Permittivity Retrieval Based on Fully Connected Neural Networks

In order to accurately and quickly extract the dielectric constant of materials, this paper designs CST software to simulate the scattering parameters corresponding to different dielectric constants under the frequency of $1 - 18\text{GHz}$ based on the traditional transmission/reflection method. These scattering parameters are used as the data source of deep learning, and the amplitude and phase of S11 and S21 are used as the input. Permittivity is used as output to construct a fully connected neural network model. Compared with linear regression, Knearest neighbor regression and random forest algorithm, the final experimental results show that the fully connected neural network inversion has certain correctness and effectiveness, and the inversion effect is the best.