VortexChat Automates Photonic Device Design with LLM Agents

The design of advanced integrated photonic devices is often a laborious process, heavily reliant on manual simulations and the specialized intuition of experts. While inverse design methods exist, they still typically require significant human oversight and lack end-to-end automation. To address these limitations, a new agentic framework called VortexChat has been developed, enabling the autonomous inverse design of photonic devices directly from natural language descriptions. VortexChat integrates a large language model (LLM) as its decision-making agent, which orchestrates topology generation, gradient-based refinement, and full-wave electromagnetic simulations. This closed-loop architecture allows the system to iteratively break down design objectives, manage computational tools, and adapt its strategies based on simulation feedback, all with minimal human input. The framework successfully met stringent performance thresholds on the Vortex100 Benchmark and autonomously designed a broadband terahertz perfect vortex beam multiplexer, which was then fabricated and experimentally validated, confirming its high efficiency and purity. This demonstrates the potential for LLM agents to take on expert design roles in photonics, paving the way for scalable, autonomous design of complex systems.