Researchers at Duke University have developed a group of artificial intelligence bots capable of addressing intricate design challenges with a proficiency akin to that of trained scientists. The findings, published on October 18 in the journal ACS Photonics, suggest that such AI systems could soon automate specialized design tasks, potentially leading to significant advancements in various fields.
Willie Padilla, the Dr. Paul Wang Distinguished Professor of Electrical and Computer Engineering at Duke, recalls a conversation with a colleague regarding a complex modeling issue in chemical reactions. He believed a standard deep learning AI could tackle the problem but lacked the time to assist. This prompted him to consider the creation of autonomous AI agents to expedite the resolution of such challenges and accelerate progress in multiple disciplines.
The study focuses on a type of problem known as an ill-posed inverse design problem, where researchers have a specific outcome in mind but face countless possible solutions without clear directions on which might be optimal. In earlier research, Padilla”s team devised methods to solve these difficult issues related to dielectric metamaterials—man-made materials engineered with specific structures to exhibit properties not found in nature. This work involved designing deep neural networks that analyzed extensive simulated data to uncover the relationships between design parameters and their corresponding effects.
In the recent study, the researchers followed a similar approach but utilized a suite of large language model (LLM) AI agents to execute the necessary tasks instead of relying on graduate students. “We aimed to create an “artificial scientist” capable of understanding metamaterial physics and autonomously working through solutions,” explained Padilla.
The agentic system comprises several specialized LLMs, each responsible for distinct functions: one compiles and organizes necessary data, another generates deep neural network code based on prior examples, and a further model verifies the work”s accuracy. An overarching LLM orchestrates communication between these agents, ensuring they work cohesively toward a solution. This system can identify when additional data points are required for better modeling or when the current model is sufficiently effective. “It can inform users about its progress and whether it needs more data or is satisfied with the decreasing error rate,” stated Dary Lu, a PhD student in Padilla”s lab who led the project.
Although the AI did not consistently outperform previous PhD students across thousands of trials, its top solutions were notably competitive. Padilla emphasizes that this research demonstrates the potential of agentic systems to address complex problems efficiently when designed thoughtfully. He believes this methodology could extend to numerous disciplines beyond computational electromagnetics.
“We are on the brink of a transformative era where such systems will enhance the productivity of highly skilled professionals,” remarked Lu. “Acquiring the skills to develop these agentic systems will be advantageous in the job market.”
Padilla added, “AI systems capable of conducting independent research and refining their methodologies will significantly advance human knowledge, producing genuinely novel outcomes at an accelerated pace.”
CITATION: “An Agentic Framework for Autonomous Metamaterial Modeling and Inverse Design.” Darui Lu, Jordan M. Malof, and Willie J. Padilla. ACS Photonics 2025. DOI: 10.1021/acsphotonics.5c01514
For more news on AI advancements in engineering, visit Duke Engineering”s website.
