Deep neural network for retrieving material's permittivity from S-parameters

This research proposed a deep neural network (DNN) model employing a multilayer feedforward artificial neural network to characterize the relative permittivity and loss tangent of a solid sample in a broad frequency range from 1 to 10 GHz. The method exploited a grounded coplanar waveguide as a measurement fixture, and a vast amount of data was obtained from full-wave simulations. The latter was used to train the proposed DNN model. We performed parametric studies to examine optimal DNN hyperparameters and improve the efficiency of the material property retrieval process. The proposed model was validated by retrieving the relative permittivity and loss tangent of a known substrate. The results show good agreement with the known reference values with a slight error of 1.2%.