TLDR: Engineers at Duke University have developed an ‘agentic system’ of AI agents, powered by large language models (LLMs), that can autonomously solve complex design problems, particularly in the field of metamaterials. Published in ACS Photonics, this research suggests a future where AI could significantly accelerate scientific discovery by automating intricate design processes, performing nearly as well as human experts in generating optimal solutions.
DURHAM, NC – A groundbreaking study from Duke University reveals the creation of ‘virtual scientists’—an innovative system of artificial intelligence agents designed to tackle advanced scientific challenges with remarkable efficiency. This development, detailed in a recent publication in ACS Photonics on October 18, 2025, could herald a new era of accelerated discovery across various scientific disciplines.
The research, led by engineers at Duke University, focuses on an ‘agentic system’ where multiple large language models (LLMs) are programmed to collaborate as a team of ‘virtual scientists.’ This collective AI is capable of autonomously designing and testing complex materials, specifically metamaterials, which are synthetic materials engineered for unique electromagnetic properties. The system operates with minimal human intervention, assigning specialized tasks to individual AI agents, such as data organization, neural network code generation, accuracy verification, and optimization of results. A central LLM orchestrates communication among these agents and monitors their progress.
Professor Willie Padilla, the Dr. Paul Wang Distinguished Professor of Electrical and Computer Engineering at Duke, initiated this research after a colleague presented a challenging problem in modeling chemical reactions. Padilla recognized the potential for AI to solve such ‘ill-posed inverse design problems’—scenarios where the desired outcome is known, but the path to achieving it involves an infinite array of potential solutions. “A few years ago, a colleague described a really challenging problem in modeling chemical reactions to me. I knew it was something a standard deep learning AI program could solve, but didn’t have time to help myself,” Padilla stated. “But it got me thinking, if we could create a group of AI agents that could solve these types of problems autonomously, it could greatly speed up the rate of advancement in many fields.”
In previous work, Padilla’s lab had developed a deep neural network and a ‘neural-adjoint’ AI method to tackle these problems for dielectric metamaterials. The new study builds upon this foundation by replacing the labor-intensive steps traditionally performed by graduate students with a suite of LLM agents. “The idea was to create an ‘artificial scientist’ that could learn metamaterial physics and work out solutions on its own,” Padilla explained.
When tested against problems previously solved by human experts, the AI system demonstrated impressive capabilities. While its average performance over thousands of trials was slightly lower than that of PhD students, its top-performing designs were strikingly close to those achieved by human experts. Researchers emphasize that in many engineering fields, the objective is to achieve one exceptional design, rather than merely a high average. This indicates that well-programmed agentic AI systems can effectively solve highly complex problems. Padilla noted, “this demonstration shows that agentic systems can solve even the most complex problems when thoughtfully and thoroughly programmed.”
Also Read:
- AI’s Transformative Impact: The Top Trends of 2025 Reshaping Daily Life
- Salesforce AI Research Unveils WALT: A Novel Framework for Web Agents to Autonomously Discover and Utilize Website-Native Tools
This innovative approach suggests that agentic AI tools could soon enhance, or even automate, research processes in diverse fields, from materials science and chemistry to biology, freeing human researchers to focus on higher-level inquiries and analysis. The ability of these virtual scientists to articulate their reasoning further adds to their potential as collaborative partners in scientific exploration.
