Chat to Chip: Large Language Model Based Design of Arbitrarily Shaped Metasurfaces

Technical Summary: Chat to Chip: Large Language Model Based Design of Arbitrarily Shaped Metasurfaces

Problem Statement
Traditional metasurface design is constrained by the computational expense of full-wave electromagnetic simulations (e.g., FDTD, FEM), which limits the exploration of complex, arbitrarily shaped configurations. While data-driven deep neural networks (DNNs) have emerged as a faster alternative, they require a laborious, task-specific workflow for every new optical function. This process involves building custom network topologies, exhaustive hyperparameter searches, and iterative engineering, which remains a barrier for researchers without extensive machine learning expertise. Furthermore, existing approaches often struggle with the inverse design of high-degree-of-freedom, randomly shaped 2D unit cells due to the "many-to-one" nature of the problem, where distinct geometries produce identical spectra, causing gradient cancellation and non-convergence in deterministic networks.

Methodology
The authors propose a "chat-to-chip" workflow that leverages pre-trained Large Language Models (LLMs) to bypass the need for custom DNN engineering. The core methodology involves:

1. Data Generation and Parameterization:

   • Arbitrarily shaped meta-atoms are generated using a 4×4 control-point grid. This grid is replicated via four-fold rotational symmetry, interpolated into a 256×256 scalar field, and binarized (threshold $t=0.5$).
   • Morphological operations (iterative opening and closing) are applied to ensure fabrication friendliness by removing isolated islands and filling holes.
   • The resulting silicon patterns (refractive index 3.5) on a glass substrate (refractive index 1.5) are simulated using Lumerical FDTD under left-handed circularly polarized (LCP) normal incidence.
   • A dataset of 45,790 designs was created, recording transmission spectra at 31 uniformly spaced wavelengths (1050 nm to 1600 nm).

2. LLM Formulation:

   • The geometric data (4×4 grid) and optical response (31-point transmission vector) are converted into natural language prompts and completions.
   • Forward Design: The prompt asks for the transmission spectrum given a grid; the completion provides the 31 numerical values.
   • Inverse Design: The prompt provides a target spectrum and asks for "one possible grid," explicitly leveraging the LLM's stochasticity to handle the many-to-one mapping problem.
   • Numerical values are rounded to three decimal places and tokenized using a byte-pair encoder consistent with the base model.

3. Model Training and Architecture:

   • The study utilizes the Meta-Llama-3.1-8B-Instruct model, quantized to 4-bit weights.
   • Parameter-efficient fine-tuning is achieved via Low-Rank Adaptation (LoRA) with rank $r=32$ and scaling factor $\alpha=32$.
   • Training is conducted on consumer-grade GPUs (NVIDIA RTX 2080 Ti) for 8 epochs using the AdamW optimizer. No architectural modifications to the LLM backbone are required.

Key Results

• Forward Prediction Accuracy: The fine-tuned Llama-3.1-8B model achieves a Mean Squared Error (MSE) of $3.4 \times 10^{-3}$ on the test set, matching the accuracy of specialized DNNs reported in literature. The model predicts spectra in approximately 2 seconds, roughly 60 times faster than full-wave simulations on the authors' CPU cluster. The accuracy is largely insensitive to the number of fine-tuning epochs beyond 5.
• Model Benchmarking: The authors benchmarked 11 open-weight LLMs across three parameter sizes (<7B, 7–9B, >9B).
  • Accuracy generally improves with model size but is not strictly monotonic; for instance, Mistral 7B outperformed the 8B Llama variant.
  • Gemma models demonstrated superior accuracy-to-cost ratios, likely due to their digit-friendly tokenizers.
  • Scaling from 7B to 72B (Qwen-2.5) yielded diminishing returns (only a $1.2 \times 10^{-3}$ MSE reduction) while significantly increasing inference time and memory requirements.
  • Reasoning-centric models (e.g., Qwen-3B, Phi-4-Reasoning) were found unsuitable for this task as they tended to produce verbose diagnostic text rather than the required numerical sequences.
• Inverse Design Performance: The LLM successfully generated diverse geometries for target spectra. By prompting for "one possible grid," the model avoided the convergence issues typical of deterministic inverse networks. Inverse-designed meta-atoms showed high fidelity to target spectra (MSE as low as $1.2 \times 10^{-6}$) and exhibited significant geometric diversity. Over 88% of inverse design test samples achieved an MSE below $1.0 \times 10^{-2}$.

Significance and Claims
The paper claims to demonstrate that 1-D token-wise LLMs can serve as a practical, "no-code" tool for both forward and inverse design of 2D arbitrarily shaped metasurfaces without requiring vision models or complex network engineering.

• Lowering Barriers: The workflow eliminates the need for bespoke network design and exhaustive hyperparameter sweeps, making data-driven nanophotonic design accessible to researchers with limited machine learning expertise.
• Efficiency: The approach provides near-real-time feedback (seconds) compared to days or weeks for traditional simulation-heavy workflows.
• Scalability: The study establishes that off-the-shelf, lightly fine-tuned LLMs can achieve high-fidelity results, suggesting that future gains may come from models with stronger numerical priors or Vision Language Models (VLMs) rather than simply scaling parameter counts.
• Paradigm Shift: The work represents a step toward a "chat-to-chip" paradigm, linking natural language interaction directly to electromagnetic modeling and enabling the rapid exploration of complex, multifunctional metasurfaces.