On-Demand Design of Metasurfaces through Multineural Network Fusion.

Area of Science:

• Metamaterials and Nanophotonics
• Computational Electromagnetics
• Machine Learning for Engineering

Background:

• Traditional metasurface design methods face limitations in creating complex, freestyle structures.
• Achieving on-demand design for specific electromagnetic responses requires efficient inverse design and forward prediction capabilities.

Purpose of the Study:

• To propose and validate a multineural network fusion method for the on-demand design of freestyle metasurfaces.
• To enable rapid generation of metasurface patterns tailored to user-defined spectral requirements.
• To demonstrate the capability of generating high-fidelity metasurface structures with specific electromagnetic characteristics.

Main Methods:

• Utilized a Wasserstein Generative Adversarial Network (WGAN) with U-net architecture for inverse structural design.
• Employed a Variational Autoencoder (VAE) for data compression and latent space mapping.
• Implemented an LSTM + Attention network for forward S-parameter spectrum prediction and validation.
• Represented vacuum and metallic materials using binary digits (0 and 1) in a 10x10 cell array.

Main Results:

• Successfully generated freestyle metasurface patterns based on user-defined spectra.
• Validated the generated patterns by simulating and comparing S-parameter spectrograms.
• Demonstrated the ability to rapidly discover high-fidelity metasurface patterns meeting specific electromagnetic responses.

Conclusions:

• The proposed multineural network fusion method significantly expands the possibilities for metamaterial design, enabling complex freestyle structures.
• The on-demand design approach effectively generates metasurface patterns with desired electromagnetic characteristics.
• The flexible fusion of multiple neural networks allows for adaptability to diverse design problems and optimization objectives.