Abstract
Large language models (LLMs) have gained significant prominence in language processing, demonstrating remarkable performance across a wide range of tasks. Recently, LLMs have been explored in various scientific fields beyond language-based tasks. However, their application in the design of nanophotonic devices remains less explored. Here, we investigate the capabilities of LLMs to address nanophotonic design problems without requiring domain-specific expertise of the user. Our findings show that an LLM with in-context learning enables nonexpert users to calculate optical responses of multilayer films via numerical simulations. Through conversational interaction and feedback between the LLM and the user, an optimal design of the multilayer films can be also produced for the user-provided target optical properties. Furthermore, we fine-tune the LLM using text-based representations of the structure and properties of optical metasurfaces. We demonstrate that the fine-tuned LLM can generate metasurface designs with target properties by reversing the input and output text. This research highlights the potential of LLMs to expedite the nanophotonic design process and to make it more accessible to a wider audience.