Pith. sign in

REVIEW 3 major objections 4 minor 5 cited by

An LLM agent that wraps VASP calculations in expert-defined workflows and guardrails completes 98% of automation tasks, up from 66% for a standalone GPT-4o, and raises result accuracy by more than 27 points.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:40 UTC pith:HE3IGIEA

load-bearing objection A useful, reusable benchmark and an honest failure analysis are the real contributions; the accuracy metric partly measures consistency with the authors' own reference calculations, and a promised pymatgen comparison is missing. the 3 major comments →

arxiv 2512.19458 v2 pith:HE3IGIEA submitted 2025-12-22 cs.AI cond-mat.mtrl-sci

VASP Agent: An Agentic Framework for Autonomous First-principles Calculations

Zeyu Xia , Jinzhe Ma , Congjie Zheng , Zhongyao Wang , Shufei Zhang , Yuqiang Li , Hang Su , P. Hu
show 6 more authors
Changshui Zhang Xingao Gong Wanli Ouyang Lei Bai Dongzhan Zhou Mao Su
This is my paper
classification cs.AI cond-mat.mtrl-sci
keywords LLM agentsVASPfirst-principles calculationmaterials simulationdensity functional theoryautonomous workflowbenchmarkcomputational materials science
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a domain-specialized LLM agent—built around a library of validated VASP workflows, modular tool components, and hierarchy-driven prompt templates—can run multi-step first-principles calculations end to end far more reliably than a standalone LLM. On a new 80-case benchmark covering relaxation, band structure, adsorption, and transition-state tasks, the agent lifts completion rates for every tested model, with GPT-4o's completion rising from 66.5% to 97.9% and accuracy from 45.7% to 73.1%. The system completes all evaluated cases, and failure analysis shows that errors which terminate fixed pipelines can be diagnosed and recovered under agentic control. If these results hold, routine materials simulations that now require expert parameter tuning could be delegated to autonomous agents, enabling high-throughput computational screening.

Core claim

By embedding domain expertise into an agentic loop rather than relying on prompt engineering alone, the system claims to make multi-step first-principles calculations reliable enough for autonomous execution. It selects a workflow from a library of expert-defined procedures, executes each step through reusable components—file I/O, command execution, parsing, and an LLM-driven parameter generator—and uses runtime evidence to enforce convergent, well-posed settings. Across six LLMs and four task types (structural relaxation, band structure, adsorption energy, transition states), the agent lifts completion and accuracy on every model; GPT-4o completion rises from 66.46% to 97.92% and accuracy f

What carries the argument

The load-bearing machinery is the workflow library, which formalizes expert procedures as sequences of modular components (file I/O, command execution, data parsing, and GetLLMAnswer). GetLLMAnswer generates VASP parameters by filling hierarchical prompt templates that embed domain background, intermediate outputs, and strict output-format constraints. This forces the LLM to produce executable, internally consistent inputs and to respect interdependencies across steps (e.g., matching IBRION with POTIM, keeping cell-relaxation settings fixed across NEB states), preventing the tag-initialization and context-loss failures observed in standalone use.

Load-bearing premise

The benchmark's 'ground truth' values are not fully external: band gaps without experimental data were determined by the authors with hybrid functionals, and adsorption/transition-state energies were computed with the authors' own VASP settings, so the reported accuracy gains may partly measure agreement with the benchmark's own conventions rather than physical correctness.

What would settle it

Hold out a set of materials whose band gaps, adsorption energies, and barriers are known from experiment or from an independent DFT implementation using different parameters; if the agent's accuracy on this blind set drops substantially, the claimed gains are in part self-consistency rather than general scientific competence.

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

If this is right

  • Open-source models integrated with the agent reach completion rates above 90%, narrowing the gap with proprietary models and making secure local deployment feasible for sensitive data.
  • Reasoning-oriented models benefit disproportionately, suggesting that long-range planning consistency is a key bottleneck that the agent helps address.
  • The agent's failure recovery means tasks that would crash a fixed pipeline can be continued, reducing wasted computation in high-throughput settings.
  • The new benchmark offers a reproducible suite of 80 tasks with reference data for evaluating future agents dedicated to computational materials science.

Where Pith is reading between the lines

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

  • Because the accuracy metric is ratio-based and the references were partly generated with the same VASP conventions the agent is coached to reproduce, a blind external test is needed to distinguish self-consistency from physical accuracy.
  • The framework depends on a hand-built workflow library, so novel tasks outside that library would require human extension; true open-ended discovery would demand the agent to construct new workflows, which it does not do.
  • The same decomposition pattern (workflow + reusable components + runtime evidence) could transfer to other simulation codes by swapping the parameter templates, turning this into a general recipe for autonomous computational experiments.
  • The persistent low transition-state accuracy indicates that for the hardest tasks the agent acts as a parameter-tuner rather than a physics reasoner; an agent that models convergence behavior could be a next step.

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

3 major / 4 minor

Summary. The manuscript presents VASP Agent, an LLM-driven agentic framework for executing VASP first-principles calculations. The framework combines a predefined workflow library, modular components (file I/O, command execution, parsing), hierarchical prompts with domain background, and deterministic runtime tools to generate INCAR parameters, run calculations, and parse outputs. The authors construct a benchmark of 80 tasks across four categories — structural relaxation (40 tasks), band structure (24), adsorption energy (10), and transition state (6) — and evaluate six LLMs with and without the agent using completion rate and a ratio-based accuracy metric. They report substantial improvements (e.g., GPT-4o completion rate rising from 66.46% to 97.92% and accuracy from 45.74% to 73.07%). The paper also analyzes failure modes and compares open-source vs proprietary and reasoning vs standard models.

Significance. If the reported gains are robust, the paper makes a useful contribution to automating routine DFT workflows. Its assets include a concrete multi-task benchmark, public code/data links, evaluation across six LLMs, and candid acknowledgment that transition-state accuracy remains low. The framework's design — deterministic tools plus domain-specialized prompts and guardrails — is sensible and likely to improve pipeline reliability. However, the evaluation methodology has two load-bearing weaknesses: no repeated trials despite LLM stochasticity, and a benchmark whose reference values are partly generated with the same parameter conventions the agent is coached to follow. These weaken the broader accuracy and physical-coherence claims. With additional validation, the work could become an important resource for the AI-for-materials community.

major comments (3)
  1. [Methods §4.2 and Results §2.3 (Figures 2–6)] All reported completion and accuracy scores are point estimates; the manuscript does not state the number of independent runs or report error bars or confidence intervals. LLM outputs are stochastic, and per-task counts are small (40 SR, 24 BS, 10 AE, 6 TS), so a single failed run can shift a completion rate by 2.5–16.7 percentage points. The claim of 'robust and consistent gains' therefore lacks statistical support. Please provide repeated runs (at least five per condition), report mean ± standard deviation or confidence intervals, and state the temperature/sampling settings.
  2. [Methods §4.1 (reference generation)] Band-gap references are taken from experiment/literature where available, but otherwise 'determined using hybrid functional calculations' by the authors; all adsorption and transition-state references are computed with the authors' chosen VASP settings (PBE, DFT-D3, 450 eV cutoff, 4×4×1 k-mesh, CI-NEB force convergence <0.05 eV/Å). Since the agent's hierarchical prompts and workflow library embed the same domain conventions, the accuracy metric (min/max ratio in §4.2.2) can be satisfied by reproducing the benchmark's own parameter choices. This does not establish physical correctness or an independent 'high-fidelity' ground truth. Please (i) report how many references are self-computed, (ii) add an independent validation set with experimental or otherwise externally determined values, or (iii) show that accuracy is insensitive to plausible variations in the reference settings.
  3. [Results §2.3 and Methods §4.2.1 (completion vs correctness)] The completion metric awards full credit for execution-level success: for band structure it requires only that the calculation 'completes without errors and produces electronic eigenvalues along specified paths'; for transition states it awards 6 of 10 points for NEB convergence regardless of barrier correctness. The paper's own results show transition-state accuracy remains 'persistently low' despite near-full completion. The abstract's phrase 'reliable end-to-end computational execution' is therefore overstated: high completion rates demonstrate pipeline robustness, not scientifically valid results. Please temper the claims or revise the completion metric to include a physical-validity check (e.g., verifying the band-gap value or barrier against a tolerance).
minor comments (4)
  1. [Discussion] The Discussion states 'benchmark dataset of over 100 materials' whereas §2.2 says '80 practical application computing scenarios'; the task counts in Table 1 sum to 80. Please reconcile the numbers.
  2. [Methods §4.2.2 (Transition State)] The text says a score is received only if the predicted value falls within a 10% relative error threshold, but the equations give a continuous min/max ratio with no threshold. Please specify which behavior is intended, and if the threshold is used, implement it in the scoring equations.
  3. [Figures 3–6] The bar charts aggregate many tasks but the manuscript does not provide a per-task, per-model numerical table. Adding such a table would support the reported percentages and improve reproducibility.
  4. [Methods §4.1 (Adsorption Energy)] Minor wording: DFT-D3 is a dispersion correction; 'further correct intermolecular forces' should be rephrased as 'account for dispersion interactions' or similar.

Circularity Check

0 steps flagged

No significant circularity: the agent's outputs are not derived from the benchmark references by construction.

full rationale

The paper's central claim is that an LLM-agent framework improves the completion rate and accuracy of VASP calculations relative to standalone LLMs. This claim is backed by an empirical benchmark, not by a derivation in which a predicted quantity is defined in terms of the reference. The accuracy metrics compare agent-produced values to independently tabulated reference values (e.g., Materials Project settings, experimental/literature band gaps, or author-computed hybrid-functional results). No equation in the paper makes the prediction equal to the reference by construction, and no model parameter is fitted to the benchmark labels. The fact that some references were generated by the authors using VASP is a benchmark-construction choice; it may affect external validity or fairness, but it does not make the evaluation circular because the agent does not receive the reference values as inputs and must still correctly generate INCAR tags, run VASP, and parse outputs. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in solely via citation. The paper is self-contained as an empirical study of an engineering system, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities or fitted scientific constants are introduced. The free parameters listed are evaluation-construction choices that directly affect the reported accuracy numbers, while the axioms are the domain assumptions needed to treat the benchmark labels as meaningful ground truth.

free parameters (4)
  • SOAP descriptor hyperparameters = cutoff=5.0 Å, nmax=8, lmax=6, sigma=0.5 Å
    Hand-selected in Methods §4.2.2 to define structural similarity for SR accuracy scoring; changing them changes reported SR accuracy.
  • TS relative-error threshold = 10%
    Chosen in Methods §4.2.2 to decide whether transition-state energy predictions contribute to accuracy; directly shapes TS accuracy scores.
  • SR reference INCAR parameters = Materials Project recommended settings
    Reference relaxed structures for the 40 SR tasks are generated with MP parameter defaults; the agent is scored by SOAP similarity to those structures, so copying MP settings can inflate accuracy.
  • k-point density and cutoff for adsorption/TS references = 1/25 Å^-1 spacing; 8x8x8 bulk and 4x4x1 slab grids; 450 eV cutoff
    Domain-practice settings used in Methods §4.1 to generate ground-truth adsorption and transition-state energies; the agent is evaluated against references produced with these choices.
axioms (4)
  • domain assumption DFT/PBE/hybrid calculations in VASP are valid ground truth for band gaps, adsorption energies, and transition-state barriers.
    Methods §4.1 treats computed VASP values as 'high-fidelity reference data' and 'ground-truth references'.
  • ad hoc to paper The predefined workflow library correctly encodes best-practice calculation procedures.
    Section 2.1 states workflows are 'based on established best practices'; the agent's success depends on this library being complete and correct.
  • domain assumption Single-run LLM outputs without repeated sampling are representative of model performance.
    Section 2.3 reports single completion and accuracy numbers per model/task with no seeds or variance, implicitly assuming LLM stochasticity does not dominate results.
  • domain assumption SOAP cosine similarity is a valid measure of structural accuracy.
    Methods §4.2.2 uses SOAP similarity to score SR accuracy; no independent validation of this metric for the task is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 10305 in / 10071 out tokens · 100791 ms · 2026-08-03T14:40:16.724387+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) are increasingly embedded in agentic frameworks for scientific discovery. First-principles materials computation imposes a demanding standard for autonomy: successful execution depends on internally consistent inputs, supervision of long-running calculations, and verified outputs. Here we present VASP Agent, a coding-agent-centered system that combines reusable domain skills, deterministic tools, workspace-state inspection, runtime evidence, and scientific guardrails to execute multi-step VASP calculations. The system is evaluated across multiple tasks including structural relaxation, bandgap calculation, equilibrium lattice constant determination, and CO/Pt(111) adsorption. VASP Agent completes all evaluated cases, and its computed numerical results are compared with those obtained using pymatgen and other agentic tools. When large deviations occur, the calculation parameters produced by VASP Agent are more appropriate than those produced by LLM-based workflows. Failure analysis shows that errors that terminate fixed pipelines can be diagnosed and recovered under agentic control.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

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

  1. MatClaw: An Autonomous Code-First LLM Agent for End-to-End Materials Exploration

    cond-mat.mtrl-sci 2026-04 conditional novelty 7.0

    MatClaw is a code-first LLM agent that autonomously executes end-to-end materials workflows by generating and running Python scripts on remote clusters, achieving reliable code generation via memory architecture and R...

  2. MatClaw: An Autonomous Code-First LLM Agent for End-to-End Materials Exploration

    cond-mat.mtrl-sci 2026-04 conditional novelty 7.0

    MatClaw shows a code-first LLM agent autonomously generating and executing workflows for ML force field training, Curie temperature prediction, and parameter search on CuInP2S6, succeeding on code but requiring interv...

  3. El Agente Quntur: A research collaborator agent for quantum chemistry

    physics.chem-ph 2026-02 unverdicted novelty 7.0

    El Agente Quntur is a new multi-agent system that uses reasoning over literature and software documentation to autonomously handle the full workflow of quantum chemistry experiments in ORCA.

  4. INCARBench: A Benchmark for Scientific Configuration in VASP INCAR by Large Language Models

    cond-mat.mtrl-sci 2026-06 unverdicted novelty 6.0

    INCARBench evaluates 19 LLMs on VASP INCAR configuration generation and repair, showing high semantic accuracy but lower scientific correctness especially for DFT+U, magnetism, and correlated materials.

  5. TSAgent: An Agentic Workflow for Autonomous Transition State Search

    physics.chem-ph 2026-05 unverdicted novelty 6.0

    TSAgent automates transition state searches at DFT accuracy via an agentic loop, reaching 83% success on 100 OC20NEB examples and 70% on 10 held-out cases versus 73% for human experts.

Reference graph

Works this paper leans on

50 extracted references · 2 canonical work pages · cited by 4 Pith papers

  1. [1]

    npj Artificial Intelligence1(1), 14 (2025)

    Zhang, Y., Khan, S.A.,et al.: Exploring the role of large language models in the scientific method: from hypothesis to discovery. npj Artificial Intelligence1(1), 14 (2025)

    2025

  2. [2]

    arXiv preprint arXiv:2406.10833 (2024)

    Zhang, Y., Chen, X., et al.: A comprehensive survey of scientific large lan- guage models and their applications in scientific discovery. arXiv preprint arXiv:2406.10833 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    arXiv preprint arXiv:2505.13259 (2025)

    Zheng, T., Deng, Z., et al.: From automation to autonomy: A survey on large language models in scientific discovery. arXiv preprint arXiv:2505.13259 (2025)

    arXiv 2025

  4. [4]

    arXiv preprint arXiv:2311.07361 (2023)

    AI4Science,M.R.,Quantum,M.A.:Theimpactoflargelanguagemodelsonscien- tific discovery: a preliminary study using gpt-4. arXiv preprint arXiv:2311.07361 (2023)

    Pith/arXiv arXiv 2023

  5. [5]

    https://arxiv.org/abs/2310.07984

    Zheng, Y., Koh, H.Y., et al.: Large Language Models for Scientific Synthesis, Inference and Explanation (2023). https://arxiv.org/abs/2310.07984

    Pith/arXiv arXiv 2023

  6. [6]

    Chemical science (2025)

    Ramos, M.C., Collison, C.J., White, A.D.: A review of large language models and autonomous agents in chemistry. Chemical science (2025)

    2025

  7. [7]

    arXiv preprint arXiv:2312.07559 (2023)

    Lála, J., O’Donoghue, O., et al.: Paperqa: Retrieval-augmented generative agent for scientific research. arXiv preprint arXiv:2312.07559 (2023)

    Pith/arXiv arXiv 2023

  8. [8]

    https://arxiv.org/abs/2404.10981

    Huang, Y., Huang, J.: A Survey on Retrieval-Augmented Text Generation for Large Language Models (2024). https://arxiv.org/abs/2404.10981

    Pith/arXiv arXiv 2024

  9. [9]

    https://arxiv.org/abs/2402.19473

    Zhao, P., Zhang, H., et al.: Retrieval-Augmented Generation for AI-Generated Content: A Survey (2024). https://arxiv.org/abs/2402.19473

    Pith/arXiv arXiv 2024

  10. [10]

    npj Computational Materials10(1), 251 (2024)

    Prince, M.H., Chan, H.,et al.: Opportunities for retrieval and tool augmented large language models in scientific facilities. npj Computational Materials10(1), 251 (2024)

    2024

  11. [11]

    arXiv preprint arXiv:2504.14891 (2025) 17

    Gan, A., Yu, H., et al.: Retrieval augmented generation evaluation in the era of large language models: A comprehensive survey. arXiv preprint arXiv:2504.14891 (2025) 17

    Pith/arXiv arXiv 2025

  12. [12]

    Bran, A.M., Cox, S., et al.: Chemcrow: Augmenting large-language models with chemistry tools (2023) arXiv:2304.05376 [physics.chem-ph]

    Pith/arXiv arXiv 2023

  13. [13]

    arXiv preprint arXiv:2409.00135 (2024)

    Zhang, H., Song, Y., et al.: Honeycomb: A flexible llm-based agent system for materials science. arXiv preprint arXiv:2409.00135 (2024)

    Pith/arXiv arXiv 2024

  14. [14]

    https://arxiv.org/abs/2402.06852

    Zhang, D., Liu, W., et al.: ChemLLM: A Chemical Large Language Model (2024). https://arxiv.org/abs/2402.06852

    Pith/arXiv arXiv 2024

  15. [15]

    https://arxiv.org/abs/2406.03949

    Zhang, K., Zeng, S., et al.: UltraMedical: Building Specialized Generalists in Biomedicine (2024). https://arxiv.org/abs/2406.03949

    Pith/arXiv arXiv 2024

  16. [16]

    Nature communications15(1), 1418 (2024)

    Dagdelen, J., Dunn, A.,et al.: Structured information extraction from scientific text with large language models. Nature communications15(1), 1418 (2024)

    2024

  17. [17]

    npj Computational Materials11(1), 194 (2025)

    Odobesku, R., Romanova, K.,et al.: Agent-based multimodal information extraction for nanomaterials. npj Computational Materials11(1), 194 (2025)

    2025

  18. [18]

    https://arxiv.org/abs/2508.20729

    Cheng, A., Zhang, L., He, G.: Re4: Scientific Computing Agent with Rewriting, Resolution, Review and Revision (2025). https://arxiv.org/abs/2508.20729

    arXiv 2025

  19. [19]

    https://arxiv.org/abs/2407.13168

    Tian, M., Gao, L., et al.: SciCode: A Research Coding Benchmark Curated by Scientists (2024). https://arxiv.org/abs/2407.13168

    Pith/arXiv arXiv 2024

  20. [20]

    https://arxiv.org/abs/2410

    Yang, Z., Liu, W., et al.: MOOSE-Chem: Large Language Models for Rediscover- ing Unseen Chemistry Scientific Hypotheses (2025). https://arxiv.org/abs/2410. 07076

    2025

  21. [21]

    https://arxiv.org/ abs/2404.07738

    Baek, J., Jauhar, S.K., et al.: ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models (2025). https://arxiv.org/ abs/2404.07738

    Pith/arXiv arXiv 2025

  22. [22]

    https://arxiv.org/abs/2504.07421

    Abaskohi, A., Ramesh, A.V., et al.: AgentAda: Skill-Adaptive Data Analytics for Tailored Insight Discovery (2025). https://arxiv.org/abs/2504.07421

    arXiv 2025

  23. [23]

    https://arxiv.org/abs/2505.13400

    Ghareeb, A.E., Chang, B., et al.: Robin: A multi-agent system for automating scientific discovery (2025). https://arxiv.org/abs/2505.13400

    Pith/arXiv arXiv 2025

  24. [24]

    Ruan, Y., Lu, C.,et al.: An automatic end-to-end chemical synthesis development platformpoweredbylargelanguagemodels.Naturecommunications15(1),10160 (2024)

    2024

  25. [25]

    arXiv preprint arXiv:2408.06292 (2024)

    Lu, C., Lu, C., et al.: The ai scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292 (2024)

    Pith/arXiv arXiv 2024

  26. [26]

    Proceedings of the National Academy of Sciences122(5), 2401238121 (2025) 18

    Musslick, S., Bartlett, L.K.,et al.: Automating the practice of science: Oppor- tunities, challenges, and implications. Proceedings of the National Academy of Sciences122(5), 2401238121 (2025) 18

    2025

  27. [27]

    Oxford University Press (2024)

    Ouzounis, C.A.: Biology’s transformation: from observation through experiment to computation. Oxford University Press (2024)

    2024

  28. [28]

    ACS Publications (2016)

    Kofke, D.A., Siepmann, J.I., et al.: Molecular modeling and simulation in JCED. ACS Publications (2016)

    2016

  29. [29]

    Progress in Biophysics and Molecular Biology117(1), 99– 106 (2015) https://doi.org/10.1016/j.pbiomolbio.2014.10.001

    Cooper, J., Vik, J.O., Waltemath, D.: A call for virtual experiments: Accelerating the scientific process. Progress in Biophysics and Molecular Biology117(1), 99– 106 (2015) https://doi.org/10.1016/j.pbiomolbio.2014.10.001

    work page doi:10.1016/j.pbiomolbio.2014.10.001 2015

  30. [30]

    arXiv preprint arXiv:2504.08621 (2025)

    Zhang, T., Liu, Z., et al.: Mooseagent: A llm based multi-agent framework for automating moose simulation. arXiv preprint arXiv:2504.08621 (2025)

    arXiv 2025

  31. [31]

    55–65 (2024)

    Sun, J., Li, A.,et al.: Chatmol copilot: An agent for molecular modeling and computationpoweredbyllms.In:Proceedingsofthe1stWorkshoponLanguage+ Molecules (L+ M 2024), pp. 55–65 (2024)

    2024

  32. [32]

    https://arxiv.org/abs/2508.07035

    Liu, J., Zhu, T., et al.: VASPilot: MCP-Facilitated Multi-Agent Intelligence for Autonomous VASP Simulations (2025). https://arxiv.org/abs/2508.07035

    Pith/arXiv arXiv 2025

  33. [33]

    Preprint / published online Aug 19, 2025 (2025)

    Han, X.-Q., Gao, Z.-F., et al.: PhysAgent: A Multi-Agent Approach to the Auto- mated Discovery of Physical Laws. Preprint / published online Aug 19, 2025 (2025). https://doi.org/10.32388/J2MXUW

    work page doi:10.32388/j2mxuw 2025

  34. [34]

    https://arxiv.org/abs/2311.17869

    Li, Y., Zhan, J.: SAIBench: A Structural Interpretation of AI for Science Through Benchmarks (2023). https://arxiv.org/abs/2311.17869

    Pith/arXiv arXiv 2023

  35. [35]

    https://arxiv.org/abs/ 2503.13503

    Qin, C., Chen, X., et al.: SciHorizon: Benchmarking AI-for-Science Readiness from Scientific Data to Large Language Models (2025). https://arxiv.org/abs/ 2503.13503

    Pith/arXiv arXiv 2025

  36. [36]

    Physical review B54(16), 11169 (1996)

    Kresse, G., Furthmüller, J.: Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Physical review B54(16), 11169 (1996)

    1996

  37. [37]

    https:// arxiv.org/abs/2412.19437

    DeepSeek-AI, Liu, A., et al.: DeepSeek-V3 Technical Report (2025). https:// arxiv.org/abs/2412.19437

    Pith/arXiv arXiv 2025

  38. [38]

    https://arxiv.org/abs/ 2410.21276

    OpenAI, Hurst, .A., et al.: GPT-4o System Card (2024). https://arxiv.org/abs/ 2410.21276

    Pith/arXiv arXiv 2024

  39. [39]

    https://arxiv.org/abs/ 2505.09388

    Yang, A., Li, A., et al.: Qwen3 Technical Report (2025). https://arxiv.org/abs/ 2505.09388

    Pith/arXiv arXiv 2025

  40. [40]

    System card, Ope- nAI (April 2025)

    OpenAI: Openai o3 and o4-mini system card. System card, Ope- nAI (April 2025). Version: April 16, 2025. https://cdn.openai.com/pdf/ 2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf

    2025

  41. [41]

    https://arxiv.org/abs/2507.06261

    Comanici, G., Bieber, E., et al.: Gemini 2.5: Pushing the Frontier with Advanced 19 Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabil- ities (2025). https://arxiv.org/abs/2507.06261

    Pith/arXiv arXiv 2025

  42. [42]

    https://www.anthropic.com/news/ claude-3-7-sonnet

    Anthropic: Claude 3.7 Sonnet. https://www.anthropic.com/news/ claude-3-7-sonnet. Accessed: 2025-11-05 (2025)

    2025

  43. [43]

    Springer

    Springer Nature: SpringerMaterials: The Landolt-Börnstein Database. Springer. Accessed: 2024-11-06 (2024)

    2024

  44. [44]

    Journal of the American Chemical Society121(33), 7644–7652 (1999)

    Bleakley, K., Hu, P.: A density functional theory study of the interaction between co and o on a pt surface: Co/pt (111), o/pt (111), and co/o/pt (111). Journal of the American Chemical Society121(33), 7644–7652 (1999)

    1999

  45. [45]

    Journal of the American Chemical Society121(34) (1999)

    Zhang, C., Hu, P., Alavi, A.: A general method for co oxidation on close-packed transition metal surfaces. Journal of the American Chemical Society121(34) (1999)

    1999

  46. [46]

    Physical review letters100(13), 136406 (2008)

    Perdew, J.P., Ruzsinszky, A.,et al.: Restoring the density-gradient expansion for exchange in solids and surfaces. Physical review letters100(13), 136406 (2008)

    2008

  47. [47]

    The Journal of chemical physics132(15) (2010)

    Grimme,S.,Antony,J.,etal.:Aconsistentandaccurateabinitioparametrization of density functional dispersion correction (dft-d) for the 94 elements h-pu. The Journal of chemical physics132(15) (2010)

    2010

  48. [48]

    ACS Catalysis 13(1), 433–444 (2022)

    Wang, J., Xu, H.,et al.: Rational design of pdag catalysts for acetylene selective hydrogenation via structural descriptor-based screening strategy. ACS Catalysis 13(1), 433–444 (2022)

    2022

  49. [49]

    ACS Catalysis6(10), 6730–6738 (2016)

    Jørgensen, M., Gronbeck, H.: First-principles microkinetic modeling of methane oxidation over pd (100) and pd (111). ACS Catalysis6(10), 6730–6738 (2016)

    2016

  50. [50]

    The Journal of chemical physics113(22), 9901–9904 (2000) 20

    Henkelman, G., Uberuaga, B.P., Jónsson, H.: A climbing image nudged elastic band method for finding saddle points and minimum energy paths. The Journal of chemical physics113(22), 9901–9904 (2000) 20

    2000