VortexChat: An agentic framework for autonomous multi-objective integrated photonic design
VortexChat is an autonomous agentic framework that leverages a large language model to orchestrate a closed-loop workflow of topology generation, gradient-based refinement, and full-wave simulation, enabling the end-to-end inverse design and successful fabrication of complex integrated photonic devices directly from natural language specifications without human intervention.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Modern technology increasingly relies on tiny chips that guide light instead of electricity, enabling faster communications, advanced imaging, and new ways to process information. These devices, known as integrated photonic circuits, are built by carving microscopic patterns into materials like silicon to control how light moves. Designing these patterns has traditionally been a slow, manual process. Engineers must guess a shape, run complex computer simulations to see how light behaves, and then tweak the design based on the results, repeating this cycle many times until the device works as intended. This approach demands deep expertise and significant time, often limiting how quickly new technologies can be developed. While computer programs exist to help find better shapes automatically, they usually still require a human expert to set the rules, choose the starting points, and decide when to stop or change course.
A team of researchers has now introduced a system called VortexChat, which aims to remove the need for that constant human supervision. This system acts as an autonomous agent, a digital worker capable of understanding a simple request in everyday language and then independently designing a complex light-guiding device from start to finish. The researchers tested this system by asking it to create a specific type of optical component: a multiplexer that splits light into different swirling beams, known as vortex beams, each carrying a unique amount of spin. The system successfully designed a working device, built it in a laboratory, and confirmed that it performed exactly as the computer predicted, all without a human engineer intervening in the design loop.
The core innovation of VortexChat lies in how it mimics the workflow of a human expert but does so with a machine's speed and consistency. Instead of a human researcher manually connecting different software tools, the system uses a large language model as a central brain. This brain listens to a natural language description of what the device should do, such as "create a device that splits light into four different swirling patterns with high efficiency." Once the goal is understood, the agent breaks the task down and coordinates three specialized tools to achieve it. First, it uses a generative model to create a rough, initial shape that looks promising. Then, it passes this shape to a refinement tool that uses physics-based math to smooth out the design and ensure it can actually be manufactured with standard factory processes. Finally, a third tool runs a full simulation to test how light would actually travel through the proposed structure.
If the simulation shows the device is not quite right, the system does not stop. Instead, it analyzes the results, decides what went wrong, and adjusts its strategy. It might tweak the parameters of the refinement tool or ask the generative model to try a completely different starting shape. This cycle of design, testing, and adjustment continues automatically until the device meets all the required performance standards. The researchers tested this approach on a challenging benchmark involving one hundred different design tasks. In these tests, the system successfully completed 71 percent of the tasks on its own, meeting strict requirements for how much light it could transmit, how pure the swirling patterns were, and how well it worked across a range of colors.
To prove that the designs were not just computer fantasies, the team took one of the autonomous designs and built it in the real world. They fabricated a device on a silicon chip, creating a structure roughly the size of a grain of sand that could split incoming light into four distinct swirling beams. When they tested this physical device in a laboratory, it performed with high efficiency and low interference between the different beams, matching the computer simulations almost perfectly. The measurements confirmed that the light emerged with the correct swirling patterns and maintained its quality across a broad range of wavelengths. This successful fabrication demonstrates that the system can navigate the gap between abstract mathematical goals and the physical realities of manufacturing, producing devices that are not only theoretically sound but also practically buildable.
The significance of this work extends beyond the specific device created. It shows that artificial intelligence can take on the role of an expert designer, handling the complex trade-offs between performance and manufacturability that usually require years of human experience. By translating a simple sentence into a fully realized, working photonic chip, the system lowers the barrier to entry for creating advanced optical technologies. While the current version was tested on a specific type of light-manipulating device, the framework is designed to be adaptable. It suggests a future where scientists and engineers can describe their goals in plain language and let an autonomous agent handle the intricate details of the design process, potentially accelerating the development of new technologies in communications, computing, and sensing.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.