Research On Performance Prediction Of SOI LDMOS With Termination Structure Based On Deep Neural Network

With the continuous development of social economy,the structural complexity of power devices is increasing,and various terminal technologies are introduced to optimize and improve device performance.However,this has also led to increasing difficulty in evaluating device performance.The current performance evaluation mainly relies on simulation tools with technology computer aided design（TCAD）as the core,but it has problems such as long simulation time,high computing power consumption,and poor convergence of simulation,especially for devices with terminal technology.The high-speed iteration of power device technology also has higher requirements for device design efficiency.Therefore,how to achieve efficient performance prediction and structure design of devices has become an urgent problem.In recent years,due to the explosive development of machine learning technology,device performance prediction and automated design by machine learning algorithms have received extensive attention and research.In this paper,the breakdown voltage（BV）and specific on-resistance(R_on,sp)prediction of SOI LDMOS with different termination techniques and the optimal drift region doping concentration design for REduced SURface Field（RESURF）SOI LDMOS are investigated using deep neural network（DNN）algorithm.The main work is as follows:（1）A DNN-based method for BV and R_on,sp of SOI LDMOS with field plate is proposed.The BV and R_on,sp data of SOI LDMOS with field plate of different structures parameters are collected.Based on the collected data,a two-stage BV prediction model based on breakdown position and a R_on,sp prediction model are established by using the DNN algorithm.The results show that,compared with the traditional TCAD simulation,the prediction accuracy for three breakdown positions is 96.89%,and the prediction errors of BV for three breakdown positions are 4.72%,3.38%,and 7.09%,respectively.The prediction error of R_on,sp is 2.55%.In addition,the BV prediction speed is increased by 1.25×10⁵ times.For the prediction of device performance using termination structures,this method reduces the prediction time while ensuring the prediction accuracy compared with traditional numerical simulation tools.More importantly,it avoids the problem of poor convergence of numerical simulation tools.（2）A DNN-based method for BV and R_on,sp of SOI LDMOS with stepped doping drift region is proposed.The BV and R_on,sp data of SOI LDMOS with stepped doping drift region of different structures parameters are collected.Based on the collected data,the BV and R_on,sp prediction models are established by the DNN algorithm.The results show that,compared with the traditional TCAD simulation,the BV prediction error is 3.75%,the R_on,sp prediction error is 3.65%,and the BV prediction speed is increased by 8.98×10⁴ times.This method effectively improves the performance prediction efficiency of SOI LDMOS with stepped doping drift region.（3）A DNN-based method for the optimal drift region doping concentration and BV at the optimal drift region doping concentration of SOI LDMOS is proposed.The optimal drift region doping concentration and the BV data of SOI LDMOS of different structures parameters are collected.Based on the collected data,the optimal drift region doping concentration and the BV prediction models are established by the DNN algorithm.The results show that the prediction error of optimal drift region doping concentration is 4.01%,and the BV prediction error is 1.56%.In the design of drift region doping concentration of SOI LDMOS,this method can effectively avoid the complicated human design process and enhance the device structure design efficiency.