Broadband Permittivity Characterization of a Substrate Material Using Deep Neural Network Trained with Full-Wave Simulations

It is crucial to know the permittivity of dielectric materials used in radio frequency (RF) components and devices because their operation frequency and loss characteristics are significantly affected by the permittivity. In this study, we propose a permittivity characterization technique based on a deep neural network (DNN). The latter was trained using data obtained from full-wave electromagnetic simulation software. With the DNN trained with more than 95% testing accuracy, the measured complex transmission coefficient of the material under test (MUT) was assigned as an input to the DNN model, and the complex permittivity of the MUT was retrieved at the output. The proposed technique was validated by measuring FR-4 epoxy resin substrates of different thicknesses. The results obtained with the DNN model showed good agreement with each other, with an error of less than 1.2% for the relative permittivity value over a broad frequency range of 1 - 10 GHz. We also compared the results with those obtained from a conventional permittivity characterization technique based on analytical solutions to highlight the effectiveness of the proposed method.