Neural network based system for RF testing of thin samples

A novel artificial neural network (ANN) based architecture is proposed for the complex permittivity determination of thin samples in RF and microwave frequency band. The proposed approach uses a coplanar waveguide sensor for the measurement of scattering coefficients of test specimens, which are then used in conjunction with the proposed ANN architecture to obtain their dielectric properties. The data for training the ANN are obtained by simulating the coplanar sensor using the CST Microwave Studio. To train the network, the gradient-descent training function with momentum is used which avoids sticking to a local minima. In addition, the regularized cross-entropy activation function is used for faster learning of the network and to address over-fitting. The proposed approach is validated by testing a number of standard samples in the designated frequency bands.