Machine-learning models for controlling a Large-Area Electronics Reconfigurable Antenna

Abstract

AI in the real world does not perform as well as in the digital realm because reality is more complicated and requires significantly more data to train. Physically integrated (PI) sensing technology introduced more invariance in the sensor data. However, knowing the spatial location of PI sensors is complicated. Previous research proposed a solution through designing and implementing a reconfigurable antenna based on large-area electronics (LAE) technology. LAE is based on deposition and processing thin films at low temperatures. This enables many kinds of sensors on plastic or polyimide substrates so that they can be flexible, conformal, and stretchable. They also enable wireless communications with sensors with fine spatial resolution. In this thesis, we examine a reconfigurable antenna made of 11 by 11 metal patches with RF switches between them. This antenna employs LAE transistors as these passive RF switches to reduce energy costs. By changing the on-off state of these RF switches, we change the current distribution of the antenna, thus changing the output radiation pattern that can help us accurately locate sensors. However, there are 2208 different configurations to simulate using EM simulation, so a deep learning algorithm is necessary. This thesis proposes two different types of machine learning algorithms: Graphic Neural Network (GNN) and Convolutional Neural Network (CNN). We show that these models can replace the EM simulation in producing expected radiation patterns from input control vector patterns, thus reducing the time needed to train the antenna.