This invention proposes a reconfigurable metasurface
beamforming design method based on CGAN and Gumbel-Sinkhorn. This method constructs a conditional
generative adversarial network (GAN) as the inverse design network. The generator takes the target
beam pattern as a conditional input to generate the corresponding encoding matrix. The Gumbel-Sinkhorn operator is introduced to discretize this continuous matrix, transforming it into an approximately doubly random
permutation matrix. The encoding is sorted and selected by multiplying this
permutation matrix with a predefined structure vector (SV), outputting a discrete 0 / 1 metasurface encoding array. A noisy forward prediction network is trained to predict the
beam pattern corresponding to a given encoding array. In the training of the inverse design network, the discretized encoding matrix output by the generator is input into the noisy-trained forward prediction network. The
mean square error between the predicted
beam pattern and the target beam pattern is calculated as the target loss term to optimize the network parameters and accurately learn the complex mapping relationship from the beam pattern to the encoding array.