Accurate Material Characterization of Wideband RF Signals via Registration-based Curve Fitting Model using Microstrip Transmission Line

Knowledge of the electrical properties, such as complex permittivity, permeability and loss tangent measurements is rapidly becoming a necessity for Nondestructive Evaluation (NDE) based material characterization. In this paper, we aim to provide a data- driven approach to estimate the wideband dielectric permittivity for a given substrate material based on the frequency responses from microstrip transmission lines fabricated with the material. We demonstrate registration aided machine learning models that adaptively use information from large simulated datasets to make improved predictions on experimental data where we have acute data scarcity. Machine learning (ML) models are trained using simulation data for several unique combinations of substrate and microstrip line dimensions and is tested on experimental data where the microstrip line are fabricated on eleven different unknown substrates. The S parameters associated with the reflection and transmission coefficients are treated as functional data across the frequency sweeps. As we had very few experimental data, along with complex non-parametric methods, we also consider low-complexity models on the frequency curves. In this aspect, dimensionality reduction techniques are considered to deal with situations in the experimental data where the number of features obtained from the frequency sweeps are much higher than the number of samples in the experimental data. We compare the efficacy of data-hungry machine learning methods with these low-complexity models. As the source of train and test data are different, registration strategies based on intercept correction are implemented. We illustrate the efficacy of registration-based varied ML techniques for lab generated experimental data and obtained encouraging results. This work is an attempt to bypass material characterization models of electromagnetic (EM) physics that is based on closed form mathematical equations and have the limitations that they can only be applied in idealized set-ups.