Imagine wanting to design an engine that runs on less fuel. You could build a virtual version of it on a computer and test different designs without ever touching a real machine. That sounds simple — but the reality is far more complicated. Building those precise computer simulations usually requires years of specialized training. Now, scientists at Argonne National Laboratory in Illinois have created a tool that could make this kind of research much easier for everyone.
The tool is called ChemGraph. It's a free, open-source software framework that uses artificial intelligence to automate the complex steps involved in chemistry and materials science calculations. Developed by researchers at the U.S. Department of Energy's Argonne lab, ChemGraph could help speed up work on better car batteries, more efficient engines, and cleaner ways to pull useful materials from the earth.
"Running such simulations often requires a doctorate's worth of knowledge and dozens of steps," said Murat Keçeli, an Argonne computational scientist who helped lead the project. For someone to model how methane molecules behave during combustion, for example, they would need to know which scientific methods to use, which software programs work together, how to prepare their data, and how to analyze the results — often running dozens of calculations in sequence.
ChemGraph gets around that problem by acting like a team of AI assistants. Different agents — specialized mini-programs — handle different parts of the job: planning the approach, running the calculations, and putting the results together. A researcher can simply describe what they want to study in plain English, and ChemGraph figures out which tools and steps are needed to get there.
The team built ChemGraph using one of the world's most powerful computers: the Aurora exascale supercomputer at Argonne's Leadership Computing Facility. They also used the ALCF Inference Service, a new platform that gives researchers easy access to large language models — the same kind of AI technology that powers chatbots like ChatGPT — but connected to high-performance computing systems.
Keçeli didn't come up with the idea overnight. About a decade ago, he was already working on ways to automate chemistry tasks. In 2017, during his postdoctoral work with Stephen Klippenstein's research group at Argonne, he created the Quantum Thermochemistry Calculator, a set of coded modules for chemistry calculations. When large language models exploded in popularity in late 2022, he saw a new opportunity. "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."
The researchers published their findings in the journal Communications Chemistry. They hope ChemGraph will lower the barriers not just for professional scientists, but also for students learning computational chemistry for the first time.