ISTANBUL

Engineers at Duke University in the US have created a team of artificial intelligence (AI) agents capable of solving advanced design problems almost as well as trained scientists -- a development they say could accelerate discovery across multiple fields, according to research published last week in ACS Photonics.

The “agentic system” uses large language models (LLMs) programmed to work together as “virtual scientists.” The system can design and test metamaterials — synthetic materials engineered to have unique electromagnetic properties — without human supervision.

The AI network assigns specialized tasks to individual agents, such as organizing data, writing neural network code, verifying accuracy, and optimizing results. A central model coordinates their communication and monitors progress.

When tested, the system’s top-performing designs closely matched those of human experts, though its average performance was slightly lower.

Researchers said the approach shows that well-programmed AI systems can independently solve highly complex problems. They believe such agentic AI tools could soon enhance — or even automate — research in fields ranging from materials science to chemistry.