Science Done on a Machine by a Machine: AI Agents in Computational Chemistry

Pavlo O. Dral; Hassan Nawaz; Arif Ullah

physics.chem-phAug 19, 2026

Science Done on a Machine by a Machine: AI Agents in Computational Chemistry

Authors: Pavlo O. Dral, Hassan Nawaz, Arif Ullah

Organizations: State Key Laboratory of Physical Chemistry of Solid Surfaces, College of Chemistry and Chemical Engineering, and Fujian Provincial Key Laboratory of Theoretical and Computational Chemistry, Xiamen University, Xiamen, Fujian 361005, China · Institute of Physics, Faculty of Physics, Astronomy, and Informatics, Nicolaus Copernicus University in Toruń, ul. Grudziadzka 5, 87-100 Toruń, Poland · Institute of Advanced Studies, Nicolaus Copernicus University in Toruń, ul. Wileńska 4, 87-100 Toruń, Poland · Aitomistic, Shenzhen 518000, China · School of Physics, Anhui University, Hefei, 230601, Anhui, China

Abstract

We are witnessing an explosion of agentic systems for computational chemistry: from four in 2024 to seventeen in 2025 and over sixty now, surveyed here. What is delegated to these systems is shifting from single calculations to whole in silico experiments and even manuscript writing. The ultimate destination is a fully autonomous AI scientist, where the entirety of computational chemistry is performed on a machine by a machine, without human supervision. Our survey shows that these systems are turning into vetted chemistry skills on general-purpose coding agents, and that they must be evaluated not only on their final answers but also on whether their calculations actually support these answers. The role of the human computational chemist is shifting from performing calculations to directing and supervising them, and the field should invest in the judgement that makes the supervision reliable: we propose a reporting standard, an evaluation reproducing published studies, a controlled comparison with general-purpose coding agents, and how to teach.

Figures & tables

Figure 1: From a chemistry question to a scientific result. The opening dialogue illustrates a researcher delegating computational work to an agent. The example is the Diels–Alder cycloaddition of 1,3-butadiene and ethene to cyclohexene. The rows distinguish the largest unit of work delegated: a single calculation; an in silico experiment that includes planning, calculation and analysis; or a study extending from a question to a written manuscript. Counts correspond to the systems surveyed in this Perspective which demonstrate each delegation level. Campaign and agenda are prospective larger scopes that no surveyed system demonstrates; 1 system has unclear scope. Chemical structures and energy diagrams are schematic and do not represent numerical results.

Figure 2: Growth of agents for computational chemistry. Each point marks one system at its first public appearance and is named beside the curve. Annual labels give the number of systems and their rate; 2026 runs only to the cutoff of 8 August 2026.

Figure 3: Reported computational chemistry capabilities. The tasks are ordered clockwise from the top in descending order of the number of systems reporting each task; sector area is not proportional to the number of systems. A system can report several tasks; percentages use the 61 surveyed systems as the denominator. All chemistry drawings are schematic. Counts use the survey cutoff of 8 August 2026.

Figure 4: How agents run chemistry software. (a) Results inform subsequent actions in an agent loop. (b) Four coexisting routes provide access to chemistry operations: predefined callable functions, tools served through the Model Context Protocol, code written and executed by an agent, and retrieved procedural skills that guide tool use. (c) The orchestration layer is distinct from these routes; counts classify surveyed systems by framework or SDK, general-purpose coding agent, custom loop or unspecified orchestration. Named frameworks and coding agents are examples from the survey, not exhaustive lists. A custom loop names the orchestration implementation, not code written by the agent.

Figure 5: Evaluation of agentic scientific work and its cost. (a) The proposed evaluation asks agents to reproduce published computational studies from the question to the report. Agents receive the question and defined allowed inputs; evaluators use the published protocol and results as reference evidence. A specialized system and a general-purpose coding agent without chemistry skills receive matched tasks, base models, software access and budgets, with repeated runs. The cost of each run is token cost plus atomistic simulation cost, including failed attempts and retries. (b) Evaluation should check whether the procedure is valid, the computation is executed correctly, and the conclusions drawn are correct. (c) Release records should include licence, runnable code and tests, model version, run traces and access; reports should also disclose scope and human interventions. (d) Teaching should develop the corresponding abilities to justify method choices, inspect raw outputs, and defend conclusions.

Figure A1: Architectures and delegation over time. (a) Orchestration of the surveyed systems, grouped by year of first appearance. (b) Number of systems at each delegation level, grouped by year. (c) Percentage of systems at each delegation level within each year. The survey cutoff is 8 August 2026.

#	Framework	Date	Scientific tasks	Architecture	Delegation	Ref.
1	ProtAgents	2024-01-27	struct. gen., vib.	fixed $\cdot$ AutoGen	calculation	Ghafarollahi and Buehler (2024)
2	ChemReasoner	2024-02-15	screening, struct. gen.	fixed $\cdot$ own harness	calculation	Sprueill et al. (2024)
3	AutoSolvateWeb	2024-03-05	struct. gen., MD, elec. struct.	fixed $\cdot$ Dialogflow	calculation	Gadde et al. (2025)
4	AtomAgents	2024-07-13	TS, struct. gen.	fixed $\cdot$ AutoGen	experiment	Ghafarollahi and Buehler (2025)
5	MDCrow	2025-02-13	MD	fixed $\cdot$ LangChain	calculation	Campbell et al. (2026)
6	Paramus	2025-02-28	elec. struct., vib., TS, excited, MD, periodic DFT, free energy, struct. gen., screening, cheminf.	fixed, MCP, skills $\cdot$ own harness	n/a	Thorsten Gressling and Paramus Transform LLC (2025)

Table 1: The 61 agentic systems for atomistic simulation surveyed here, ordered by date of first public appearance and surveyed to a cutoff of 8 August 2026. Scientific tasks are those the system performs on its authors’ own account, read paper by paper: an implemented capability is recorded even where the worked example is missing, but a task the authors place in future work, or one performed by a program they merely cite, is not. The Delegation column is what the paper demonstrates in its body; aspirations and future work are deliberately not tabulated. That column uses the five-position delegation scale defined above; a system that occupies more than one position at once carries a compound entry and is counted at the larger. Architecture names how the agent reaches its capabilities and what runs it: predefined tool functions, the Model Context Protocol, retrieved skills or agent-written code, then, after the separator, the software framework the agent itself runs on. The tool tokens are predefined functions, the Model Context Protocol, retrieved procedural skills, agent-written code, or a combination. A dash marks a system whose agents have no callable tool functions. A question mark marks a value the paper does not record, and n/a a value for which no source exists, the system having no paper at all.

Explore similar work

CompMat-Bench: Benchmarking AI Agents for Computational Materials Science
Sep 30, 2026Chenmu Zhang, Levi Felix, Jun-Jie Zhang +6Materials Science Artificial Intelligence Agents
Department of Materials Science and NanoEngineering Rice University, Houston, TX 77005, USA
El Agente Quntur: A research collaborator agent for quantum chemistry
Date pendingJuan B. Pérez-Sánchez, Yunheng Zou, Jorge A. Campos-Gonzalez-Angulo +12Quantum Chemistry Chemistry
Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada +6
Can Coding Agents Reproduce Findings in Computational Materials Science?
May 1, 2026Ziyang Huang, Yi Cao, Ali K. Shargh +15Coding Agents Reproducibility
Johns Hopkins University

Sep 30, 2026, cs.AIJ/K move · Enter open · S save

CompMat-Bench: Benchmarking AI Agents for Computational Materials Science

Chenmu Zhang, Levi Felix, Jun-Jie Zhang, Xingfu Li+5

Department of Materials Science and NanoEngineering Rice University, Houston, TX 77005, USA

Materials Science Artificial Intelligence Agents Scientific Discovery Methodology

Open paper AskPDF arXiv

Evaluating AI agents on scientific research tasks is constrained by the time and resources required for the underlying experiments or calculations. In computational materials research, repeating the same expensive simulations across agents and trials can make evaluation impractical. We introduce CompMat-Bench, a benchmark of 94 tasks derived from recently published computational materials studies, each asking agents to complete a step toward achieving the study's scientific goal. We reproduce the research steps in advance and assess agents on preparing inputs and analyzing outputs for expensive simulations, so expensive simulations can be avoided during evaluation. The reproduced inputs and results serve as ground truth for grading agents with fixed rules, without an LLM judge. The benchmark supports four evaluation conditions: single tasks and workflows composed of related tasks, each with full or reduced methodological guidance. With full guidance on single tasks, agents based on three LLMs demonstrate the ability to complete individual materials research steps, with pass rates of 66.0-90.4% across 94 tasks. Both longer workflows and reduced guidance can limit agent performance, but in different ways for different agents: they lower the pass rates of the weaker agents, whereas the strongest agent falls only when a long workflow is combined with reduced guidance. Failure analysis attributes most failures to scientific errors rather than to errors in software usage. CompMat-Bench provides a basis for comparing agents on the steps of real materials research and for analyzing agent failure modes.

Abstract

Explore similar work

CompMat-Bench: Benchmarking AI Agents for Computational Materials Science

El Agente Quntur: A research collaborator agent for quantum chemistry

Can Coding Agents Reproduce Findings in Computational Materials Science?