Ultra-broadband terahertz absorber via deep learning

The traditional design of terahertz metasurface absorbers requires software for device modeling, simulation analysis, and parameter optimization, which is extremely time-consuming and labor-intensive. In this paper, we use deep learning methods to design an ultra-wideband terahertz metasurface absorber, which has a terahertz wave absorption over 90% in the frequency range of 6.31-16.23 THz and is insensitive to incident wave polarization. This method demonstrates the ability to quickly, accurately, and bidirectionally predict the geometric parameters and frequency domain response, greatly saving time and computational resources in designing metasurface-based absorbers. The time required to complete a metasurface absorber structure prediction is less than 3 ns. This method can be widely applied to the design of other terahertz metasurface devices, such as polarization conversion, focusing, and reflection.