Tackling Non-linearity in Cavity Perturbation using Machine Learning Approach

Abstract

This paper presents a unique design of a cylindrical cavity to identify the dielectric materials with high resolution using the supervised machine learning algorithm. The mountable design of an aluminum-based cylindrical cavity records a quality factor of more than 9000 and provides an easy assembly of samples to be tested. The linear region of the cavity precisely provides the identification of dielectric samples in the range of 1 to 20 within $\sim$ 99 % of accuracy using the standard cavity formulation. On the other hand, the proposed machine learning approach works effectively in the non-linear region of the cavity and predicts the dielectric properties accurately in the wide range dielectric constant starting from 20-45 with a typical error of 0.35 %. The non-linearity of the cavity output is modeled using the cascade feedforward architecture of Artificial Neural Network (ANN) for multiinput variables extracted from the simulations. The model is trained using a well-known Bayesian regularization algorithm with an adequate number of samples and subsequently tested over a large sample of novel test input. The mean square error of test samples in the range of 10-4 and correlation coefficient (R) near 1 demonstrates the effectiveness of the approach in dielectric testing using the proposed cavity.