Comparison of Machine Learning Regression Models for Coupled Microstrip Line Parameter Synthesis

This paper presents the development of a machine learning model for the synthesis of key geometric and electrophysical parameters of coupled microstrip transmission lines. It presents the results of a comparative analysis of the most common machine learning models for regression tasks, including linear and polynomial regression models, decision trees, random forests, and neural networks. Model training for comparative analysis was performed using both central and graphics processing units. The analysis showed that the use of a graphics accelerator significantly reduced training time only for the neural network model. The values of the elements of the per-unit-length parameter matrices of the coupled microstrip line are fed to the input of the models, and the output synthesizes key geometric and electrophysical parameters. Model verification was performed on test data not used in training, with an analysis of errors both within and outside the training range. A comparison of polynomial regression and neural network models with Wheeler’s analytical expression is also presented. Results showed that the polynomial regression and neural network models exhibit the lowest error compared to the true data.