Practical aspects of complex permittivity reconstruction with neural-network-controlled FDTD modeling of a two-port fixture.

The paper discusses characteristics of a new modeling-based technique for determining dielectric properties of materials. Complex permittivity is found with an optimization algorithm designed to match complex S-parameters obtained from measurements and from 3D FDTD simulation. The method is developed on a two-port (waveguide-type) fixture and deals with complex reflection and transmission characteristics at the frequency of interest. A computational part is constructed as an inverse-RBF-network-based procedure that reconstructs dielectric constant and the loss factor of the sample from the FDTD modeling data sets and the measured reflection and transmission coefficients. As such, it is applicable to samples and cavities of arbitrary configurations provided that the geometry of the experimental setup is adequately represented by the FDTD model. The practical implementation of the method considered in this paper is a section of a WR975 waveguide containing a sample of a liquid in a cylindrical cutout of a rectangular Teflon cup. The method is run in two stages and employs two databases--first, built for a sparse grid on the complex permittivity plane, in order to locate a domain with an anticipated solution and, second, made as a denser grid covering the determined domain, for finding an exact location of the complex permittivity point. Numerical tests demonstrate that the computational part of the method is highly accurate even when the modeling data is represented by relatively small data sets. When working with reflection and transmission coefficients measured in an actual experimental fixture and reconstructing a low dielectric constant and the loss factor the technique may be less accurate. It is shown that the employed neural network is capable of finding complex permittivity of the sample when experimental data on the reflection and transmission coefficients are numerically dispersive (noise-contaminated). A special modeling test is proposed for validating the results; it confirms that the values of complex permittivity for several liquids (including salt water acetone and three types of alcohol) at 915 MHz are reconstructed with satisfactory accuracy.