Neural network enabled metasurface design for phase manipulation.

Area of Science:

• Metasurfaces and Nanophotonics
• Computational Electromagnetics
• Artificial Intelligence in Optics

Background:

• Artificial metasurfaces offer precise control over electromagnetic wave phase, enabling advanced optical devices like metalenses and holographic systems.
• Designing complex metasurfaces with numerous geometric parameters is challenging using conventional simulation methods due to time constraints.
• Deep learning models can learn intricate relationships between nanostructure geometry and electromagnetic responses.

Purpose of the Study:

• To develop a deep learning approach for the inverse design of metasurfaces with on-demand phase characteristics.
• To demonstrate the capability of neural networks in handling multiple geometric parameters for metasurface design.
• To achieve simultaneous phase and group delay prediction for achromatic metalens applications.

Main Methods:

• Utilized deep neural networks with innovative training methods to predict phase values based on six geometric parameters.
• Implemented a direct inverse design methodology for metasurfaces based on deep learning predictions.
• Extended the deep learning model for simultaneous prediction of phase and group delay to address achromatic requirements.

Main Results:

• Successfully demonstrated accurate phase prediction using deep neural networks with six geometric parameters.
• Achieved the first direct inverse design of metasurfaces for specific phase requirements.
• Showcased simultaneous phase and group delay prediction, enabling near-zero group delay dispersion for achromatic metalenses.

Conclusions:

• Deep learning significantly enhances the design capabilities for complex metasurfaces.
• The developed deep learning approach facilitates efficient and accurate inverse design of optical devices.
• This work paves the way for creating ultra-compact, high-performance metalenses and other nanophotonic devices.