An Accurate Neural Network-Based Consistent Gate Charge Model for GaN HEMTs by Refining Intrinsic Capacitances

Abstract

Neural network-based capacitance models are accurate, but some of them are not charge-conservative. In this work, a novel consistent gate charge model for GaN high electron mobility transistors is presented based on neural networks. The equivalent circuit parameters are extracted using the multiobjective gray wolf optimizer-based hybrid method, which improves the accuracy of parameter extraction. To obtain more reliable data sets for accurate neural network-based modeling, the outliers in the extracted intrinsic capacitances are automatically detected and removed using the isolation forest technique. The gate charge is obtained by integrating the capacitances with the voltages at different temperatures. A neural network is used to model the bias- and temperature-dependent gate charges, and the intrinsic capacitance formulation is obtained by taking the partial derivative of the gate charge function with respect to the voltages. The proposed model is charge-conservative and requires no transcapacitances. The large-signal model is implemented in the Advanced Design System and verified by small- and large-signal measurements. Good agreement is obtained between the measurements and simulations.