Deep learning-based inverse design of multi-functional metasurface absorbers

Area of Science:

• Physics
• Materials Science
• Electrical Engineering

Background:

• Terahertz (THz) perfect absorbers are crucial for various applications.
• Developing THz absorbers with desired properties is computationally intensive and time-consuming.
• Existing design methods often lack efficiency and speed.

Purpose of the Study:

• To develop a rapid and accurate method for designing THz perfect absorbers.
• To integrate simulated annealing (SA) with deep learning (DL) for accelerated metasurface design.
• To enable both forward prediction of absorption spectra and backward design of absorber structures.

Main Methods:

• A novel approach combining a simulated annealing (SA) algorithm with deep learning (DL) acceleration.
• Utilizing a forward neural network (FNN) for predicting absorption spectra from metasurface geometry.
• Employing backward design to derive absorber structures from desired spectral responses.

Main Results:

• Achieved an 80,000-fold increase in computational speed compared to full-wave solvers.
• Successfully designed low-frequency, high-frequency, and broadband absorbers in the 4-16 THz range.
• Demonstrated high accuracy with an error margin below 0.02 and a design time of only 10 minutes.

Conclusions:

• The integrated SA-DL approach offers a highly efficient and accurate method for THz perfect absorber design.
• This model presents a novel technique for metasurface design applicable to THz frequencies and potentially other domains.
• The method significantly reduces design time and computational cost, paving the way for practical THz device development.