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

Researchers present an LLM-based design agent that orchestrates the optimization of silicon-on-insulator directional couplers by coordinating eigenmode solvers and FDTD simulations without performing calculations itself. The agent achieved a 50/50 optical splitter with 0.498 cross-fraction accuracy against a 0.500 target, demonstrating effective human-AI collaboration in photonic device engineering.

This research represents a meaningful advancement in how artificial intelligence can augment specialized engineering domains without requiring LLMs to possess deep physics knowledge. Rather than attempting to train neural networks on complex electromagnetic simulations, the authors designed an LLM to serve as an intelligent orchestrator—proposing design parameters, evaluating convergence criteria, and coordinating deterministic solvers that handle the actual physics computation. This hybrid approach leverages the LLM's natural language processing and reasoning capabilities while preserving the accuracy and interpretability of established numerical methods.

The work addresses a persistent challenge in computational photonics: designing optical components requires iterating through parameter spaces guided by both analytical understanding and numerical validation. Traditional approaches rely on human engineers or computationally expensive optimization algorithms. The symmetric phase-matched coupler design space provided sufficient analytical structure to guide the LLM's proposals, while the eigenmode solver and FDTD validation ensured physical consistency. The discovery that residual errors collapsed into a single constant phase offset attributable to excess coupling length demonstrates that the coupled system achieved genuine self-consistency rather than ad-hoc convergence.

The practical implications extend beyond this specific device. The methodology—using LLMs as intelligent parameter proposers within deterministic solver frameworks—offers a template for other engineering domains involving numerical simulation. This could accelerate design cycles in photonics, RF engineering, and materials science by reducing the expertise barrier and human iteration time. The approach also maintains transparency and reproducibility, critical for hardware design where physical validation is non-negotiable. Future applications might integrate this pattern with broader design automation pipelines.

• →LLMs can effectively coordinate complex engineering workflows by proposing parameters and managing convergence without performing the physics calculations themselves.
• →The hybrid approach achieved target 50/50 optical splitting performance with residual error of only 0.0017, validating the methodology.
• →Self-consistency was established by using common 2D effective-index models across both eigenmode and FDTD solvers, eliminating discrepancy sources.
• →Discovered invariant excess coupling length of 2.837 micrometers across varying design parameters enables predictable length corrections.
• →The framework provides a reusable template for LLM-orchestrated design in other numerical simulation-heavy engineering domains.