Stanford’s New Agentic AI Platform MetaChat Is Rapidly Transforming Advanced Optics and Metasurface Design

Stanford University engineers have introduced a powerful new agentic AI platform that could significantly change how advanced optical devices are designed. The framework, called MetaChat, combines ultra-fast computational models with multiple self-directed AI agents to dramatically speed up the design of metasurfaces, a class of cutting-edge photonic devices. The research, published in Science Advances in 2025, points toward a future where complex optical engineering tasks that once took months can now be completed in minutes.

Metasurfaces are already considered one of the most promising developments in modern optics. These devices are built from nanoscale structures, so small that hundreds of thousands of them could fit on a grain of sand. Despite their tiny size, metasurfaces can manipulate light in extraordinary ways, enabling technologies like hyper-realistic holograms, compact lenses for smartphones, next-generation sensors for autonomous robots, and lightweight augmented reality glasses. However, designing them has traditionally been slow, computationally expensive, and limited to a small pool of highly specialized experts.

MetaChat was created to tackle these exact challenges.

Why Designing Metasurfaces Is So Difficult

At the heart of metasurface design is the need to precisely control how light interacts with materials at the nanoscale. Engineers must account for complex electromagnetic behavior governed by Maxwell’s equations, which describe how electric and magnetic fields evolve over time. To test a single design, engineers typically run detailed electromagnetic simulations that can take tens of minutes per run on a conventional computer.

Designing a full metasurface means repeating these simulations thousands of times, adjusting geometries, materials, and configurations through trial and error. For large or advanced devices, this process can easily stretch into weeks or even months, making experimentation slow and innovation costly.

This is where MetaChat’s approach becomes especially interesting.

The Computational Engine Behind MetaChat

The foundation of MetaChat is a deep-learning-based physics solver called FiLM WaveY-Net. Instead of relying on traditional numerical solvers, this neural network directly learns how to solve Maxwell’s equations. Once trained, FiLM WaveY-Net can evaluate how light interacts with nanoscale structures in milliseconds, making it more than a thousand times faster than conventional simulation methods.

This speed changes everything. It allows design ideas to be tested almost instantly, enabling real-time iteration instead of long waits between simulation runs. Without this computational breakthrough, the rest of the MetaChat framework would not be possible.

What Makes MetaChat “Agentic”

What truly sets MetaChat apart is its use of multiple AI agents, each assigned a specific role in the design process. Some agents behave like optics designers, others act as materials experts, and each one has a clearly defined workflow. These agents are powered by large language models but are not limited to rigid, pre-programmed steps.

Instead, MetaChat gives these agents agency. They can evaluate their own decisions, reflect on previous design choices, and change direction when something is not working. This ability to self-reflect allows the system to avoid dead ends that often plague automated design tools.

The agents also communicate with one another, query databases for material properties, interact with simulation tools like FiLM WaveY-Net, and ask clarifying questions to the human user when needed. All of this happens through a chat-based interface, making the system accessible even to users who are not experts in optical simulation software.

A Real-World Design Test

To demonstrate MetaChat’s capabilities, the researchers tested it on dozens of real optics and photonics problems. One notable example involved designing a metal lens capable of focusing blue light to one point and red light to another point simultaneously. This is a challenging task that normally requires careful optimization and deep expertise.

Using MetaChat, the process unfolded collaboratively. The materials expert agent identified suitable materials by querying a database. The design agent configured nanoscale building blocks and evaluated their performance using the ultra-fast solver. Along the way, the system asked targeted questions to refine the design requirements.

The final result was a downloadable, fabrication-ready design produced in just 11 minutes, with performance comparable to state-of-the-art devices created through traditional methods.

What This Means for Optics and Photonics

The implications of MetaChat extend far beyond one experiment. There is currently a global shortage of skilled optical designers, even as demand for photonic systems continues to grow. Fields like optical computing, astronomy, advanced imaging, and sensing all rely on increasingly complex optical components.

By lowering the barrier to entry, MetaChat could make advanced optical design accessible to a much broader range of researchers, engineers, and students. Instead of spending months learning specialized simulation tools, users can focus on defining goals and evaluating results, while the AI handles much of the heavy lifting.

Importantly, the researchers emphasize that this technology is not meant to replace humans. Human insight remains essential for asking the right questions, identifying unrealistic results, and guiding the overall design process. MetaChat is best understood as a high-speed collaborator, not an autonomous replacement.

The Bigger Picture of Agentic AI in Science

MetaChat also serves as a broader example of how agentic AI systems could transform scientific research. The same principles—fast surrogate models combined with specialized, self-reflective AI agents—could be applied to other disciplines that rely on heavy computation, such as materials science, fluid dynamics, or structural engineering.

If researchers across fields build similar high-speed computational tools and agent-based frameworks, it could enable rapid cross-disciplinary collaboration and push the limits of what is computationally feasible today.

Understanding Metasurfaces and Their Future

To fully appreciate MetaChat’s importance, it helps to understand why metasurfaces matter so much. Unlike traditional bulky lenses and mirrors, metasurfaces achieve their effects using flat, nanostructured layers. This allows optical systems to become smaller, lighter, and more efficient.

As manufacturing techniques continue to improve, metasurfaces are expected to play a central role in future consumer electronics, medical imaging systems, and space-based telescopes. Tools like MetaChat could significantly accelerate the path from theoretical designs to real-world devices.

A New Direction for Engineering Workflows

MetaChat represents a shift in how complex engineering problems are approached. Instead of linear workflows dominated by manual simulation and optimization, this platform enables interactive, real-time collaboration between humans and AI agents. The result is a faster, more flexible, and potentially more creative design process.

As agentic AI systems mature, platforms like MetaChat may become standard tools in engineering labs, changing not just how devices are designed, but who can design them.

Research paper:
https://doi.org/10.1126/sciadv.adx8006