Shaped Beampattern Synthesis of Planar Arrays With Fully Convolutional Networks

Abstract

In this letter, a method based on an encoder–decoder fully convolutional network is proposed for the efficient synthesis of planar arrays fulfilling user-defined beampattern masks. The decoder introduces a novel sub-pixel convolutional layer to yield an accurate and fast upsampling regression from the array excitations to the radiated pattern. The encoder is trained to minimize the deviation between the desired shaped beampattern and the actual one. By collaborating with the pretrained decoder, an efficient training of the encoder can be realized to yield a set of array excitations that fit the design objectives. Representative numerical results are reported to assess the effectiveness and efficiency of the proposed method.