Research On ANN Model Of GaN HEMT Small Signal Equivalent Circuit Based On GWO

Gallium nitride（Ga N）high electron mobile transistor（HEMT）is widely used in circuit design for high frequency operation due to its wide band gap and high electron mobility characteristics.A high-precision scattering（S）parameter model is required for efficient circuit design.An Artificial neural network（ANN）method based on small signal equivalent circuit knowledge can be used to obtain a high-precision S-parameter model,which has the capability of predicting S-parameters under unmeasured bias voltage.In view of this,this thesis mainly studies the following contents:（1）A knowledge-based ANN modeling method is proposed by using the empirical model S_emp to represent the S-parameters of the small signal equivalent circuit under multi-bias conditions.Based on the commonly used multilayer perceptron（MLP）,to improve the accuracy and generalization ability of ANN model by optimizing the output of MLP with extreme learning machine（ELM）,the MLP-ELM structure is proposed.In this thesis,the number of neural layers and neurons in MLP is optimized by taking the index F as the objective function based on the characteristics of few hyper-parameters and strong robustness of gray wolf optimization（GWO）.Utilizing the fast learning advantages of ELM,the network structure of ELM is clarified based on the mean square error（MSE）of dataset,thereby ensuring the accuracy requirements of the MLP-ELM structure.The proposed modeling framework provides a theoretical support for the accurate modeling of ANN.（2）An accurate ANN empirical model S_emp is established.Firstly,the performance of the GWO is improved by adding an external iteration method,so that a small signal equivalent circuit modeling method of the Ga N HEMT based on the GWO is established,and the model error caused by the capacitance empirical constraint of the complex circuit topology and the frequency dependence of the intrinsic parameters is effectively reduced.The well agreement between the simulated S parameters and the experimental S parameters verifies the accuracy of the training data and the test data needed by the empirical model,which lays a foundation for the accurate modeling of the model.Then,the ANN empirical model is established according to the MLP network structure optimization method based on GWO and the network structure determination method based on ELM.Finally,the accuracy of the empirical model and the validity and practicability of the MLP-ELM structure are verified by the comparative analysis of the relative error of S parameters and the MSE of the dataset.（3）Based on ANN experience model,a knowledge-based ANN model is established.The well agreement between the simulated S-parameters and the experimental S-parameters shows the accuracy of the model and verifies the ability of the model to accurately predict the S-parameters without measured bias voltage.This thesis compares the ANN modeling method based on single-bias equivalent circuit and the pure-ANN modeling method.Through comparative analysis of the relative error of S parameters and the dataset MSE,it is concluded that the method proposed in this thesis is more accurate in modeling Ga N HEMT scattering parameters.In addition,knowledge-based ANN modeling methods are more accurate than pure ANN modeling methods,and the increase in information contained in knowledge can improve the accuracy of knowledge-based ANN models.