Pith. sign in

REVIEW 5 major objections 8 minor 1 cited by

TopoMAS: Large Language Model Driven Topological Materials Multiagent System

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multi-agent workflow lets a 72-billion-parameter model match language models three times its size on topological-materials tasks while using fewer tokens, and the same loop produced the candidate material SrSbO3.

desk verdict Useful engineering with a plausible efficiency story; the headline accuracy number is not trustworthy until TopoQA's provenance is disclosed. read the letter →

arxiv 2507.04053 v1 pith:G74U3BLK submitted 2025-07-05 cond-mat.mtrl-sci cs.AI

Baohua Zhang , Xin Li , Huangchao Xu , Zhong Jin , Quansheng Wu , Ce Li This is my paper
classification cond-mat.mtrl-scics.AI
keywords topologicalmaterialsmulti-agentsystemlargelanguagemodelsknowledgegraphdiscoveryfirst-principlescalculationcrystalstructuregenerationLLMefficiency
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

TopoMAS is a multi-agent system that wraps a single large language model in a team of specialized agents: one queries the Materials Project database, another searches the arXiv literature, one interrogates a curated topological-materials knowledge graph, one generates new crystal structures, and a computing agent prepares, runs, and validates density-functional-theory calculations. The paper's central claim is that this orchestration lets a relatively small model, Qwen2.5-72B, sit alongside much larger models: it scores 94.55% accuracy on the system's 110-question TopoQA benchmark, consumes only 74.3–78.4% of the tokens that Qwen3-235B uses, and responds about twice as fast. The same closed loop, in which validated computation results are written back into the knowledge graph, produced a previously unknown material, SrSbO3, classified by first-principles calculation as a topological crystalline insulator. If the claims hold, the significance is practical and architectural: scientific reasoning in a specialized domain can be bought with task decomposition and tool use rather than with raw model scale, and the discovery pipeline can run end-to-end under natural-language control.

What carries the argument

The load-bearing object is the three-tiered multi-agent architecture itself. A Core Agent orchestrates an execution layer of five specialized agents — the MP Agent for database retrieval through six generated API tools, the Lit. Agent for literature search, the KG Agent that translates natural language into graph queries over the curated TopoKG knowledge graph, the MG Agent for crystal-structure generation, and the CP Agent for first-principles computation — coordinated through a reasoning-and-acting loop and, in the computation subsystem, a state-machine workflow that cycles through Planning, Selection, Action, and Review phases with checkpointing and automated error recovery. The second load-bearing mechanism is the self-evolving knowledge graph: every validated computation is extracted into structured records and stored, so a later query about the same material is answered instantly from the graph instead of re-running the calculation. The paper's efficiency hypothesis is that these mechanisms substitute for parameter count: tool access and orchestrated decomposition let a smaller model perform at the level of a larger one.

What would settle it

Rebuild the TopoQA evaluation after removing every question whose correct answer can be obtained by a single graph lookup or a single API call against TopoKG or the Materials Project; if accuracy on the remaining multi-hop questions drops far below the reported 94.55%, the reasoning claim fails. Independently, recompute SrSbO3's electronic structure with a different first-principles package or a different topological-invariant routine and check whether the topological-crystalline-insulator classification survives.

Watch

Extended reading notes

Core claim

The paper reports that a hierarchical multi-agent architecture can automate the whole topological-materials discovery pipeline: interpreting a user's query, retrieving data from five integrated sources, generating candidate crystal structures, executing VASP density-functional-theory calculations with fault-tolerant scheduling, and writing the validated results back into a knowledge graph for future reuse. Its headline empirical findings are that the lightweight Qwen2.5-72B model attains 94.55% accuracy on a 110-question topology benchmark against 90.90% for Qwen3-235B and 96.40% for DeepSeek-V3, while using fewer tokens and half the latency of Qwen3; that the agent-driven property-retrieval pipeline roughly triples the previous best MAD:MAE score on the LLM4Mat-Bench subset (14.421 versus 5.317); and that the workflow generated and then computationally verified SrSbO3 as a topological crystalline insulator. The authors conclude that model scale is not the decisive factor in specialized domains, and that the system constitutes a transferable paradigm for computation-driven materials discovery.

Load-bearing premise

The TopoQA benchmark's 110 ground-truth answers were created independently of the TopoKG knowledge graph that the KG Agent retrieves from — if the questions were derived from that graph or from the same underlying databases, the reported accuracy would measure template translation and database coverage rather than reasoning.

Editorial extensions

If this is right

  • Model scale is not decisive in specialized domains: a 72-billion-parameter model embedded in the agent pipeline matches or nearly matches 200-billion-parameter models, so capability can be bought with orchestration rather than parameters.
  • Computed results written back into the knowledge graph make the system faster over time: repeated queries about an already-analyzed material are answered from the graph, bypassing redundant density-functional-theory runs.
  • The same framework is claimed to be transferable to other materials-science domains, extending the query-reason-calculate-update loop beyond topological properties.
  • The SrSbO3 result, if correct, demonstrates that the pipeline can generate a plausible new crystalline compound and computationally certify its topology in an end-to-end workflow.

Reading between the lines

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

  • The efficiency advantage of the smaller model is most plausible for tasks whose bottleneck is tool selection and token overhead; on open-ended reasoning without retrieval tools, the gap between small and large models may reappear.
  • The TopoQA accuracy figure should be read as an upper bound on system capability until the benchmark is shown to require multi-step reasoning that cannot be satisfied by a single graph query.
  • If the pattern holds, the design principle — invest in orchestration, tools, and memory rather than raw parameters — is testable in adjacent materials domains such as battery and catalyst discovery, where the same retrieval-and-compute loop applies.
  • SrSbO3 is a computational prediction; its status as a topological insulator will not be settled until the phase is reproduced experimentally or by an independent calculation chain.
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

5 major / 8 minor

Summary. This manuscript presents TopoMAS, a hierarchical multi-agent LLM system for topological materials research. The system integrates Materials Project, arXiv, a curated knowledge graph (TopoKG), crystal generators (Conv-CDVAE, CrystalFormer), and VASP/SymTopo computation into a closed loop with automatic knowledge graph updates. The main claims are: (1) on the 110-question TopoQA benchmark, Qwen2.5-72B achieves 94.55% accuracy while using fewer tokens than Qwen3-235B and DeepSeek-V3; (2) on LLM4Mat-Bench's Materials Project subset, the MP Agent's API retrieval achieves a MAD:MAE of 14.421 versus 5.37 for MatBERT-109M; (3) the system guided identification of SrSbO3 as a topological crystalline insulator, confirmed by first-principles calculations.

Significance. The token-efficiency result is practically relevant if the evaluation is sound: it suggests that a 72B model with well-designed ReAct tooling can match a 235B model on structured query tasks. The closed-loop DFT workflow and automatic TopoKG updates are useful engineering components. However, the TopoQA accuracy is vulnerable to circularity (the KG Agent retrieves from the same graph that may have generated the ground truths), and the LLM4Mat-Bench comparison is against predictive models rather than equivalent retrieval systems. The SrSbO3 novelty claim is not substantiated. The architectural idea is promising, but the quantitative evidence as presented needs re-framing and additional controls.

major comments (5)
  1. [§2.2, §3.1.1, §3.3] The TopoQA dataset description in §3.1.1 does not state how the 110 questions and their 'true answers' were constructed, while §2.2 describes TopoKG as containing exactly the queried attributes (space groups, crystal systems, band gaps, topological classifications) and §2.3 describes the KG Agent as answering by translating natural language into Cypher queries over TopoKG. If the ground truths were taken from TopoKG or from the same source databases that populate it, the reported 94.55% accuracy measures graph retrieval and template matching rather than reasoning or domain knowledge. The paper must disclose the provenance of the questions and answers, and if they are dependent on TopoKG, either re-evaluate on an externally sourced benchmark or explicitly re-label the metric as retrieval success.
  2. [§3.2, Table 1] The comparison with MatBERT-109M and LLM-Prop is not like-for-like. Those baselines are predictive models tested on material representations (CIF, composition, text), whereas the MP Agent directly queries Materials Project APIs for properties of Materials Project entries, and the test subset itself is drawn from Materials Project. The 168.4% regression 'improvement' is therefore largely mechanical and expected. Please reframe this result as a retrieval-versus-prediction comparison and, ideally, add a baseline that also has API access to Materials Project.
  3. [§4 vs Table 2] The conclusion states that Qwen2.5-72B 'consumes 74-83% fewer tokens than larger models', but Table 2 reports total token counts of 566,258 for Qwen2.5-72B versus 722,308 for Qwen3 and 681,960 for DeepSeek-V3. These are 78.4% and 83.0% of the larger models' totals, i.e., 21.6% and 17.0% fewer. The direction and magnitude of the conclusion's claim are inconsistent with the paper's own data and should be corrected.
  4. [§3.3, Table 2] With only 110 TopoQA questions, 94.55% vs 96.40% corresponds to 104 vs 106 correct answers, and the difference is within sampling uncertainty. The text should not assert without qualification that DeepSeek-V3 'demonstrates the highest accuracy' or that Qwen2.5-72B is 'near-state-of-the-art'; a statistical test or confidence intervals are needed before ordering the models.
  5. [Abstract, §3.5.2] The manuscript calls SrSbO3 a 'novel topological phase' but provides no evidence that SrSbO3 is absent from existing crystal databases or prior literature. The DFT+SymTopo calculation confirms a topological crystalline insulator classification for the generated structure, not uniqueness or novelty. Please add a database/literature search or soften the novelty claim.
minor comments (8)
  1. [§3.1.1] There is a typo in 'topological classiffcations'; also specify the number of questions per property type in TopoQA.
  2. [§3.1.2] Clarify how 'expert-validated answer matching' was performed for TopoQA (e.g., exact string match, normalized comparison, or an LLM judge), since this directly affects the reported accuracy values.
  3. [§3.2] State whether the baseline models (MatBERT-109M, LLM-Prop) were evaluated on the same cleaned 8,579-entry subset; if not, the comparison in Table 1 is not controlled.
  4. [§3.4] The use of DeepSeek R1-0528 as the 'third-party' evaluator for TopoOQ should include an assessment of judge bias and inter-judge agreement with human experts, given that LLM-based evaluation of another LLM is not known to be unbiased.
  5. [§2.3] Provide more detail on the 'over 50 pre-encoded description-Cypher pattern pairs' and the template matching algorithm, as the KG Agent's performance depends on this pre-encoded library.
  6. [General] No data or code availability statement is included; please provide links to TopoKG, the question sets, and the agent code to support reproducibility.
  7. [§2.1, Figure 1] The architecture diagram appears low-resolution in the current version; ensure figure labels are legible in the final PDF.
  8. [References] Reference 52 (Jain et al.) has a placeholder year field; complete the citation.

Circularity Check

1 steps flagged · score 6.0 of 10

LLM4Mat-Bench 'prediction' reduces to MP-API retrieval of the ground-truth labels; TopoQA provenance is unstated, so the headline accuracy is not independently shown to be non-circular.

  1. self definitional [Section 3.1.1 (LLM4Mat-Bench test subset) and Section 3.2 (reported improvement)]
    "For our testing dataset, we selected 10% of the test set from the LLM4Mat-Bench paper, specifically entries sourced from the Materials Project. ... Our system has made significant progress: regression performance is enhanced by 168.4% (WTD. Avg. MAD:MAE = 14.421) ... These improvements are attributed to the MP Agent's six specialized tools: ... This focused execution proves that our framework can accurately convert natural language queries into optimized API workflows."

    The test entries and their property labels come from the Materials Project, and the MP Agent's design is to 'execute structured data retrieval and extraction from the Materials Project' via API tools (Section 2.3). At evaluation time the agent therefore returns the very database record that defines each ground-truth value; the MAD:MAE and AUC measures accordingly reduce to API-call success and JSON parsing. Comparing this to MatBERT/LLM-Prop, which must predict the same properties from Composition/CIF/text without accessing MP, is not a comparison of predictive models: the improvement is forced by letting the 'predictor' read the answer key. This is a constructional reduction, not evidence of multi-agent reasoning or domain knowledge.

full rationale

The clearest demonstrated circularity is the LLM4Mat-Bench evaluation: the paper selects test entries 'sourced from the Materials Project' and then attributes its near-perfect scores to the MP Agent's 'optimized API workflows' that retrieve from the same Materials Project. By construction, the agent's output is the ground-truth record, so the reported 168.4% improvement over text-based predictors is a retrieval-coverage result rather than a property-prediction result. The TopoQA headline (94.55% accuracy) is more concerning: TopoQA's '110 questions and true answers' cover space groups, band gaps, and topological classifications, which are exactly the attributes stored in TopoKG, and the KG Agent answers by querying TopoKG. However, the paper never states whether the TopoQA ground truths were constructed from TopoKG, from the same underlying databases, or independently from expert knowledge, so this remains an unproven risk rather than a demonstrated circular step. The SrSbO3 case is not circular: the workflow generates candidates conditioned on user-specified elements and space group, then verifies them with independent VASP/SymTopo first-principles calculations. Self-citations to the authors' prior TopoChat work are legitimate prior art, but they would become load-bearing only if TopoQA were shown to be derived from the self-built TopoKG. Overall, one central benchmark claim reduces by construction, giving partial circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated; SrSbO3 is a candidate material, not an invented mediator or force. The agent framework and TopoKG are software artifacts, not entities needing independent falsifiable handles.

free parameters (1)
  • Cypher template library size = over 50 templates (exact number not given)
    The KG Agent's two-stage query optimization relies on a hand-crafted library of description-Cypher pattern pairs (Section 2.3). Its coverage determines how well natural-language questions map to correct knowledge graph queries; no ablation measures sensitivity.
assumptions (5)
  • domain assumption Materials Project database entries are accurate and current
    MP Agent retrieves properties and the LLM4Mat-Bench evaluation fetches ground-truth-like values from the same database (Sections 2.2, 2.3).
  • domain assumption SymTopo correctly computes topological invariants from VASP outputs
    The SrSbO3 topological crystalline insulator classification rests on SymTopo's analysis of DFT wavefunctions (Sections 2.4, 3.5.2, Stage 3).
  • domain assumption Generated crystal structures from CrystalFormer and Conv-CDVAE are physically valid and chemically reasonable
    The MG Agent returns candidates such as SrSbO3 without an explicit stability filter beyond the generator's prior (Section 2.3).
  • domain assumption TopoQA ground-truth answers are independent of the TopoKG retrieval source
    The benchmark's validity depends on this independence, which the paper does not demonstrate (Sections 2.2, 3.1.1).
  • domain assumption DeepSeek R1-0528 scores are a reliable third-party quality measure
    Open-ended TopoOQ quality scores come from a single LLM judge with no human validation or inter-rater check (Section 3.1.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of TopoMAS: Large Language Model Driven Topological Materials Multiagent System." pith.science (2026). https://pith.science/paper/G74U3BLK

@misc{pith2026250704053,
  author       = {Pith},
  title        = {Pith review of: TopoMAS: Large Language Model Driven Topological Materials Multiagent System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G74U3BLK}},
  note         = {Machine review of arXiv:2507.04053}
}
read the original abstract

Topological materials occupy a frontier in condensed-matter physics thanks to their remarkable electronic and quantum properties, yet their cross-scale design remains bottlenecked by inefficient discovery workflows. Here, we introduce TopoMAS (Topological materials Multi-Agent System), an interactive human-AI framework that seamlessly orchestrates the entire materials-discovery pipeline: from user-defined queries and multi-source data retrieval, through theoretical inference and crystal-structure generation, to first-principles validation. Crucially, TopoMAS closes the loop by autonomously integrating computational outcomes into a dynamic knowledge graph, enabling continuous knowledge refinement. In collaboration with human experts, it has already guided the identification of novel topological phases SrSbO3, confirmed by first-principles calculations. Comprehensive benchmarks demonstrate robust adaptability across base Large Language Model, with the lightweight Qwen2.5-72B model achieving 94.55% accuracy while consuming only 74.3-78.4% of tokens required by Qwen3-235B and 83.0% of DeepSeek-V3's usage--delivering responses twice as fast as Qwen3-235B. This efficiency establishes TopoMAS as an accelerator for computation-driven discovery pipelines. By harmonizing rational agent orchestration with a self-evolving knowledge graph, our framework not only delivers immediate advances in topological materials but also establishes a transferable, extensible paradigm for materials-science domain.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Agentic Exploration of Physics Models

    cs.AI 2025-09 conditional novelty 7.0 of 10

    A general-purpose LLM agent can discover physics models, including ODEs and spin Hamiltonians, by autonomously choosing experiments and fitting hypotheses to numeric data.

Reference graph

Works this paper leans on

62 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    M. Z. Hasan and C. L. Kane, Reviews of modern physics, 2010, 82, 3045--3067

    work page 2010

  2. [2]

    Qi and S.-C

    X.-L. Qi and S.-C. Zhang, Reviews of modern physics, 2011, 83, 1057--1110

    work page 2011

  3. [3]

    Kitagawa, M

    T. Kitagawa, M. A. Broome, A. Fedrizzi, M. S. Rudner, E. Berg, I. Kassal, A. Aspuru-Guzik, E. Demler and A. G. White, Nature communications, 2012, 3, 882

    work page 2012

  4. [4]

    C.-K. Chiu, J. C. Teo, A. P. Schnyder and S. Ryu, Reviews of Modern Physics, 2016, 88, 035005

    work page 2016

  5. [5]

    Bradlyn, L

    B. Bradlyn, L. Elcoro, J. Cano, M. G. Vergniory, Z. Wang, C. Felser, M. I. Aroyo and B. A. Bernevig, Nature, 2017, 547, 298--305

    work page 2017

  6. [6]

    B. A. Bernevig, T. L. Hughes and S.-C. Zhang, science, 2006, 314, 1757--1761

    work page 2006

  7. [7]

    Fu and C

    L. Fu and C. Kane, APS March Meeting Abstracts, 2008, pp. V10--001

    work page 2008

  8. [8]

    Vergniory, L

    M. Vergniory, L. Elcoro, C. Felser, N. Regnault, B. A. Bernevig and Z. Wang, Nature, 2019, 566, 480--485

    work page 2019

Show all 62 references
  1. [9]

    Armitage, E

    N. Armitage, E. Mele and A. Vishwanath, Reviews of Modern Physics, 2018, 90, 015001

    2018

  2. [10]

    Zhang, Y

    T. Zhang, Y. Jiang, Z. Song, H. Huang, Y. He, Z. Fang, H. Weng and C. Fang, Nature, 2019, 566, 475--479

    2019

  3. [11]

    Z. Wang, H. Hua, W. Lin, M. Yang and K. C. Tan, Crystalline Material Discovery in the Era of Artificial Intelligence, 2025, https://arxiv.org/abs/2408.08044

    2025

  4. [12]

    T. Xie, X. Fu, O.-E. Ganea, R. Barzilay and T. Jaakkola, arXiv preprint arXiv:2110.06197, 2021

    2021 arXiv

  5. [13]

    R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu and Y. Liu, Advances in Neural Information Processing Systems, 2023, 36, 17464--17497

    2023

  6. [14]

    R. Jiao, W. Huang, Y. Liu, D. Zhao and Y. Liu, arXiv preprint arXiv:2402.03992, 2024

    2024 arXiv

  7. [15]

    Z. Cao, X. Luo, J. Lv and L. Wang, arXiv preprint arXiv:2403.15734, 2024

    2024

  8. [16]

    Y. Chen, X. Wang, X. Deng, Y. Liu, X. Chen, Y. Zhang, L. Wang and H. Xiao, arXiv preprint arXiv:2408.07608, 2024

    2024 arXiv

  9. [17]

    Flam-Shepherd and A

    D. Flam-Shepherd and A. Aspuru-Guzik, arXiv preprint arXiv:2305.05708, 2023

    2023 arXiv

  10. [18]

    Sriram, B

    A. Sriram, B. Miller, R. T. Chen and B. Wood, Advances in Neural Information Processing Systems, 2024, 37, 46025--46046

    2024

  11. [19]

    B. K. Miller, R. T. Chen, A. Sriram and B. M. Wood, Forty-first International Conference on Machine Learning, 2024

    2024

  12. [20]

    X. Luo, Z. Wang, J. Lv, L. Wang, Y. Wang and Y. Ma, arXiv preprint arXiv:2412.11693, 2024

    2024

  13. [21]

    Ye, H.-M

    C.-Y. Ye, H.-M. Weng and Q.-S. Wu, Computational Materials Today, 2024, 1, 100003

    2024

  14. [22]

    X. Luo, Z. Wang, P. Gao, J. Lv, Y. Wang, C. Chen and Y. Ma, npj Computational Materials, 2024, 10, 254

    2024

  15. [23]

    Z. Qiu, L. Jin, Z. Du, H. Chen, Y. Cen, S. Sun, Y. Mei and H. Zhang, arXiv preprint arXiv:2409.06191, 2024

    2024 arXiv

  16. [24]

    Cao and L

    Z. Cao and L. Wang, CrystalFormer-RL: Reinforcement Fine-Tuning for Materials Design, 2025, https://arxiv.org/abs/2504.02367

    2025

  17. [25]

    Z. Li, S. Liu, B. Ye, D. J. Srolovitz and T. Wen, Active Learning for Conditional Inverse Design with Crystal Generation and Foundation Atomic Models, 2025, https://arxiv.org/abs/2502.16984

    2025 arXiv

  18. [26]

    Yamazaki, W

    S. Yamazaki, W. Nong, R. Zhu, K. S. Novoselov, A. Ustyuzhanin and K. Hippalgaonkar, Multi-property directed generative design of inorganic materials through Wyckoff-augmented transfer learning, 2025, https://arxiv.org/abs/2503.16784

    2025 arXiv

  19. [27]

    C. Ye, Y. Wang, X. Xie, T. Zhu, J. Liu, Y. He, L. Zhang, J. Zhang, Z. Fang, L. Wang, Z. Liu, H. Weng and Q. Wu, Materials discovery acceleration by using condition generative methodology, 2025, https://arxiv.org/abs/2505.00076

    2025

  20. [28]

    C. Zeni, R. Pinsler, D. Z \"u gner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabb \'e , S. Ueda et al., Nature, 2025, 1--3

    2025

  21. [29]

    C. K. Joshi, X. Fu, Y.-L. Liao, V. Gharakhanyan, B. K. Miller, A. Sriram and Z. W. Ulissi, arXiv preprint arXiv:2503.03965, 2025

    2025 arXiv

  22. [30]

    X.-Q. Han, Z. Ouyang, P.-J. Guo, H. Sun, Z.-F. Gao and Z.-Y. Lu, Chinese Physics Letters, 2025

    2025

  23. [31]

    K. L. K. Lee, C. Gonzales, M. Spellings, M. Galkin, S. Miret and N. Kumar, Proceedings of the SC'23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis, 2023, pp. 51--59

    2023

  24. [32]

    K. Das, B. Samanta, P. Goyal, S.-C. Lee, S. Bhattacharjee and N. Ganguly, Proceedings of the AAAI Conference on Artificial Intelligence, 2023, pp. 7323--7331

    2023

  25. [33]

    R. Dong, Y. Song, E. M. Siriwardane and J. Hu, Advanced Intelligent Systems, 2023, 5, 2300141

    2023

  26. [34]

    Huang, R

    H. Huang, R. Magar, C. Xu and A. B. Farimani, arXiv preprint arXiv:2308.16259, 2023

    2023 arXiv

  27. [35]

    A. N. Rubungo, C. Arnold, B. P. Rand and A. B. Dieng, arXiv preprint arXiv:2310.14029, 2023

    2023 arXiv

  28. [36]

    G. Song, D. Fu, Z. Qiu, Z. Yang, J. Dai, L. Ma and D. Zhang, arXiv preprint arXiv:2312.09744, 2023

    2023 arXiv

  29. [37]

    T. Xie, Y. Wan, Y. Liu, Y. Zeng, S. Wang, W. Zhang, C. Grazian, C. Kit, W. Ouyang, D. Zhou et al., arXiv preprint arXiv:2412.11970, 2024

    2024 arXiv

  30. [38]

    Ansari and S

    M. Ansari and S. M. Moosavi, Digital Discovery, 2024, 3, 2607--2617

    2024

  31. [39]

    B. Hu, S. Liu, B. Ye, Y. Hao and T. Wen, arXiv preprint arXiv:2411.16416, 2024

    2024

  32. [40]

    N. H. Park, T. J. Callahan, J. L. Hedrick, T. Erdmann and S. Capponi, arXiv preprint arXiv:2408.11793, 2024

    2024 arXiv

  33. [41]

    Ghafarollahi and M

    A. Ghafarollahi and M. J. Buehler, arXiv preprint arXiv:2407.10022, 2024

    2024 arXiv

  34. [42]

    Bazgir, Y

    A. Bazgir, Y. Zhang et al., AI for Accelerated Materials Design-ICLR 2025, 2025

    2025

  35. [43]

    Ghafarollahi and M

    A. Ghafarollahi and M. J. Buehler, arXiv preprint arXiv:2410.13768, 2024

    2024 arXiv

  36. [44]

    Z. Ni, Y. Li, K. Hu, K. Han, M. Xu, X. Chen, F. Liu, Y. Ye and S. Bai, arXiv preprint arXiv:2411.08063, 2024

    2024 arXiv

  37. [45]

    Ni and M

    B. Ni and M. J. Buehler, Extreme Mechanics Letters, 2024, 67, 102131

    2024

  38. [46]

    T. Song, M. Luo, X. Zhang, L. Chen, Y. Huang, J. Cao, Q. Zhu, D. Liu, B. Zhang, G. Zou et al., Journal of the American Chemical Society, 2025, 147, 12534--12545

    2025

  39. [47]

    Lai and Y

    Z. Lai and Y. Pu, arXiv preprint arXiv:2504.08810, 2025

    2025 arXiv

  40. [48]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford et al., arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  41. [49]

    H. Xu, B. Zhang, Z. Jin, T. Zhu, Q. Wu and H. Weng, arXiv preprint arXiv:2409.13732, 2024

    2024

  42. [50]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang et al., arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  43. [51]

    Hafner, Journal of computational chemistry, 2008, 29, 2044--2078

    J. Hafner, Journal of computational chemistry, 2008, 29, 2044--2078

    2008

  44. [52]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder et al., APL materials, 2013, 1, year

    2013

  45. [53]

    Mavroudis, Preprints, 2024

    V. Mavroudis, Preprints, 2024

    2024

  46. [54]

    Y. He, Y. Jiang, T. Zhang, H. Huang, C. Fang and Z. Jin, Chinese Physics B, 2019, 28, 087102

    2019

  47. [55]

    H. Xiao, H. Wu and X. Chi, International Conference on Networks for Grid Applications, 2008, pp. 35--42

    2008

  48. [56]

    Z. R. T. N. J. L. S. Robert M. Hanson, Jaime Prilusky, Israel Journal of Chemistry, 2013, pp. 207--216

    2013

  49. [57]

    A. N. Rubungo, K. Li, J. Hattrick-Simpers and A. B. Dieng, Machine Learning: Science and Technology, 2025, 6, 020501

    2025

  50. [58]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan et al., arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  51. [59]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv et al., arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  52. [60]

    Choudhary and B

    K. Choudhary and B. DeCost, npj Computational Materials, 2021, 7, 185

    2021

  53. [61]

    chemist agents

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi et al., arXiv preprint arXiv:2501.12948, 2025 mcitethebibliography main.tex0000664000000000000000000016750615032230400011230 0ustar rootroot [twoside,twocolumn,9pt] article extsizes [super,sort&...

    2025 arXiv

  54. [62]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.