Calculation of per-unit-length parameters of multiconductor microstrip transmission lines using machine learning models

The calculation of the per-unit-length RLGC parameters of microstrip transmission lines is an integral part of the analysis and design of microwave devices. For coupled multi-conductor structures, this task requires accounting for mutual electromagnetic interactions and is typically solved using numerical electromagnetic methods, which provide high accuracy. However, the significant computational cost of such methods substantially limits their use in repeated calculations, which are typical for automated design, parametric analysis, and optimization tasks. With the increasing demands for speed in modern computer-aided design systems, developing alternative approaches that significantly reduce computation time while maintaining the required accuracy has become an urgent task. One promising direction is the application of machine learning technologies to build surrogate models, enabling rapid calculation of transmission line parameters based on data obtained from numerical simulations or measurements. The objective of this work is to investigate the applicability of machine learning methods for the accelerated computation of per-unit-length RLGC matrices of single and coupled microstrip transmission lines. The study considers the use of polynomial regression, decision tree, random forest, and artificial neural network models for predicting RLGC matrices. Reference data were obtained via electrostatic modeling using the method of moments (MoM) for 96,000 line configurations at a frequency of 1 GHz. It was shown that the artificial neural network model provides the best performance, achieving a mean relative error of 1,06% compared to numerical simulations. Additionally, the predicted RLGC parameters were validated by analyzing the transient response of transmission lines to an impulse excitation, confirming the physical consistency of the obtained results. It is demonstrated that using the neural network model allows a reduction in computation time by at least a factor of three compared to direct simulations, while maintaining acceptable accuracy, which confirms the potential of machine learning methods for modern computer-aided design systems.