Surrogate modeling and simulation of antenna parameters via artificial intelligence

It is observed that the amount of data used in modeling problems does not always directly and positively affect modeling performance. Although advanced statistical sampling methods like Latin Hypercube Sampling (LHS) are preferred for determining input data, model performance can be negatively impacted when the dataset is either too small or excessively large. This thesis aims to better understand this interaction by focusing on the data-driven modeling of a cavity-backed Vivaldi antenna operating in the Ku-band (12–18 GHz) frequency range, specifically within the 10 GHz to 20 GHz range, using a multi-layer perceptron (MLP) model. Ku-band antennas are widely preferred in satellite communication systems due to their advantages in regional attractive force and atmospheric permeability. However, the inherently wideband operating characteristic of such antennas poses a significant challenge for accurate and reliable data-driven model creation. Furthermore, the presence of numerous input parameters in the antenna's structural design introduces the significant problem known as the "curse of dimensionality" during the model training process. For this reason, the LHS method was meticulously employed for the selection of input parameters to ensure model effectiveness and efficiency. In this study, an MLP model was developed for the 10 GHz to 20 GHz frequency range. The model's output aimed to accurately predict the frequency-dependent variation of the S11 parameter, obtained through full-wave electromagnetic (EM) simulations. The results obtained have clearly demonstrated that the model's predictions are in high agreement with the actual simulation outputs. This successful modeling methodology is adaptable and offers a valuable approach not only for the Vivaldi antenna type examined but also for different antenna configurations and general electromagnetic modeling problems.