Duke University researchers have engineered an advanced AI system capable of solving complex design problems with an efficiency close to that of human scientists. This breakthrough hints at a future where AI could significantly accelerate innovation across various scientific disciplines.
Engineers at Duke University have developed a team of AI bots that can autonomously solve intricate design problems almost as proficiently as trained scientists. This study, recently published in ACS Photonics, suggests that AI could soon take on narrow, yet sophisticated, design challenges, unleashing a wave of rapid progress across many fields.
“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,” Willie Padilla, the Dr. Paul Wang Distinguished Professor of Electrical and Computer Engineering at Duke, said in a news release. “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.”
The specific type of challenge their AI system tackles is an ill-posed inverse design problem, where researchers know the intended outcome but face countless possible solutions without clear guidance on which might be optimal.
Previously, Padilla and his team successfully addressed such challenges for creating dielectric (metal-free) metamaterials. These synthetic materials, with unique electromagnetic properties derived from their structure, rather than chemistry, posed intricate design parameters that AI could efficiently manage.
In the new study, the researchers advanced their techniques by employing a suite of large language model (LLM) AI agents to handle the design tasks autonomously. Unlike earlier efforts where human graduate students had to perform repetitive steps, these AI agents could manage the entire process.
“The idea was to create an ‘artificial scientist’ that could learn metamaterial physics and work out solutions on its own,” Padilla added.
This “agentic system” comprises multiple specialized LLMs. One manages data organization, another generates deep neural network code, a third verifies accuracy, and yet another applies the lab’s “neural-adjoint” method to refine solutions. An overarching LLM oversees communication and progress among these agents, essentially mimicking the decision-making process of a human scientist.
“It will literally tell you if it is running into diminishing returns and needs to generate more data or if it’s happy with how the error rate is dropping and needs to continue iterating,” added Dary Lu, a doctoral student in Padilla’s lab who led the project. “It’s similar to the intuition that a scientist needs to develop over time and was probably the hardest part to program.”
Testing their system against previously solved inverse design problems, the researchers found that the AI’s best designs were comparable to those produced by human experts. Although the AI’s average results did not surpass those of doctoral students, its top solutions were sufficiently close, and in this field, achieving even a single optimal design is a significant success.
Padilla believes this method demonstrates the potential for AI-driven productivity in scientific research.
“Having AI systems that can conduct their own research and improve on their own methods will start making significant gains to push human knowledge,” Padilla said. “At large scales and on significantly faster timelines, these systems will soon be able to produce truly novel results.”
Lu sees a promising future for such AI applications.
“We are right on the cusp of where systems like these will be able to enhance the productivity of highly skilled workers,” he added. “Being able to build these agentic systems is a valuable skillset to have going into the job market.”
Source: Duke Pratt School of Engineering