Ultra-broadband terahertz absorber via deep learning

Area of Science:

• Metasurfaces
• Terahertz Technology
• Artificial Intelligence

Background:

• Traditional terahertz metasurface absorber design is complex, requiring extensive software-based modeling, simulation, and optimization.
• These conventional methods are time-consuming and computationally expensive, hindering rapid development.

Purpose of the Study:

• To develop a deep learning-based methodology for the rapid and accurate design of terahertz metasurface absorbers.
• To demonstrate the efficiency and versatility of AI in predicting geometric parameters and frequency domain responses for metasurface devices.

Main Methods:

• Utilized deep learning algorithms to predict the geometric parameters and frequency domain response of terahertz metasurface absorbers.
• Developed a method capable of bidirectional prediction, linking structural design to absorption characteristics.

Main Results:

• Achieved over 90% terahertz wave absorption in an ultra-wideband range (6.31-16.23 THz).
• The designed absorber is insensitive to incident wave polarization.
• Metasurface absorber structure prediction completed in less than 3 nanoseconds.

Conclusions:

• Deep learning offers a significantly faster and more efficient alternative to traditional methods for designing terahertz metasurface absorbers.
• The proposed AI-driven approach drastically reduces design time and computational costs.
• This methodology is broadly applicable to the design of various terahertz metasurface devices, including those for polarization conversion, focusing, and reflection.