Bidirectional deep neural network-assisted design of terahertz vortex beam metasurfaces

Vortex beams, whose topological charges form an infinite-dimensional Hilbert space, theoretically offer unlimited communication capacity and hold great potential for future 6G applications. Addressing the challenges of time-consuming and inefficient traditional terahertz metasurface design methods, we propose a bidirectional deep neural network approach for designing all-dielectric transmissive vortex beam metasurfaces. The proposed metasurface was fabricated using 3D printing technology, with experimental results confirming its effective generation of terahertz vortex beams carrying the intended topological charge number l = -2. The study demonstrates that the trained bidirectional deep neural network can predict the phase and transmission coefficient for a single metasurface unit within 4 μs, significantly reducing design complexity. This approach saves substantial time and computational resources for designing terahertz vortex beam metasurfaces, providing an efficient pathway for rapid terahertz metasurface development.