End-to-end multichannel holographic metasurface inverse design using an enhanced bi-directional deep neural network

We present an end-to-end inverse-design system that combines a periodicity-aware modified gradient descent (MGD) with an enhanced bidirectional neural network (EBI-DNN) to rapidly generate phase distributions and quickly get meta-atom parameters for multifunctional terahertz (THz) holographic metasurfaces. The MGD converts six grayscale letters into a smooth phase hologram, and the EBI-DNN, trained on 7.4 k samples in 40.8 s, 1.8 × faster than the conventional model, maps the phase to silicon pillar geometries parameters. This system completes designs within seconds and produces a 100 × 100 pixel, six-channel metasurface operating at 1 THz that projects the images "YT," "CO," and "DB" onto two focal planes with a peak signal-to-noise ratio (PSNR) of 15.7 dB, root-mean-square (RMS) error of 1.72. The system's combined advantages of computational speed, structural accuracy, and compact form factor offer a practical pathway toward large-area, multifunctional metasurface technologies.