Knowledge‐based neural network with Bayesian optimization for efficient nonlinearRFdevice modeling

Abstract This paper proposes a novel approach for modeling the nonlinear behavior of microwave devices, with a particular focus on addressing the challenging problem of modeling compound semiconductor transistors under large‐signal excitation. The approach utilizes a knowledge‐based neural network (KBNN) to construct a frequency‐domain behavioral model of the transistor, which maps incident and scattered waves from a coarse model to a fine model in cases where the coarse model fails to match the behavior of a new device. This neural network‐based method aims to achieve better alignment between the extracted model parameters obtained from the coarse model and the behavior of the fine model. To improve mapping efficiency and accuracy, Bayesian optimization techniques are employed to automatically adjust the hyper parameters in the KBNN within a custom‐defined hyper parameter space. The proposed method uses simulation and measurement of a 0.25 um GaN HEMT device operating at 8 GHz, with VGS at −2 V and VDS at 18 V. Furthermore, the proposed model exhibits good interpolation capability at different input power levels, indicating its broad applicability in the design of high‐speed device models. Specifically, the accuracy of the simulation data and test data reached −53.83 and −51.68 dB, respectively, with the test data used in model training being less than 7.5% of the simulation data.