Engineers at Duke University have developed a group of AI bots capable of solving complex design problems nearly on par with a trained scientist. The system, described in the journal ACS Photonics, represents a major step toward automating intricate scientific challenges, particularly in fields like materials science and engineering. The research focuses on what are known as ill-posed inverse design problems—situations where scientists know the desired outcome but face an infinite number of possible designs with no clear path to the best solution. In this case, the team aimed to design dielectric metamaterials—engineered structures that manipulate electromagnetic waves in ways not found in nature—without relying on metals. Previously, Padilla’s lab used deep neural networks trained on tens of thousands of simulations to map relationships between design parameters and electromagnetic responses. They then developed a "neural-adjoint" method that could work backward from a target outcome to find optimal designs. In the new study, the researchers replaced much of the manual work with an agentic system—a team of large language model (LLM) AI agents. Each agent handles a specific task: gathering and organizing data, writing code for a deep neural network from scratch, verifying accuracy, and applying the neural-adjoint method. An overarching LLM coordinates communication between agents and monitors progress. What sets this system apart is its ability to self-assess. It can determine whether it needs more simulation data to improve its model or if it’s making sufficient progress. It can also explain its reasoning at any stage, informing users whether it’s hitting diminishing returns or on track toward a solution. Dary Lu, a Ph.D. student who led the project, emphasized that teaching the AI to mimic scientific intuition—the ability to know when to stop iterating or when to gather more data—was one of the most difficult challenges. When tested on the same problems previously solved by human experts, the AI system did not outperform graduate students on average across thousands of trials. However, its top-performing designs were remarkably close to those created by humans. In scientific discovery, one high-quality solution can be more valuable than many average ones. The results suggest that with careful design, agentic AI systems can match human-level problem-solving in specialized domains. The team believes the approach can be extended to other areas beyond computational electromagnetics, including drug discovery, climate modeling, and robotics. Padilla sees this as a turning point. “We’re on the verge of systems that can significantly boost the productivity of skilled researchers,” he said. “These agentic systems aren’t just tools—they’re collaborators that can learn, adapt, and improve over time.” As AI continues to evolve, the ability to build and manage such systems is becoming a critical skill. The researchers believe that in the near future, these AI-driven systems will not only accelerate scientific progress but also generate genuinely novel discoveries at unprecedented speeds.