Extrapolation of Radiation Pattern with Neural Networks: A Paradigm with LSTM-based and Generative Adversarial Networks

The radiation pattern (RP) specification is an important graphical representation of diverse quantities such as directivity, gain, or electric field/power density in various antenna designs. Hence, optimizing the RP will effectively influence the overall performance of any communication system. calculating the RP in both the E-plane and H-plane is time-consuming and requires additional effort with simulations, since the calculations require the knowledge of the surface current on the overall structure. To tackle this drawback, we propose impressive methodologies for achieving the RPs through neural network-based approaches: generative adversarial network (GAN), and long short-term memory (LSTM)-based deep neural network (DNN). These two networks are strong enough to predicting the RP specifications at specific frequencies. To prove the effectiveness of the proposed method, a frequency-selective surface structure operating at the X-band is designed and afterward, the RPs are predicted through the two proposed networks (i.e., GAN and LSTM-based DNN) at 10.5 GHz which shows good agreement.