Alternate Optimization with Deep Learning to Design Beam Deflector Under Aperiodic Near-Field Coupling Conditions

When designing metasurface beam deflectors, the near-field coupling under aperiodic conditions is significantly different from that under periodic conditions, and the electromagnetic (EM) responses are distorted which finally leads to an inaccurate design result. A deep learning (DL) model is proposed to estimate the EM responses of each metasurface element in the deflector under aperiodic conditions, and an alternate optimization (AO) process is used to determine the arrangement for elements so that the real responses of these elements can be aligned with the design target. The proposed method is validated by both transmissive and reflective deflectors in microwave bands. Simulation results show that the DL model can estimate the responses of elements accurately in aperiodic conditions. Also, compared with conventional optimization with time-consuming full-wave simulation, the AO based on the DL model can design the deflector in just several iterations with 3° error of phase responses.