Neural networks for microwave characterization of material samples in rectangular cavities

In order to characterize material samples of different sizes at microwaves, use was made of a rectangular metallic cavity, which is partially filled by these material samples, having arbitrary locations. Subsequently, by using a multi-layer perceptron (MLP) network, its dielectric constant, dielectric losses and its amount could be accurately extracted from measurements of the magnitude of the scattering parameter |S₁₁| only, since the phase information was not available. The input for this network is generated by a proper preprocessing of the simulated and measured magnitude of the return loss |S₁₁|. The investigations indicate very good agreement of the simulated and measured data, thus, a simulation-based training of neural networks and an subsequent parameter extraction of the material samples from |S₁₁|-measurements was possible with high accuracy.