Precise Modeling of Active Component in Ka-Band Using Deep Neural Network Trained with S21 Data

This paper proposes a method to precisely determine the R-L-C characteristics of active components, namely PIN diode and varactor diode, with an in-house developed deep neural network (DNN) approach. We use a microstrip transmission line as the measurement fixture, which also contains a DC biasing circuit to activate the component under measurement. The above structures are modeled and simulated in a full-wave electromagnetic simulator (HFSS). A through-reflect-line (TRL) calibration is applied to obtain accurate responses of the device under test (DUT), e.g., PIN diode. A vast amount of simulated transmission coefficient (S21) data of the DUT is generated to train the DNN model, and an optimal model is considered to have a 95% testing accuracy. From this optimal model, we proceed with the testing data from simulation and obtain its R-L-C characteristics. A good agreement was obtained between the properties of the configured DUT, in the simulation, and predicted results, by the DNN model, in a broad frequency range from 24 to 40 GHz (Ka-Band).