Optimizatio N And Performance Prediction Of TFET Based On Neural Network

With the scaling-down of metal oxide-semiconductor-field effect transistor(MOSFET),the short channel effects(SCEs)becomes more and more serious,which leads to the sharp increase of the transistor leakage current in off state and the increase of the power consumption of the device.In addition,the thermal electron emission mechanism makes the sub-threshold swing of the transistor difficult to break through the limit of 60m V/dec,which makes it difficult to further reduce the supply voltage and dissipated power.The tunnel field-effect transistor(TFET)based on the band to band tunneling mechanism has a lower OFF-state current and a subthreshold swing of less than 60m V/dec,which can overcome the influence of SCEs effectively,so the power consumption of the device can be effectively reduced.However,the major drawbacks of TFETs are its low ON-state current(I_on)and ambipolar conduction(I_amb),so further research on TFETs is still needed.In order to meet the needs of the rapid development of TFET,this paper proposes a novel method of using neural network to predict and optimize the performance of existing TFETs to speed up TFETs research.The main research results of this paper are as follows:(1)Firstly,feature vectors were determined and training data were collected for Gate-All-Around TFET(GAA-TFET),and then Convolutional Neural Networks(CNN)with forward design and reverse design were built respectively.According to the error between the network prediction results and the real value,the prediction method based on the neural network is more convenient and efficient than the traditional numerical simulation method based on the empirical trial-and-error method,and the prediction accuracy is higher than the machine learning method.Therefore,the performance of the device can be optimized and predicted by adjusting the forward design or reverse design network input to predict whether the corresponding device performance or structure meets the design goal.(2)In this paper,a Ge_0.9Si_0.1/Si heterojunction double-gate TFET with T-channel heterogate dielectric(HJ-HGD-DGTFET)structure is presented.The Ge_0.9Si_0.1/Si heterojunction and Pocket layer can not only increase the conduction current of the device,but also improve the subthreshold characteristics of the device.The simulation results show that the ON-state(V_D=V_G=1V)current reaches 9.59×10^-5A/?m,the I_on/I_off reaches 10¹²,and the average subthreshold swing is 18m V/dec.In addition,hetero-gate-dielectric(HGD)is used to inhibs the bipolar current,and the ambipolar current is only 1.531×10^-17A/?m.It shows that the structure not only has a higher I_on,but also can effectively alleviate ambipolar behavior.(3)The proposed optimization and performance prediction method of TFET based on neural network is applied to the proposed HJ-HGD-DGTFET.The Mean Relative Errors(MRE)of the forward design neural network for the prediction of I_on/off,I_amb,V_th and SS_avgare 0.0149,0.0052,0.0379 and 0.0806,respectively.The MRE of the reverse design neural network for the prediction of Si O₂ thickness,source dopant concentration and germanium mole fraction are 0.096,0.0048 and 0.095,respectively.The prediction accuracy of the two methods is higher than that of the typical machine learning methods.By connecting the theory of convolutional neural network with the performance prediction and optimization in the design of TFETs,this paper creatively proposes the optimization and performance prediction method of TFET based on neural network.The method can optimize the device structure and predict the device performance more efficiently,quickly and automatically,which also explores a new method for the design of micro-electronic devices.