New agentic AI platform accelerates advanced optics design

(Press-News.org) Hyper-realistic holograms, next-generation sensors for autonomous robots, and slim augmented reality glasses are among the applications of metasurfaces, emerging photonic devices constructed from nanoscale building blocks.

Now, Stanford engineers have developed an AI framework that rapidly accelerates metasurface design, with potential widespread technological applications. The framework, called MetaChat, introduces new computational tools and self-reflective AI assistants, enabling rapid solving of optics-related problems. The findings were reported recently in the journal Science Advances.

“This combination of agentic AI with high-speed surrogate computational models is new,” said Jonathan Fan, the paper’s senior author and an associate professor of electrical engineering. “We have multiple agents talking with those tools and the user to do complex design tasks. That’s a really big opportunity.”

Simulating metasurfaces Metasurfaces hold potential for innovations in imaging and sensing. The devices are made of building blocks so tiny that hundreds of thousands could cover a grain of sand. The nano-sized components can bend light in ways that are not possible with traditional “bulky” optics. For example, integrated into a phone camera, they can beam a dot array onto a user’s face, generating a high-resolution hologram that can be used for identification.

But such fine-scale design comes with challenges for optical engineers. The designs require extensive material knowledge, geometrical precision, and heavy-duty computing. Engineers use simulations that model how electrical and magnetic fields are produced and change over time, and the simulations must be run thousands of times as the design undergoes trial and error. On a typical computer, a single simulation can take tens of minutes.

“It really snowballs into something that takes on the order of weeks or even months to do for large devices,” said Robert Lupoiu, the study’s first author and an electrical engineering PhD student.

Accelerating optical designs The team sought out a more efficient process, with the help of AI.

First, they built the engine powering the new framework: a deep-learning neural network that solves Maxwell’s equations, which govern electric and magnetic fields. The new solver, named Feature-wise Linear Modulation (FiLM) WaveY-Net, can run a simulation more than a thousand times faster than conventional methods, solving equations in just milliseconds.

Then, the team built AI agents that play the roles of optics designers and materials experts. By prompting large language models, they outlined each agent’s job and workflow. While existing AI-based designers follow a pre-determined flowchart-like process, that can sometimes lead to dead ends or poor results. To improve decision-making, the team used prompts to give the AIs agency, including the ability to self-reflect. “In one round of design, it can go back to itself and think about what it did,” said Lupoiu. “Then it can make the next best decision without a predefined template of what it must do.”

MetaChat combines the computing tools and AI agents together, with a chat interface for users to make design requests. The team tested MetaChat with dozens of optics and photonics engineering problems. In one test, Lupoiu asked MetaChat to design a metal lens that can simultaneously focus blue light to one point, and red light to another point. From there, the materials expert AI agent queried a database to identify materials with the properties needed. The AI designer configured the tiny blocks and pinged Lupoiu with clarifying questions. That rapid back-and-forth is made possible by FiLM WaveY-Net, which evaluates the performance of each building block in fractions of a second. Altogether, it took 11 minutes to produce a downloadable design that was comparable to state-of-the-art devices.

“Seeing the agents being able to figure out the optical design tasks on their own, and then asking for input at strategic times to come up with a truly useful design for the user was mind-blowing,” said Lupoiu. “To see a future in engineering and science that will be driven by AI in much the same way is super exciting.”

Applying AI agents across industries The researchers found that MetaChat could design advanced optical devices in collaboration with human users in real time. MetaChat could provide access to specialized optical design knowledge across fields such as optical computing and astronomy, added Fan. “There’s a real shortage of optical designers,” said Fan. “There’s a huge need for people to build various types of photonic systems, so I think this is something that can really help with that.”

In the future, similar systems with autonomous AI agents could accelerate technology in other areas. Researchers in other fields could develop their own specialized, self-reflective AI agents. This could potentially allow rapid cross-disciplinary collaboration. “If everyone put in an effort to create these high-speed scientific computing and optimization tools, the collection of agents and these tools will really be pushing the limits of computation and innovation, ” said Fan.

Such platforms won’t make humans obsolete, though. “The goal is to utilize insights from people,” said Fan. “It takes a person to ask the right questions and to identify when something is not right.”

--

Fan is also a member of Stanford’s Bio-X and Wu Tsai Neurosciences Institute.

Other Stanford co-authors included PhD students Chenkai Mao, Yixuan Shao, and Tianxiang Dai; Kofi Edée of the Université Clermont Auvergne is also a co-author.

The research was funded by the National Aeronautics and Space Administration, the National Science Foundation under award, and Stanford’s Knight-Henessy Scholars and Graduate Fellowship programs.

END