Automated and high speed freeform metasurface design using an agentic large language model framework

We present MetaChat, an innovative text-based interface leveraging large language models for nanophotonics simulation and design. This platform integrates a decentralized ensemble of physics-augmented, high-speed surrogate full-wave solvers, enabling real-time interactive design across various incident source field angles and wavelengths. MetaChat introduces a novel method for conditioning machine learning surrogate solvers using light-weight conditional control networks, enhancing adaptability for specialized applications. We demonstrate its capabilities by designing large-area metasurfaces composed of scatterers co-optimized using the adjoint variables method, translating semantic design objectives into quantitative figures of merit. As a proof of concept, we showcase the rapid design of a multifunctional TiO2 metasurface operating across multiple visible wavelengths. This new design paradigm promises to streamline workflows in multi-physics computing, making advanced design tools more accessible to practitioners.