Physics-driven unsupervised deep learning network for programmable metasurface-based beamforming.

Area of Science:

• Electromagnetics
• Materials Science
• Artificial Intelligence

Background:

• Programmable metasurfaces dynamically manipulate electromagnetic (EM) waves.
• Designing digital coding patterns for metasurfaces is crucial for beamforming.
• Traditional optimization methods for coding patterns are computationally intensive and time-consuming.

Purpose of the Study:

• To propose a novel physics-assisted deep learning (DL) model for rapid calculation of digital coding patterns.
• To enable efficient beamforming with programmable metasurfaces by simplifying the design process.
• To reduce the computational complexity and time required for optimizing metasurface coding patterns.

Main Methods:

• Developed a physics-assisted deep learning model integrating an extended macroscopic model for digital coding metasurfaces.
• The model computes radiation patterns rapidly based on coding patterns, ensuring physics compliance.
• The DL model calculates the required digital coding pattern in milliseconds from a desired beam depiction.

Main Results:

• The proposed DL model successfully calculates coding patterns for generating a wide range of EM beams.
• Experimental validation demonstrated good beamforming performance for both single-beam and dual-beam scenarios.
• The method significantly reduces the time for coding pattern generation from hours to milliseconds.

Conclusions:

• The physics-assisted DL model offers a highly efficient and accurate approach for designing programmable metasurfaces.
• This breakthrough accelerates the development and application of dynamic EM wave manipulation technologies.
• The proposed method overcomes the limitations of traditional optimization algorithms in metasurface design.