TLDR: Engineers at Duke University have developed a groundbreaking system of AI bots capable of solving complex design problems with a proficiency comparable to human scientists. This advancement, published in ACS Photonics, utilizes agentic systems of large language models to automate challenging ‘ill-posed inverse design problems,’ potentially accelerating scientific discovery across various fields.
A team of engineers at Duke University has achieved a significant breakthrough in artificial intelligence, constructing a group of AI bots that can collaboratively solve intricate design problems almost as effectively as a seasoned human scientist. This pioneering research, published online on October 18, 2025, in the journal ACS Photonics, suggests a future where AI could automate niche yet critical design challenges, paving the way for rapid advancements in numerous scientific disciplines.
The initiative was sparked by Dr. Willie Padilla, the Dr. Paul Wang Distinguished Professor of Electrical and Computer Engineering at Duke. Padilla recalled a particularly challenging problem in modeling chemical reactions that he knew a standard deep learning AI could tackle but lacked the time to address himself. This led him to conceptualize a system of autonomous AI agents that could resolve such complex issues, thereby significantly accelerating the pace of scientific progress.
The specific type of challenge addressed by Padilla and his team is known as an ‘ill-posed inverse design problem.’ These problems are characterized by a clear desired outcome but an overwhelming, often infinite, number of potential solutions, offering no clear path to the optimal one. In previous work, Padilla’s lab had found ways to solve these problems for dielectric metamaterials, which are synthetic materials with engineered features that produce properties not found in nature.
Building on this foundation, the researchers in their new paper programmed a suite of large language model (LLM) AI agents to undertake the extensive legwork typically performed by graduate students. This ‘agentic system’ was designed to function as an ‘artificial scientist,’ capable of learning metamaterial physics and independently deriving solutions.
When tested against ill-posed inverse problems previously solved by the lab, the AI system demonstrated remarkable capabilities. While its average performance over thousands of trials did not surpass that of PhD students, its best solutions were strikingly close to human-generated optimal designs. As Padilla noted, ‘in this area, one great design is the goal,’ highlighting the system’s potential to yield high-quality results. This demonstration underscores that thoughtfully and thoroughly programmed agentic systems can indeed tackle even the most complex scientific problems.
Also Read:
- AI-Powered ‘Virtual Scientists’ Developed at Duke University Promise to Accelerate Research
- University of Michigan Unveils New Framework for Evaluating Human-AI Teamwork
The implications of this research are profound. The ability of AI systems to autonomously conduct research and refine their methodologies heralds a paradigm shift in the scientific landscape. By embracing these advancements, scientists, engineers, and researchers could see significantly accelerated progress, uncovering new realms of knowledge through efficiencies achieved at unprecedented scales and speeds.
