Inverse design of cross-polarization conversion metasurfaces using convolutional neural networks

In order to solve the inefficiency problem in the traditional manual design process of cross-polarization conversion metasurface, this paper proposes the optimization design method of cross-polarization conversion metasurface based on deep learning. First, a pseudo-ellipticity angle is proposed to characterize the cross-polarization conversion; second, combined with the coded metasurface, the inverse design of the cross-polarization conversion metasurface is realized based on the deep learning convolutional neural network (CNN) and the electromagnetic simulation software; and finally, a series of structural arrangements of the metasurface units are obtained. Through the demonstration example, a series of cross-polarization conversion metasurfaces are optimally designed, and these metasurfaces achieve narrow-band cross-polarization performance within 8-16 GHz, which verifies the feasibility of the inverse design method.