由于GaN HEMT的寄生电容在漏源电压的影响下具有非线性特征。本文提出一种于模拟寄生电容参数与栅源电压的映射关系的反向传播神经网络（BPNN）参数拟合法。基于GaN HEMT大信号参数模型和数据手册提取寄生电容关于漏源电压变化的参数取样点作为神经网络模型的输入层。该神经网络模型通过将误差值反馈至输出端以调整神经元的权重，以产生一个能够拟合出系统映射关系的神经网络训练集，以预测GaN HEMT实际开启状态下栅源电压和寄生电容的关系，并将所得实验数据连接至输入层进行传递。最后得到BPNN模型的预测相对误差分别为9%，证明了该方法模型预测的准确性。该方法简化了物理模型对参数提取的复杂度，可用于分析和预测GaN HEMT器件的误导通情况。

This article proposes a backpropagation neural network(BPNN) parameter fitting method to simulate the mapping relationship between the parasitic capacitance parameters and the gate source voltage of GaN HEMT due to its nonlinear characteristics under the influence of drain source voltage. The input layer of the neural network model is based on the GaN HEMT large signal parameter model and data manual, which extracts the parameter sampling points of parasitic capacitance related to the variation of leakage source voltage. This neural network model adjusts the weights of neurons by supplying error values to the output end to create a neural network training set that can adapt to the system mapping relationship, predict the connection between gate source voltage and parasitic capacitance in the actual open state of GaN HEMT, and link the acquired experimental data to the input layer for transmission. The BPNN model yielded a relative prediction error of 9%, demonstrating the efficacy of this method in predicting the behaviors of GaN HEMT devices. This technique simplifies the complexity of parameter extraction from physical models, thus facilitating the analysis and prediction of misleading situations.

TN710

国家自然科学基金(No.61974056); 江苏省重点研发社会发展项目(No. BE2020756)