CoChat
Sign Up
Cited by 9Machine Learning in Materials ScienceAdvanced Chemical Physics StudiesArtificial Intelligence Applications

Multi-agentic AI framework for end-to-end atomistic simulations

Open access
Vriza, Aikaterini, Kornu, Uma, Koneru, Aditya, Chan, Henry, Sankaranarayanan, Subramanian K. R. S.·2026·Digital Discovery
DOI 10.1039/d5dd00435g
Open this paper in CoChat — free accountOpen PDF

Abstract

Autonomous multi-agent AI system coordinates specialized agents to perform complex materials property calculations from natural language prompts.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper introduces a multi-agent LLM framework that autonomously performs end-to-end atomistic simulations, from structure generation and interatomic-potential discovery to LAMMPS execution on HPC and post-simulation analysis. It demonstrates static and dynamic property predictions across elemental and alloy systems, with results reported as comparable to human experts.

Key findings

  • The framework automates the full MD pipeline using specialized agents and tools including Atomsk, LAMMPS, Phonopy, OVITO, Playwright web scraping, ASE, and custom HPC scripts, coordinated through shared memory and human-in-the-loop control.
  • For gold, the agent produced an fcc lattice constant of 4.078 Å and cohesive energy of −3.789 eV/atom, consistent with human expert predictions; reported lattice-constant examples include Ti at 2.866 Å, Ni-Cu at 3.58177 Å, Si at 5.43094 Å, and Au-Cu at 3.8465 Å by agent vs 3.8388 Å by human.
  • Elastic constants for FCC gold, using a 5×5×5 supercell with 500 atoms, EAM potential Au_u3.eam, and strain δ = 0.001, yielded bulk modulus 166.89 GPa, primary shear modulus 44.73 GPa, and Poisson ratio 0.46; phonon dispersions for Au and NiCu were reported as identical between agent and human.
  • The melting-point workflow, using the interface/coexistence method and vision agents, identified an AuCu melting point at approximately 1604 K from a sharp heat-capacity peak; GPT o3 vision was most accurate, while GPT-4o and Qwen3-VL-8B-Thinking were slightly lower but reliable.

Methods

  • Multi-agent architecture built on AutoGen (AG2), with agents for structure creation, LAMMPS input generation, input review, potential discovery, web scraping, elastic constants, melting point, Phonopy, results analysis with vision, HPC execution, and administration.
  • Static workflows included lattice/cohesive energy via energy minimization over lattice constants, elastic constants via finite-deformation templates, and phonon dispersion via Phonopy finite displacements; the dynamic workflow used a solid-liquid interface/coexistence approach with NVT/NPT heating and OVITO frame inspection.
  • Validation compared agent outputs against human experts using the same initial structures and potential files, with %error = abs(A − H)/H × 100 and human results treated as ground truth; systems included Au, Fe, Ti, Si, Ni-Cu, and Au-Cu, run on the Carbon HPC cluster with LAMMPS version 29 Aug 2024 Update 3.

Limitations

  • Stated: dependency management and environment reproducibility can be brittle across platforms or HPC systems; the authors currently use curated environments and call for containerization and environment self-checks.
  • Stated: trustworthiness and explainability of LLM-based reasoning remain open challenges; future work should add transparent logging, agent-level confidence scores, and possibly symbolic reasoning, and complex tasks such as phonon dispersion, thermal conductivity, or defect energetics may need adaptive sampling, uncertainty quantification, and self-correction.
  • Stated: the choice of interatomic potential dominates computed properties and requires human scientific oversight; the agent should not be held responsible for errors from the underlying potential. Evident: the provided text contains a placeholder “average errors below x%” and many quantitative comparisons are relegated to supplementary tables, so the main text alone does not fully report all validation numbers.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1039/d5dd00435g

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 28 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The phonon dispersion for cubic gold was successfully computed and visualized along the Γ → X → W → K → Γ → L→ U path.

    “Finally, the phonon dispersion was successfully computed and visualized in a band.pdffile along the Γ → X → W → K → Γ → L→ U path for the cubic gold system, completing the workflow (Figure 4b).”
    Quote in the paperNot a substantial claim
  • Findingnone given

    The system achieved average errors below x% compared to the results of human LAMMPS simulation experts.

    “The system achieved average errors below x% compared to the results of human LAMMPS simulation experts.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The human output is treated as the correct value (general truth).
  • Findingown data

    The generated elastic constant matrices confirm that the agentic system correctly calculated the stiffness tensor for gold in GPa.

    “The generated elastic constant matrices are shown in Figure 3b and Supplementary Fig. 6 confirms that the agentic system has correctly calculated the stiffness tensor for gold in Gpa.”
    Quote in the paperSubstantial claim
  • Findingown data

    The agentic workflow reproduces the expected thermodynamic signature of melting.

    “Our agentic workflow not only reproduces the expected thermodynamic signature of melting but also highlights how multi-agent orchestration-spanning structure preparation, simulation setup, execution, and analysis-can seamlessly integrate with physics-based MD to deliver accurate, reproducible predictions of key material properties.”
    Quote in the paperSubstantial claim
  • Findingown data

    The GPT o3 model consistently yielded the most accurate image-based reasoning results.

    “We used the GPT o3 model as the primary vision agent, as it consistently yielded the most accurate image-based reasoning results.”
    Quote in the paperSubstantial claim
  • May combine two claims. 21
  • Classified as probably not a substantial claim. 25
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingnone given

    The results produced by the agents show strong agreement with those obtained by a human expert, highlighting the reliability of the agentic approach.

    “The results produced by the agents show strong agreement with those obtained by a human expert, highlighting the reliability of the agentic approach.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The human output is treated as the correct value (general truth).

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    In the gold phonon dispersion workflow, the LAMMPS input script initially failed due to unrecognized or invalid commands, and these errors were progressively corrected, leading to successful relaxation and generation of a relaxed structure.

    “A LAMMPS input script was created to relax the gold structure, but the simulation initially failed due to unrecognized or invalid commands (after read pause 0, then after minimize).These errors were progressively corrected, leading to successful relaxation and generation of a relaxed structure.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The heat capacity curve exhibits a sharp peak at approximately 1604 K, which corresponds to the alloy's melting point.

    “the heat capacity curve is computed from fluctuations in enthalpy, and it exhibits a sharp peak at approximately 1604 K, which corresponds to the alloy's melting point.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    GPT-4o and Qwen3-VL-8B-Thinking showed slightly lower accuracy than GPT o3 for complex atomic structures.

    “While both GPT-4o and Qwen showed slightly lower accuracy for complex atomic structures, they remain reliable and consistent, making them a promising and accessible alternative for future multimodal agentic workflows.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The system is capable of autonomous decision-making and error recovery.

    “although the system is capable of autonomous decision-making and error recovery, the trustworthiness and explainability of some AI-driven actions, particularly LLM-based reasoning, remain open challenges.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The elastic constants matrix was computed by the multi-agent system after setting the variable up equal to 0.001 and 0.01.

    “Elastic constants matrix computed by the multi-agent system after setting the variable up equal to 0.001 and 0.01.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The key component of the process is the integration of the results analyzer agent with two vision models assigned different observatory tasks and based on the o3 vision model.

    “The key component of the process is the integration of the results analyzer agent with t wo vision models which are assigned with different observatory tasks and are based on o3 vision model.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    By combining automation, reproducibility, and human-in-the-loop control, the framework lowers the barrier to widespread adoption of scalable, AI-driven discovery tools in materials science.

    “By combining automation, reproducibility, and human-in-the-loop control, our framework lowers the barrier to the widespread adoption of scalable, AI-driven discovery tools in materials science.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    Unlike prior works largely focused on single-agent systems, the framework distributes tasks across agents that can reason, self-correct, and collaborate dynamically via agent-to-agent communication through a shared memory and execution interface.

    “Unlike prior works which are largely focused on single-agent systems, our framework distributes tasks across agents that can reason, self-correct, and collaborate dynamically via agent-to-agent communication through a shared memory and execution interface.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    These agentic workflows aim to leverage intelligent agents to orchestrate complex computational tasks, streamline parameter exploration, and autonomously extract insights, thereby accelerating scientific discovery and reducing human intervention in labor-intensive simulation pipelines.

    “These workflows aim to leverage intelligent agents, often powered by machine learning or AI, to orchestrate complex computational tasks, streamline parameter exploration, and autonomously extract insights, thereby accelerating scientific discovery and reducing human intervention in traditionally labor-intensive simulation pipelines.21,22”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [21]

      Not checked in the free run.

    • cites [22]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Recent advances in large language models and agent-based AI systems offer promising routes toward automating complex scientific workflows that lower the barrier for scientific usage by non-expert users.

    “Recent advances in large language models (LLMs) and agent-based AI systems offer promising routes toward automating complex scientific workflows that lower the barrier for scientific usage by non-expert users.[12][13][14][15][16]”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [12]

      Not checked in the free run.

    • cites [13] source

      Not checked in the free run.

    • cites [14]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    There has been a significant surge in development of agentic pipelines designed to automate experimental workflows, simulation tools, and subsequent data analysis.

    “There has been a significant surge in the development of agentic pipelines designed to automate experimental workflows, simulation tools and the subsequent data analysis processes.[17][18][19][20]”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [17] source

      Not checked in the free run.

    • cites [18]

      Not checked in the free run.

    • cites [19]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The fragmented tool collection limits accessibility for non-specialists and poses a significant challenge for integrating simulations into autonomous or data-driven materials design pipelines.

    “It also limits accessibility for non-specialists and poses a significant challenge for integrating simulations into autonomous or data-driven materials design pipelines.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Atomistic simulation workflows remain fragmented, manual, and highly dependent on domain expertise.

  • May combine two claims.

    Interpretationreasoning

    The system can respond autonomously to simulation errors, adapt tool selection based on material context, and incorporate human guidance for learning and explainability.

    “This enables our system to respond autonomously to simulation errors, adapt tool selection based on material context, and incorporate human guidance for learning and explainability.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Variability in computed properties across different potential parameterizations stems from intrinsic limitations and transferability of the potential itself, rather than from the agent's operation.

    “Such variability stems from the intrinsic limitations and transferability of the potential itself, rather than from the agent's operation.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The choice of interatomic potential exerts the largest influence on all computed physical properties, often leading to significant variability in lattice constants, cohesive energies, and elastic moduli across different parameterizations.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    AtomAgents applied a multimodal GPT-4 based multi-agent system, where each agent acts like planner, scientist, and engineer to automate alloy design coordinating MD workflows.

    “AtomAgents, on the other hand applied multimodal GPT-4 based multi-agent system, where each agent is designed to act like planner, scientist and engineer to automate alloy design coordinating MD workflows.26”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26]

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    Each agent is provided with a specialized set of domain-specific tools and a clearly defined role.

    “Each agent is provided with a specialized set of domain-specific tools and a clearly defined role.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    For consistent comparison and due to the deterministic nature of LAMMPS simulations, the human expert evaluation used the same initial structure and same potential files as those generated by the Structure Agent and the Potential Agent.

    “For consistent comparison and due to the deterministic nature of the LAMMPS simulations, the human expert evaluation used the same initial structure and same potential files as those generated by the Structure Agent and the Potential Agent respectively.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: LAMMPS simulations are deterministic.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DynaMate offers a modular multi-agent pipeline for scientific workflows built on a flexible Python template that leverages LLM agents to automate molecular simulations across diverse domains mainly focused on solvents and metal-organic frameworks.

    “Similarly, DynaMate offers modular multi-agent pipeline for scientific workflows built on a flexible Python template that leverages LLM agents to automate molecular simulations across diverse domains mainly focused on solvents and metal-organic frameworks.25”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Existing agentic systems often rely on static templates, lack flexibility in input generation, and do not fully support classical MD workflows involving multiple diverse tasks ranging from structure generation, interatomic potential selection by searching the web, visual inspection of structures, or trajectory analysis.

    “While each of these agentic systems contribute valuable ideas, they often rely on static templates, lack flexibility in input generation, and do not fully support classical MD workflows involving multiple diverse tasks ranging from structure generation, interatomic potential selection by searching the web, visual inspection of the structures, or trajectory analysis.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [24]

      Not checked in the free run.

    • cites [25]

      Not checked in the free run.

    • cites [26]

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    Upon detection, the reviewer agent collaborates with the relevant task agent to regenerate or fix input scripts, and this autocorrection loop is repeated until the simulation is executed without any errors.

    “Upon detection, the reviewer agent collaborates with the relevant task agent (e.g., potential agent, LAMMPS input agent) to regenerate or fix the input scripts.This autocorrection loop is repeated until the simulation is executed without any errors.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The complementary static and dynamic modes of operation allowed benchmarking the system across equilibrium property calculations and time-dependent processes, ensuring robustness and generality of the pipelines.

    “Together, these complementary modes of operation allowed us to benchmark the system across both equilibrium property calculations and time-dependent processes, ensuring robustness and generality of the pipelines.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MDCrow introduced a single-agent system built on LangChain, leveraging GPT-4o and Llama3-405B to automate biomolecular MD workflows using tools like OpenMM, MDTraj, and PDBFixer.

    “For example MDCrow introduced a single agent system built on LangChain, 23 leveraging GPT-4o and Llama3-405B to automate biomolecular MD workflows using tools like OpenMM, MDTraj, and PDBFixer.24”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [23] source

      Not checked in the free run.

    • cites [24]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The agentic approach was tested for the exemplary systems with all results and comparison with human-derived plots shown in Supplementary Fig. 8-13.

    “The agenting approach was tested for the exemplary systems with all the results and comparison with the human derived plots shown in the Supplementary Fig. 8910111213.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The human output is treated as the correct value (general truth).

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LAMMPS is a widely adopted open-source molecular dynamics simulation package used for modeling and analyzing materials at the atomic scale.

    “We utilize LAMMPS, a widely adopted open-source molecular dynamics (MD) simulation package, for modeling and analyzing materials at the atomic scale.8”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [8]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Each expert agent operates within a defined role, equipped with domain-specific functions and a shared memory context for coordination.

    “Each expert agent operates within a defined role, equipped with domainspecific functions and a shared memory context for coordination.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DREAMS framework was used for DFT calculations on a materials benchmark, demonstrating comparable performance to human DFT experts.

    “More recently, DREAMS framework was used for DFT calculations on a materials benchmark, demonstrating comparable performance to human DFT experts.27”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    A hierarchical framework integrated with GPT-4 autonomously plans, executes, and debugs DFT-based computations using tools like ORCA and xTB.

    “demonstrated a hierarchical framework for complex computational and quantum chemistry workflows by integrating with GPT-4 to autonomously plan, execute, and debug DFT-based computations using tools like ORCA and xTB.26”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [26]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Such architectures reduce the contextual load on any single agent and enable complex workflows to be decomposed into manageable subtasks, mitigating context-window limitations observed in single-agent designs.

    “Such architectures reduce the contextual load on any single agent and enable complex workflows to be decomposed into manageable subtasks, mitigating context-window limitations observed in single-agent designs.26”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [26]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The heating simulation starting from room temperature and ramping beyond the expected melting point marks the upper bound of the melting point for the given interatomic potential.

    “This marks the upper bound of the melting point for the given interatomic potential.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The LAMMPS input reviewer agent monitors output logs from each simulation and parses for known failure signatures, including invalid commands, numerical instabilities, and missing potential files or incorrect atom types.

    “Furthermore, the LAMMPS input reviewer agent monitors the output logs from each simulation and parses for known failure signatures, including invalid commands (e.g., misnamed keywords or deprecated syntax), numerical instabilities (e.g., atom loss, box deformation) and missing potential files or incorrect atom types.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Multi-agent orchestration ensures reproducibility, modularity, and scalability across different workflows.

    “Our results highlight that multi-agent orchestration is not only capable of reproducing core methodologies in LAMMPS but also ensures reproducibility, modularity, and scalability across different workflows.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The multi-agent framework demonstrates substantial improvements in automation, modularity, and reproducibility for atomistic simulations.

    “While our multi-agent framework demonstrates substantial improvements in automation, modularity, and reproducibility for atomistic simulations, several limitations and opportunities for enhancement remain.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation

    Melting properties play a key role for the in-depth understanding of a material as they affect the synthesis method, processing and performance in various areas of application.

    “Melting properties play a key role for the in-depth understanding of a material as they affect the synthesis method, processing and performance in various areas of application.41”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [41]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The approach is intentionally designed to be framework-agnostic.

    “our approach is intentionally designed to be frameworkagnostic.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Premisenone given

    Visual inspection of structures during melting is a critical part for understanding the process, and human experts have an understanding of how a melted structure should look like.

    “A critical part for the understanding of the process is the visual inspection of the structures during the melting, which is performed by human experts that have an understanding on how a melted structure should look like.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The Au lattice constant and cohesive energy results are consistent with predictions by expert human users.

    “The results yield a stable fcc Au structure with lattice constant 4.078 Å and cohesive energy -3.789 eV/atom which is consistent with predictions by expert human users.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Human expert results are treated as the ground truth for evaluating the agent's accuracy.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    The predefined functions, system messages, and agent roles can be readily transferred or reconfigured within other pipelines, making the framework portable and easy to integrate with emerging ecosystems.

    “The predefined functions, system messages, and agent roles that we developed can be readily transferred or reconfigured within other pipelines, making the framework portable and easy to integrate with emerging ecosystems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The framework is easily customizable and can potentially serve a wide range of simulation tasks for soft materials.

    “Although we have demonstrated our workflow for inorganic alloys and materials, our framework is easily customizable and can potentially serve a wide range of simulation tasks for soft materials as well.”
    Quote in the paperNot a substantial claim
39 of 39

  • The paper relies on this without support.

    Atomistic simulation workflows remain fragmented, manual, and highly dependent on domain expertise.

    “However, despite their wide adoption and availability, these simulation workflows remain fragmented, manual, and highly dependent on domain expertise.”

  • The paper relies on this without support.

    The choice of interatomic potential exerts the largest influence on all computed physical properties, often leading to significant variability in lattice constants, cohesive energies, and elastic moduli across different parameterizations.

    “Note that the choice of interatomic potential exerts the largest influence on all computed physical properties, often leading to significant variability in lattice constants, cohesive energies, and elastic moduli across different parameterizations.”

  • The authors state this as an assumption.

    The human output is treated as the correct value (general truth).

    “The human output is treated as the correct value (general truth).”

  • The paper relies on this without support.

    LAMMPS simulations are deterministic.

    “due to the deterministic nature of the LAMMPS simulations”

  • The authors state this as an assumption.

    Human expert results are treated as the ground truth for evaluating the agent's accuracy.

    “which we treat as the ground truth.”

  • The paper relies on this without support.

    Human experts have an understanding of how a melted structure should look like.

    “which is performed by human experts that have an understanding on how a melted structure should look like.”
6 of 6

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Cites

Flagged first, then by citations

Research works printed but not linked: 12

  1. J. D.Evans , in Computer Simulation of Porous Materials: Current Approaches and Future Opportunities , ed. K. Jelfs , The Royal Society of Chemistry , 2021
  2. M.Griebel , S.Knapek and G.Zumbusch , Numerical Simulation in Molecular Dynamics: Numerics, Algorithms, Parallelization, Applications , Springer Publishing Company Incorporated , 1st edn, 2007
  3. S.Yang , S.Batzner , R.Gao , M.Aykol , A.Gaunt , B.Mcmorrow , D.Rezende , D.Schuurmans , I.Mordatch , E.Dogus and C. G.Deepmind , Generative Hierarchical Materials Search
  4. J.Gottweis , W.-H.Weng , A.Daryin , T.Tu , A.Palepu , P.Sirkovic , A.Myaskovsky , F.Weissenberger , K.Rong , R.Tanno , K.Saab , D.Popovici , J.Blum , F.Zhang , K.Chou , A.Hassidim , B.Gokturk , A.Vahdat , P.Kohli , Y.Matias , A.Carroll , K.Kulkarni , N.Tomasev , Y.Guan , V.Dhillon , E. D.Vaishnav , B.Lee , T. R. D.Costa , J. R.Penadés , G.Peltz , Y.Xu , A.Pawlosky , A.Karthikesalingam and V.Natarajan , Towards an AI Co-Scientist , 2025
  5. T.Schick , J.Dwivedi-Yu , R.Dessì , R.Raileanu , M.Lomeli , E.Hambro , L.Zettlemoyer , N.Cancedda and T.Scialom , Toolformer: Language Models Can Teach Themselves to Use Tools , NIPS '23: Proceedings of the 37th International Conference on Neural Information Processing Systems , 2023 , pp. 68539–68551
  6. M.Muhoberac , A.Parikh , N.Vakharia , S.Virani , A.Radujevic , S.Wood , M.Verma , D.Metaxotos , J.Soundararajan , T.Masquelin , A. G.Godfre , S.Gardner , D.Rudnicki , S.Michael and G.Chopra , State and Memory is All You Need for Robust and Reliable AI Agents , 2024
  7. A.Vriza , M. H.Prince , T.Zhou , H.Chan and M. J.Cherukara , Operating advanced scientific instruments with AI agents that learn on the job , 2025
  8. Q.Campbell , S.Cox , J.Medina , B.Watterson and A. D.White , MDCrow: Automating Molecular Dynamics Workflows with Large Language Models , 2025
  9. Y.Zou , A. H.Cheng , A.Aldossary , J.Bai , S. X.Leong , J. A.Campos-Gonzalez-Angulo , C.Choi , C. T.Ser , G.Tom , A.Wang , Z.Zhang , I.Yakavets , H.Hao , C.Crebolder , V.Bernales and A.Aspuru-Guzik , El Agente: An Autonomous Agent for Quantum Chemistry , 2025
  10. B.Liu , X.Li , J.Zhang , J.Wang , T.He , S.Hong , H.Liu , S.Zhang , K.Song , K.Zhu , Y.Cheng , S.Wang , X.Wang , Y.Luo , H.Jin , P.Zhang , O.Liu , J.Chen , H.Zhang , Z.Yu , H.Shi , B.Li , D.Wu , F.Teng , X.Jia , J.Xu , J.Xiang , Y.Lin , T.Liu , T.Liu , Y.Su , H.Sun , G.Berseth , J.Nie , I.Foster , L.Ward , Q.Wu , Y.Gu , M.Zhuge , X.Liang , X.Tang , H.Wang , J.You , C.Wang , J.Pei , Q.Yang , X.Qi and C.Wu , Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems , 2025
  11. Q.Wu , G.Bansal , J.Zhang , Y.Wu , B.Li , E.Zhu , L.Jiang , X.Zhang , S.Zhang , A.Awadallah , R. W.White , D.Burger , C.Wang , AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation , in, COLM 2024 , 2024
  12. W.Chen , Y.Su , J.Zuo , C.Yang , C.Yuan , C.-M.Chan , H.Yu , Y.Lu , Y.-H.Hung , C.Qian , Y.Qin , X.Cong , R.Xie , Z.Liu , M.Sun and J.Zhou , AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors , in, ICLR , 2024

Cited by

Sorted by citations

Related

Authors
Vriza, A., Kornu, U., Koneru, A., Chan, H., & Sankaranarayanan, S. K. R. S.
Affiliations
  • Argonne National Laboratory (US)
  • University of Illinois Chicago (US)
ORCID
Venue
Digital Discovery
Publisher
Royal Society of Chemistry (RSC)
Published
2026
Volume
5
Issue
1
Pages
440-452
ISSN
2635-098X (ISSN-L)
DOI
10.1039/d5dd00435g
OpenAlex
W7111352005
Semantic Scholar
6ee1affc67202debc0f1cd9f47a3af757f10e303
Cited by
8
References
28
Citations by year
2026: 8
FWCI
1.70
Type
journal article
Language
en
Open access
Licence
Locations
Digital Discovery (publishedVersion, cc-by)
Funders
  • U.S. Department of Energy (DE‐ AC02‐06CH11357, -AC02-05CH11231, AC02‐06CH11357, 05CH11231, DE AC02 05CH11231, 34532, 06CH11357, DE-AC02-)
  • National Energy Research Scientific Computing Center (DE–AC02–05CH11231, 05CH11231, AC02-05CH11231)
  • Office of Science (AC02‐06CH11357, DEAC02–05CH11231, AC02-05CH11231, DE-AC02, 34532, 06CH11357, DE-AC02-06CH11357)
  • Basic Energy Sciences (DE-AC02, DE-AC02-05CH11231., DE-AC- 02-06CH11357, AC02‐06CH11357, AC02–05CH11231, 06CH11357, 34532)
  • Argonne National Laboratory (DE- AC02-06-CH11357, DE-AC02, AC02–06CH11357, DE-AC02-05CH11231, 06CH11357)
  • Lawrence Berkeley National Laboratory (AC02‐05CH11231, DE-AC02-05CH11231, DE-AC02-06CH11357, 05CH11231)
Topics
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Physics and Astronomy › Atomic and Molecular Physics, and Optics › Advanced Chemical Physics Studies
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
Keywords
multi-agent AI, atomistic simulation, natural language prompts, autonomous agents, end-to-end simulation, materials informatics

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review