Research On Small-signal Model Of GaN HEMT Based On Machine Learning Algorithms

With the rapid development of communication technology,the devices based on second-generation semiconductor technology have been gradually unable to meet the application requirements,and the gallium nitride（GaN）heterojunction structure of high electron mobility transistor（HEMT）with high power,high frequency and large breakdown voltage is popular in the market,and is regarded as one of the most promising semiconductor technologies.Therefore,research on modeling of GaN device characteristics has become extremely important.Among many modeling technologies,behavioral model has become a research hotspot due to its low complexity.Machine learning,which is widely used in other modeling fields,has also been applied to device modeling by a large number of scholars and has achieved excellent results.This thesis focuses on three small-signal behavior models of GaN HEMT based on machine learning algorithms,and analyzes and selects appropriate models for quantitative analysis of small-signal kink effect（KEs）.Focusing on the small-signal behavior of GaN devices,this thesis proposed an accurate small-signal model of GaN HEMT based on support vector regression（SVR）algorithm.Compared with the equivalent circuit model（ECM）,the proposed model is simple to extract and has great fitting accuracy.The performance of SVR model is validated using data in the 1GHz-10GHz frequency band of an 8×125μm GaN HEMT device with a gate feature size of 0.25μm.The results show that the SVR model has higher fitting accuracy than ECM,and the error is reduced by more than 60%.Under the generalization conditions,the relative error E22 is controlled within 5.9%,which has good prediction ability.The above research results have been published in International Journal of Numerical Modeling:Electronic Networks Devices and Fields.Aiming at the insufficiency of generalization ability of the small signal behavior model of GaN devices,this thesis proposes a small-signal accurate model of GaN HEMT devices based on the long-short term memory（LSTM）network.By introducing the gate parameters,the ability of the algorithm to deal with long sequence problems is enhanced and better model prediction performance is obtained.Testified on the same device under test（DUT）,the relative error E22 of LSTM model under generalization conditions is controlled within 4.7%,and the accuracy is improved by 20%compared with SVR model,which can well fit the behavior trend of S-parameter.These findings have presented at 2022 International Conference on Microwave and Millimeter Wave Technology（ICMMT）.Aiming at the quantitative analysis of the small signal kink effect of GaN device,this thesis proposes a precise model of GaN HEMT devices based on gate recurrent unit（GRU）network.The small-signal kink effect in the 0.5GHz-65GHz frequency band of a 0.25μm GaN HEMT with a 1.5mm large gate periphery on a Si C substrate is analyzed.Compared with the LSTM model,the training parameters of the GRU model are reduced by 25%,the accuracy is improved by more than 40%under the generalization conditions,the relative error E22 is controlled within 2.7%,proving its generalization performance is further improved.The GRU model also shows excellent performance in the modeling of h21,and the mean square error（MSE）under all test conditions remained on the level of 1×10^-5.According to the quantitative analysis based on kink parameters,all kinks change significantly after the removal of parasitic parameters,and the kink in S22 becomes more obvious,while the two kinks in h21 disappeare directly.Kink in S22 decrease with the increase of temperature,but kinks in h21 are insensitive to temperature.In addition,the size and shape of kinks change significantly with the change of the bias voltage.The above research results have been submitted to IEEE Transactions on Computer-Aided Design of Intergrated Circuits and Systems and applied for national invention patent.In conclusion,this thesis developed modeling of the small-signal behavior under various machine learning algorithms for GaN HEMT,and used the model to analyze the small-signal kink effect in detail,and determines the influence of various operating conditions on kink,which has positive significance to the design of the device.