  55. [55]

    Ling J, Hutchinson M, Antono E, DeCost B, Holm E A and Meredig B 2017Materials Discovery10 19–28

  56. [56]

    Tsuruta H and Kumagai M 2025 MatPROV: A provenance graph dataset of material synthesis extracted from scientific literatureNeurIPS 2025 Workshop on AI for Accelerated Materials Design

    2025

  57. [57]

    Kim H, Jeon T, Choi S, Hong J H, Jeon D W, Baek G Y, Kwak G W, Lee D H, Bae J, Lee Cet al.2025 Towards fully-automated materials discovery via large-scale synthesis dataset and expert-level llm-as-a-judgeACM International Conference on Information and Knowledge Managementpp 1302–1312

    2025

  58. [58]

    Zhang J, Yin C, Farbiz F, Jafary-Zadeh M and Sing S L 2025Virtual and Physical Prototyping20 e2592732

  59. [59]

    Deshmukh K, Riensche A, Bevans B, Lane R J, Snyder K, Halliday H S, Williams C B, Mirzaeifar R and Rao P 2024Materials & Design244113136

  60. [60]

    Lederbauer M, Betala S, LI X, Jain A, Sehaba M E A, Channing G, Germain G, Leonescu A, Flaifil F, Amayuelas A, Nozadze A, Schmid S P, Zaki M, Ethirajan S K, Pan E, Franckel M L D, Duval A, Krishnan N M A and Gleason S P 2025 Lemat-synth: a multi-modal toolbox to curate broad synthesis procedure databases from scientific literatureNeurIPS 2025 Workshop on ...

    2025

  61. [61]

    Salgado J E, Lerman S, Du Z, Xu C and Abdolrahim N 2023npj Computational Materials9

  62. [62]

    Oviedo F, Ren Z, Sun S, Settens C, Liu Z, Hartono N T P, Ramasamy S, DeCost B L, Tian S I, Romano Get al.2019npj Computational Materials5

  63. [63]

    Paulus B and Biskup T 2023Digital Discovery2234–244

  64. [64]

    Statt M J, Rohr B A, Guevarra D, Suram S K, Morrell T E and Gregoire J M 2023Scientific Data10 184

  65. [65]

    Swain M C and Cole J M 2016Journal of Chemical Information and Modeling561894–1904

    1904

  66. [66]

    Mavracic J, Court C J, Isazawa T, Elliott S R and Cole J M 2021Journal of Chemical Information and Modeling614280–4289

  67. [67]

    Polak M P and Morgan D 2024Nature Communications151569

  68. [68]

    Dagdelen J, Dunn A, Lee S, Walker N, Rosen A S, Ceder G, Persson K A and Jain A 2024Nature Communications151418

  69. [69]

    Hira K, Zaki M, Krishnan Net al.2025arXiv preprint arXiv:2509.10448

    Pith/arXiv arXiv

  70. [70]

    Li C, Xian Y, Zhou Y, Ding X, Sun J and Xue D 2026Advanced Materialse20478

  71. [71]

    Debnath A, Krajewski A M, Sun H, Lin S, Ahn M, Li W, Priya S, Singh J, Shang S, Beese A M, Liu Z K and Reinhart W F 2021Journal of Materials Informatics13

  72. [72]

    Reeves-McLaren N and Christensen S M L 2026Journal of Materials Chemistry A14276–283

  73. [73]

    Berry J and Christofidou K A 2025Materials Science and Technology41773–790

  74. [74]

    Wang A Y T, Murdock R J, Kauwe S K, Oliynyk A O, Gurlo A, Brgoch J, Persson K A and Sparks T D 2020Chemistry of Materials324954–4965

  75. [75]

    Chang R, Wang Y X and Ertekin E 2022npj Computational Materials8

  76. [76]

    Xu P, Ji X, Li M and Lu W 2023npj Computational Materials9

  77. [77]

    Che L, He Z, Zheng K, Si T, Ge M, Cheng H and Zeng L 2023Materials Today Communications37 107531 16 IOP PublishingJ. Phys. Matervv(yyyy) aaaaaa Authoret al

  78. [78]

    Wong R, Tran A, Dovgyy B, Maldonado C S and Pham M S 2025Materials & Design256114301

  79. [79]

    Gianola D S, della Ventura N M, Balbus G H, Ziemke P, Echlin M P and Begley M R 2023Current Opinion in Solid State and Materials Science27101090

  80. [80]

    Mouritz A P 2012Introduction to aerospace materials(Elsevier) ISBN 0857095153

Showing first 80 references.