ChemGraph as an agentic framework for computational chemistry workflows.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Atomistic simulations are crucial but require extensive expert knowledge for setup, execution, and validation.
• Current workflows in computational chemistry and materials science are often complex and time-consuming.

Purpose of the Study:

• To present ChemGraph, an AI-powered agentic framework designed to streamline and automate computational chemistry and materials science workflows.
• To leverage graph neural networks and large language models for efficient and intuitive scientific simulations.

Main Methods:

• Developed ChemGraph, an agentic framework integrating graph neural network foundation models and large language models (LLMs).
• Utilized LLMs for natural language understanding, task planning, and scientific reasoning.
• Evaluated performance across 13 benchmark tasks using various LLMs and a multi-agent decomposition strategy.

Main Results:

• Smaller LLMs (GPT-4o-mini, Claude-3.5-haiku, Qwen-2.5-14B) showed proficiency in simpler workflows.
• Complex tasks benefited from larger LLMs.
• A multi-agent framework enabled GPT-4o to achieve perfect accuracy and allowed smaller LLMs to match or surpass single-agent GPT-4o performance on benchmarks.

Conclusions:

• ChemGraph effectively automates and simplifies complex computational chemistry and materials science tasks.
• Task decomposition via a multi-agent system significantly enhances the performance of both large and small LLMs.
• The framework offers an intuitive interface, reducing the need for expert knowledge in running simulations.