Computers have made it easier than ever before to design the perfect material for a given problem: Scientists can create a virtual version and simulate how that material will behave. Building these atomically precise simulations, however, typically requires deep expertise in computational chemistry. At the U.S. Department of Energy's (DOE) Argonne National Laboratory, researchers have developed a kind of shortcut, streamlining scientific workflows using artificial intelligence (AI).
ChemGraph is an open-source, publicly available framework that automates some of the steps required when performing calculations for materials science and chemistry. This could help accelerate efforts such as boosting engine efficiency, extracting critical materials and making better batteries. The framework was described recently in the journal Communications Chemistry.
The team developed ChemGraph using resources at the Argonne Leadership Computing Facility (ALCF), including the Aurora exascale supercomputer and the ALCF Inference Service, a first-of-its-kind platform that gives researchers cloud-like access to a broad range of large language models (LLMs) on the facility's high performance computing systems. The ALCF is a DOE Office of Science user facility.
A Team of AI Assistants for Complex Science
ChemGraph's purpose is to lower the barriers to innovation for both scientists and students. Let's say you want to design a gas turbine engine that derives more power from less fuel. For that, you need to understand various aspects of methane combustion, such as the exact conditions that will help get the most value out of the gas. Computer simulations will help answer questions about how methane molecules behave as they go through the combustion process.
Running such simulations often requires a doctorate degree's worth of knowledge and dozens of steps. You need the theoretical background to know which scientific methods to use for a study. You need to identify which software is compatible with those methods. You must then prepare your input file (the data) and navigate the software to get results. Then you will put those results into a separate tool for further analysis, running sequential calculations, fine-tuning parameters and comparing results along the way before arriving at a conclusion. This is your workflow.
Materials scientists and chemists have the theoretical and experimental background to write research papers and carry out real-world experiments, but they may not have the technical savvy or staff to run these workflows. ChemGraph assigns different parts of a workflow to agents, which are akin to assistants that specialize in different tasks, such as planning, executing work or aggregating data.
About a decade ago, Argonne computational scientist Murat Keçeli had already been working on automating some of the tasks involved in chemistry through rule-based automation. Computer scientists have long used this strategy to achieve leaps in productivity - at its most basic, think of macros on a computer that can bundle steps together and execute them in one keystroke. In 2017, during his postdoctoral work with Argonne Distinguished Fellow Stephen Klippenstein's group, Keçeli developed the Quantum Thermochemistry Calculator, a series of coded modules for thermochemistry calculations, before moving on to other projects.
Then ChatGPT, the generative AI powered by LLMs, emerged in late 2022. "When this large language model breakthrough happened, I thought, 'I should go back to that workflow automation,'" Keçeli said. "Basically, we wanted to put all of our expert knowledge about workflows into an agent-based automation that you could talk to through an LLM."
ChemGraph uses LLMs to provide a natural language interface to its agent-based automation. A researcher can state the scientific problem in plain language, and the framework maps that request onto a sequence of computational tasks, software tools and analyses needed to produce the result.
Argonne researchers designed ChemGraph to call only the right types of tools and libraries to minimize the risk of hallucination, a well-known phenomenon in which generative AI fabricates answers.
"We don't want the large language model to just answer the questions," said Thang Duc Pham, an Argonne postdoctoral fellow and ChemGraph co-creator. "We want it to run physics-based simulations and get an answer for you, instead of just relying on what it knows." He noted that this capability is also useful in cases where a problem has not been studied yet and new data is needed for a hypothesis.
ChemGraph complements DOE's Genesis Mission, a national initiative to accelerate science through AI. Even when computational chemists run simulations, problems inevitably surface somewhere along the workflow, Keçeli noted. ChemGraph aims to simplify a complicated process and minimize hassle so that scientists can focus on their research goals.
Force Multipliers: AI Agents, Human Collaborators and ALCF
The ChemGraph team, which also included Aditya Tanikanti (a former Argonne computer scientist now at DOE's SLAC National Accelerator Laboratory), initially built the framework with a single agent. But they saw that it began to fail when problems reached a certain level of complexity. They also realized that some tasks could be handled by smaller language models, while others required more sophisticated reasoning LLMs. Multiple agents could do the same job more efficiently.
"If you use only one type of LLM for everything, then you risk wasting money and allotted compute time," Keçeli said. "We found that we could start with a big model for workflow planning and then revert to smaller models for execution tasks."
The team used ALCF's Inference Service to access powerful open-weight models on facility systems rather than through external cloud providers, helping reduce cost and address data-security concerns. They also leveraged the Aurora supercomputer to run the computationally demanding quantum chemistry simulations embedded in ChemGraph, underscoring the complementary roles of AI inference and large-scale high performance computing in the framework.
With its ability to make computational chemistry more accessible, ChemGraph is already seeing interest at universities. Professors can use it as a teaching tool and students can use it to explore their own research questions. Because ChemGraph is open source, it is also adaptable to tasks beyond the ones in the initial release.
"We have already added one new feature to ChemGraph through a hackathon last fall, and as we collaborate with more people, we are hoping to expand ChemGraph's capabilities beyond our own expertise," Pham said.
In one recent collaboration at Argonne, researchers adapted ChemGraph for X-ray absorption near-edge structure (XANES) simulation and analysis, helping automate a spectroscopy workflow from user requests through simulation, data processing and curation. In another effort with ALCF researchers, ChemGraph was extended to coordinate a high-throughput materials screening workflow on Aurora, demonstrating a path toward scalable, AI-driven scientific automation on exascale supercomputers.
Ultimately, the goal of any scientific simulation is to obtain results that translate to a real-world advance. This is the promise of autonomous discovery: Better simulations on computers translate to fewer failed experiments in the lab, bringing ideas to life faster.
"Our dream for ChemGraph is to make it available as a service for ALCF users through a chatbot-style interface," Keçeli said. "In the long run, we hope to make it increasingly autonomous, able to plan, execute and refine complex computational workflows with minimal user intervention, so scientists can focus on the scientific questions they want to answer."
Work on ChemGraph was supported by DOE's Office of Science.