Machine learning based design of gap waveguides

Abstract

The development of wireless network technologies and the increasing demand for these technologies encourage engineers to design at higher frequencies. As industry 4.0, autonomous vehicles, space applications such as 5G and starlink require instantaneous high data transfer, there is a popularity towards gigabit and terabit transmission systems. The increasing demand for data transfer has caused the frequency spectrum currently used to be insufficient and a design and research trend towards higher frequencies has emerged. There are different design difficulties at high frequencies. Hardware costs of high frequency systems seem to be a big problem for companies. Low cost and low loss structures should be used for RF components in data transmission. Waveguide and microstrip lines cannot meet the requirements of these systems at high frequencies. According to the researchers, gap waveguide structures will meet these requirements. Gap waveguide-style artificial magnetic conductor structures are difficult to analyze and design, and maxwell solutions in EM Simulators take a lot of time. In general, maxwell solutions take a lot of time, regardless of EM structure. Therefore, researchers benefit from different fields of study such as machine learning for a quick solution. This study has been prepared for this purpose and presents a machine learning-based study for gap waveguide structures. In the first stage of the thesis, existing studies about gap waveguide structures are examined and the history of gap waveguide and its status in the literature are explained. Then, the effects of changes in pin radius, pin height, period and gap height parameters on Mod-1 and Mod-2 frequencies in the dispersion diagram of the unit cell that make up the gap waveguide were analyzed. In order to use the data set in Python Orange program, parametric sweep was made in the CST Studio Suite program. The resulting data has been adapted to the machine learning program. Machine learning algorithms were researched and trial studies were conducted on the Python Orange program. In the second stage of the thesis, firstly, in the Python Orange program, the information of the unit cell's pin radius, pin height, period and gap height parameters are set as the input of the system with this data set machine learning algorithms, Mode-1 and Mode-2 frequencies are taken as the output of the system. In other words, a regression study was conducted between the size parameters of the system and the information of the output parameters. Thus, the stop band frequencies of a unit cell that provides these dimensions of the algorithms in which random size information is entered have been found. In this study, predictions were made with Random Forest, Tree, Gradient Boosting, AdaBoost, kNN, Neural Network and SVM algorithms and regression studies. It was seen that Gradient Boost algorithm for Mode-1 frequency and AdaBoost algorithm for Mode-2 frequency gave the best results. Then, Mod-1 and Mod-2 frequency information was determined as the input of the system, and the pin radius, pin height, period and gap height parameters that make up the unit cell were configured as the output of the system. Thus, a regression study was carried out between the stopband frequency information (Mode-1 and Mode-2 frequencies), which is the input of the system, and the dimension parameters, which are the output. In this way, the size parameters of the unit cell that can best provide the Mode-1 and Mode-2 frequencies randomly entered into the system have been found. In this study, regression studies were performed with Random Forest, Tree, Gradient Boosting, AdaBoost, kNN, Neural Network, SVM, Stochastic Gradient Descent and Linear Regression algorithms. At the end of this study, Gradient Boost gave the best results for pin height and gap height, and Random Forest algorithm gave the best results for pin radius and pin period. In both studies, \%75 of the data set produced by parametric sweep was used to train the algorithms and the remaining \%25 was used to test the algorithms. Then, the algorithm optimization process was carried out by changing the learning rate, method and tree values of the algorithms. First, the kNN algorithm was started by changing the metric system, and then the optimization was made by changing the neighborhood numbers. The effects of these changes were observed by changing the number of trees, learning rate values and method in the Gradient Boosting algorithm. Optimization was made by changing the learning rate and tree values in the Adaboost algorithm, and the number of nodes and layers in the neural network algorithm. Regarding the validation of these studies, firstly, it started with working with dimension parameters as input and Mode-1 and Mode-2 frequencies as output. Then, gap height, pin radius, height and period parameters were obtained as the output of Mod-1 and Mod-2 frequencies as random inputs to the machine learning system. The results obtained in both studies were analyzed in CST Studio Suite and it was observed that the results obtained were similar to the results obtained from machine learning algorithms. In the last stage of Thesis, analysis was made with the full model in the CST Studio Suite program. The data obtained as a result of the analysis supports the study.