An LLM-Orchestrated Agent for Directional-Coupler Design with Self-Consistent Eigenmode and FDTD Validation

The article presents a design agent utilizing a Large Language Model (LLM) to orchestrate the design of a silicon-on-insulator (SOI) 2x2 directional coupler, where the LLM proposes gap values and assesses convergence while deterministic solvers handle the physics. The design process incorporates a frequency-domain eigenmode solver for coupling coefficient estimation and a Finite-Difference Time-Domain (FDTD) stage for validation, achieving a cross fraction of 0.498, closely aligning with the target of 0.500. This approach demonstrates the potential of LLMs in enhancing the design workflow of photonic devices by integrating numerical simulation techniques, which could streamline development for practitioners in the field.