Design of Sparse Antenna Arrays Using the Physics-Aware Generative Adversarial Network

An unsupervised machine learning method, referred to as the physics-aware generative adversarial network (PAGAN), is proposed for synthesizing a desired beam pattern by optimizing and configuring the feeding for each element, array spacing, and number of elements. By merging the related physical formulas for antenna synthesis into the loss function, the physics-aware neural network (PANN) generates the solution space, comprising the amplitude and phase of the array factor; then, the confrontation between expected and predicted beam patterns directs PAGAN with the PANN module to produce the desired radiation pattern. In the process, we integrate physical information and penalty terms to improve physical interpretability and guide the model toward consistent physical solutions, respectively. In addition, a new masking and encoding technique is established to guarantee the continuity of differentiation in the solution space and handle variable-length inputs. As shown by numerical examples, compared with existing methods, our proposed unsupervised machine learning method achieves the desired radiation pattern with fewer antennas automatically and in a faster and more efficient way.