Engineers at Duke University have developed a group of artificial intelligence (AI) programs that can solve complex design problems almost as well as a trained scientist.
These agentic systems use large language models (LLMs) to work together on tasks. The goal is to automate difficult design challenges, which could speed up progress in many fields. The research focuses on ill-posed inverse design problems, where the desired outcome is known, but there are countless possible solutions with no clear best choice. For example, designing materials with specific properties can be tricky because many designs could work.
In this study, the AI was used to design metamaterials, which are man-made materials with unique properties created by their structure, not their chemical makeup. Metamaterials can control things like light or electricity in special ways. The AI system includes several LLMs, each handling a specific job, like organizing data, writing code for a deep neural network, or checking results for accuracy. A main LLM oversees the process, ensuring the programs communicate and adjust as needed. The system can tell if it needs more data or if it’s making good progress toward the solution.
How the AI System Works
The AI was tested on the same design problems previously solved by human researchers in the lab. While the AI’s average results were not as good as those of experienced students, its best designs were very close to the human solutions. This shows the AI can produce high-quality results for complex tasks. The system’s ability to explain its process and adjust its approach makes it act like an “artificial scientist.” This could help automate niche tasks, freeing up human researchers for other work. The approach may also apply to fields beyond metamaterials, like medicine or engineering, potentially leading to faster discoveries and new technologies.
The engineers have described the methods and results of this study in a paper published in ACS Photonics.