Deep learning-assisted design of a tunable broadband absorber for the terahertz region

To accelerate the design of terahertz metamaterial absorbers, this study developed a fully connected forward neural network that maps known parameters to electromagnetic responses. This network can predict terahertz wave absorption rates based on structure parameters, achieving a mean squared error of only 0.002662 on the validation set. Using this forward network, a tunable broadband terahertz absorber was designed. The absorber consists of a vanadium dioxide (VO 2 )-gold composite metasurface and a polytetrafluoroethylene (PTFE) substrate. When VO 2 is in its metallic state, the absorption rate exceeds 90% within the 1.31–3.50 THz frequency range, reaching a peak absorption rate of 99.79% at 3.0 THz, while maintaining polarization insensitivity and stability under wide-angle incidence. The method enables absorber prediction and design significantly reducing design time and computational resources. It can be applied to high-speed communication, advanced sensing, and stealth technology, providing a practical solution for developing tunable broadband terahertz absorbers.