Chat to chip: large language model based design of arbitrarily shaped metasurfaces

Area of Science:

• Nanophotonics
• Computational Electromagnetics
• Artificial Intelligence

Background:

• Metasurface design traditionally relies on computationally intensive full-wave simulations, limiting exploration of complex geometries.
• Data-driven methods, like neural networks, accelerate meta-atom design but require task-specific training and architecture searches.
• Existing approaches necessitate significant effort for new optical functions and lack user-friendly interfaces.

Purpose of the Study:

• To investigate the efficacy of pre-trained large language models (LLMs) for designing arbitrarily shaped metasurfaces.
• To establish a
• chat-to-chip
• workflow linking natural language to electromagnetic modeling for nanophotonic device design.
• To benchmark various LLMs for spectral prediction and inverse design tasks.

Main Methods:

• Utilized 1-D token-wise LLMs trained on descriptive inputs of metasurface geometries.
• Fed LLMs with geometric descriptions to learn physical relationships for spectral prediction and inverse design.
• Benchmarked open-weight LLMs, analyzing the correlation between model size and predictive accuracy.

Main Results:

• Demonstrated that LLMs can accurately predict metasurface spectra and perform inverse design tasks.
• Identified performance-accuracy relationships for LLMs at the billion-parameter scale.
• Showcased the practicality of 1-D token-wise LLMs for designing 2-D arbitrarily shaped metasurfaces.

Conclusions:

• LLMs offer a viable, efficient alternative to traditional simulation-based methods for metasurface design.
• The proposed
• chat-to-chip
• workflow significantly enhances user-friendliness in data-driven nanophotonics.
• This approach represents a significant step towards accessible and intuitive nanophotonic device engineering.