Research On The Behavior Model Of RF Power Device Based On Machine Learning Algorithms

High electron mobility transistors(HEMTs)using modulation-doped gallium nitride(GaN)heterojunction structures have been rapidly developed in the past decade.GaN based technology is considered as one of the most promising semiconductor technologies for high frequency and high power applications.Accurate modeling of GaN devices is essential for computer-aided circuit design,especially for highefficiency power amplifiers(PAs)and monolithic microwave integrated circuits(MMIC)design.Among all of these modeling methods,behavioral models stand out as a research hotspot due to their high accuracy and low complexity.In recent years,artificial intelligence technology represented by machine learning method has gradually become popular,and powerful machine learning algorithms have been widely and successfully employed in the field of circuit designing.The application of the machine learning based methodologigs to build device behavioral models has been also highly valued,and this thesis is focus on this research area.In the device behavior modeling of this work,the first thing carried out is smallsignal modeling.A 1-W GaN HEMT is selected and the characteristics of the device are analyzed,an 18-component small-signal equivalent circuit model is established,and then a small-signal behavior modeling method based on the support vector regression(SVR)technique is proposed.Compared with the classical equivalent small-signal circuit model,the proposed model which adds an error correction technique based on the SVR technique,maintains high accuracy over a wide frequency range from 1 GHz to 10 GHz.The validation is done by comparing with multi-bias S-parameter data measured on an 8 × 125 ?m GaN HEMT device with a gate feature size of 0.25 ?m,and the error of the proposed model is reduced by more than 50% when compared with the traditional equivalent circuit model for all those four S-parameter at different bias conditions.Then,two different machine learning algorithms are investigated in this thesis,including support vector regression for solving linear models and long short-term memory neural networks for dealing with long sequence problems.In this paper,smallsignal behavior models based on SVR and long short-term memory neural network(LSTM)are developed using the idea of black-box modeling techniques,respectively.The validation is carried on the same devices,and both models provide highly accurate prediction results,they can both control their error lower than 1%,even for the strongly nonlinear S12 and S21 behaviors cases.Compared with the SVR technique,the LSTM network technique shows better modeling performance and significantly shorter model extraction time.In the nonlinear behavior modeling case,the LSTM neural network is applied for the first time to model the large-signal behavior of GaN HEMT devices.The basic principles of the network and the detailed modeling process are introduced,and the optimization algorithm,the Adam method,which suit most for this model,is selected.Simulation and experimental validation results are given for a 10 W GaN HEMT device,for example,with input power range from-20 d Bm to +30d Bm.Both extrapolation and interpolation capability are tested in the same input power range.In the simulation,the average relative error(ARE)is less than 1% when the input power interval reaches 3d B;in experimental test,the ARE is still controlled below 2% when the input power interval even reaches 10 d B;the model also verifies the extrapolation and interpolation capability of the frequency,and the ARE is controlled below 4% when it reaches 5 GHz.The strong prediction capability of the model is verified,and the effectiveness of establishing the behavior model of GaN HEMT devices with LSTM technique is demonstrated.