Researchers at Argonne National Laboratory have developed an open-source artificial intelligence framework that allows scientists to design and run computational chemistry workflows using natural-language instructions.
ChemGraph connects large language models with established scientific software and simulation methods. A researcher can describe a problem in plain language, and the system translates the request into a sequence of tasks, including generating molecular structures, selecting simulation methods, preparing inputs, running calculations, and summarizing results.
The framework could make atomistic simulations more accessible to researchers who lack extensive experience with specialized computational chemistry tools. Potential applications include the development of catalysts, energy-storage materials, pharmaceuticals, and materials for critical-mineral and combustion research. The research appeared in Communications Chemistry.
Dividing complex workflows among AI agents
Computational chemistry simulations can require researchers to coordinate multiple software packages and correctly configure numerous parameters. Errors during setup or execution can produce inaccurate results, failed calculations, or unnecessary computing costs.
ChemGraph uses specialized AI agents to divide that work. A planner agent breaks a scientific request into smaller tasks, executor agents call the required simulation tools, and an aggregator agent combines the outputs into a final result.
Rather than relying solely on information generated by a large language model, ChemGraph directs the model to use physics-based simulation tools and scientific libraries. This tool-calling approach is intended to reduce the risk that the system will produce an unsupported answer instead of performing the requested calculation.
In one demonstration, a user asked ChemGraph to calculate the reaction enthalpy of methane combustion at 400 K. The system converted the chemical names into molecular representations, generated atomic coordinates, performed thermodynamic calculations, and returned the resulting enthalpy change.
Testing single- and multi-agent performance
The researchers evaluated ChemGraph across 360 instances representing 13 tasks, ranging from converting chemical names into machine-readable molecular structures to calculating thermodynamic properties.
Smaller language models performed well on relatively simple workflows. GPT-4o-mini and Claude-3.5-haiku achieved accuracy above 80 percent on tasks involving a limited number of tool calls. Performance declined as workflows grew more complex and the models had to process larger volumes of intermediate data.
Dividing complex calculations among multiple agents improved the results. On a reaction enthalpy task, the accuracy of GPT-4o-mini increased from 40 percent in the single-agent configuration to 87 percent in the multi-agent configuration. Claude-3.5-haiku also reached 87 percent, up from 67 percent.
For a Gibbs free energy task, GPT-4o-mini improved from 49 percent to 87 percent, while Claude-3.5-haiku increased from 69 percent to 93 percent. GPT-4o completed both complex tasks with 100 percent accuracy in the multi-agent tests.
Operational considerations for research labs
ChemGraph does not automate physical experiments. Instead, it provides an orchestration layer for computational work, allowing researchers to create and modify simulation workflows without manually coding each step.
For lab managers overseeing computational research, the framework illustrates how breaking AI-enabled workflows into defined roles can improve performance and make outputs easier to trace. It also highlights the need to preserve intermediate calculations, tool calls, software versions, parameters, and error messages so researchers can review how the system reached a result.
Advanced Lab Management Certificate
The Advanced Lab Management certificate is more than training—it’s a professional advantage.
Gain critical skills and IACET-approved CEUs that make a measurable difference.
The researchers reported that models still made mistakes, including selecting incorrect tool arguments, confusing molecular properties, and omitting values when summarizing results. ChemGraph therefore requires scientific review rather than autonomous acceptance of its outputs.
Argonne provides the framework in a controlled software container to restrict unintended interactions with host systems. The team also used open-weight language models hosted on Argonne computing infrastructure to address costs and data-security concerns associated with external commercial systems.
The current version remains an initial release. Argonne researchers plan to expand its compatibility with additional tools and support larger, longer-running simulations. They have also begun adapting ChemGraph for spectroscopy analysis and high-throughput materials screening.
This article was created with the assistance of Generative AI and has undergone editorial review before publishing.
