Deep-learning based broadband reflection reduction metasurface.

Area of Science:

• Metamaterials and Nanophotonics
• Electromagnetics
• Artificial Intelligence in Engineering

Background:

• Traditional reflection reduction metasurface (RRM) design relies on inefficient trial-and-error methods.
• Metasurfaces offer potential applications in stealth technology but require optimized designs.
• Current design processes are time-consuming and limit the exploration of broadband functionalities.

Purpose of the Study:

• To establish an intelligent design methodology for broadband reflection reduction metasurfaces (RRMs).
• To leverage deep learning for accelerated prediction and inverse design of metasurface structures.
• To overcome the limitations of traditional RRM design approaches.

Main Methods:

• Development of a forward prediction deep learning network to forecast metasurface polarization conversion ratio (PCR) rapidly.
• Construction of an inverse deep learning network to derive structural parameters from a target PCR spectrum.
• Arrangement of polarization conversion units in a chessboard layout for broadband RRM realization.

Main Results:

• The forward network predicts PCR in milliseconds, surpassing traditional simulation speeds.
• The inverse network enables immediate derivation of structural parameters for desired PCR spectra.
• Achieved a broadband RRM with relative bandwidths of 116% (reflection < -10 dB) and 107.4% (reflection < -15 dB).

Conclusions:

• A novel, efficient deep learning methodology for designing broadband RRMs has been established.
• The intelligent design approach significantly accelerates the development of polarization converters.
• The demonstrated RRM exhibits superior bandwidth performance compared to previous designs, validating the deep learning methodology.