| Electronic design automation(EDA)tools are the critical part of the design of power integrated circuits.As the core component of power integrated circuits,lateral power devices are the key factors that determine the performance,cost,manufacturability,and integration of both power integrated circuits and on-chip power systems.The design of lateral power devices is highly dependent on the combination of EDA tools and human experience,so it is difficult to achieve efficient automatic design.In recent years,due to the explosion of big data and the improvement of computer hardware equipment,intelligent EDA design through machine learning algorithm modeling has received wide attention.Thus,this thesis takes the design and manufacturing method of silicon on insulator(SOI)lateral power device as the research object,conducts the research of the following part,i.e.,the prediction of breakdown performance,the automated design of device structure,and the lithography hotspot detection model during the layout design process.By applying machine learning algorithms to the design and research of SOI lateral power devices,the research aims to provide support for the design and manufacturing technology of power semiconductor devices.The main research contents and contributions of this thesis are presented as follows:1.The establishment of breakdown performance prediction model of SOI lateral power device.(1)The establishment of prediction classification models for the breakdown location of lateral power devices.The important factors affecting the breakdown performance of lateral power devices is analyzed and summarized.Then,with the important factors as input objects,the model for breakdown location prediction is proposed by using deep neural network(DNN)algorithm.This method establishes the classification model through information forward propagation and error backward feedback process of the network framework.The results show that,compared with the TCAD simulation,the accuracy of the breakdown location prediction is as high as 97%.Besides,it can provide the breakdown location simply and efficiently.It shows that the classification model can accurately give the breakdown location,which greatly improves the performance prediction speed.(2)The establishment of prediction regression models for the breakdown voltage(BV)of lateral power devices.The breakdown voltage corresponding to different breakdown locations is analyzed first,and then a two-stage prediction framework,which uses deep neural network or Gaussian process regression(GPR)to train the regression model,is proposed.The results show that the average error between the breakdown voltage prediction and the targets is less than 5%,and the model can output breakdown voltage within 1 second.Moreover,compared with simulation tools,the prediction model based on machine learning can greatly reduce the prediction time and completely avoid potential nonconvergence.2.The automated structure design method of SOI lateral power device.(1)Automatic structure design of power device based on unconstrained conditions.The factors that affect the BV and on-resistance(Ron)of lateral power devices are analyzed first.Then,by using the factors as the input,the BV and Ron prediction models are established.Next,based on the prediction models,a fully automatic structure design method of power device is proposed.Given the design specifications,including BV and Ron,this method can achieve the automated structure design by utilizing the Bayesian optimization framework.The framework can achieve the automatically iteratively design by referring the prior information until the performance meets the target.The results show that designs obtained from our optimization framework fall within 5% range from the desired specifications,which effectively avoids the complicated human design process.Moreover,the proposed method can also greatly improve design efficiency,reduce design time,and thus provide an intelligent approach for efficient device design.(2)Automated structure design of power device based on REduced SURface Field(RESURF)constraints.First,the influence of RESURF effect on the breakdown performance is analyzed,and then the effect is introduced into the automatic optimization process of the device structure.By using the Bayesian optimization with inequality constraints,the automated design method of RESURF power device is proposed.In addition,through considering the practical application scenarios,the design thresholds are introduced to optimize the design result.Moreover,the 3-D analytical models of breakdown voltage and on-resistance are analyzed,and a scheme of using model initialization instead of random initialization is proposed for the initialization of Bayesian optimization.The results show that the optimized designs are all located within the RESURF boundary,which can achieve fully depleted condition when the breakdown occurs and avoid the occurrence of breakdown at the P+N junction.In addition,the addition of the design threshold significantly improves the design quality.Meanwhile,the model initialization scheme also effectively improves the design efficiency.3.Lithography hotspot detection model for power semiconductor device manufacturing.The establishment of lithography hotspot prediction model based on Inception module.The traditional convolutional neural network(CNN)modeling scheme for lithography hotspot detection is analyzed first,and then this thesis proposes a new parallel framework based on the Inception module.For the feature extraction part,the framework uses two parallel Inception modules to replace the traditional stacked convolutional layer.For the feature information classification part,the framework uses global average pooling to replace the traditional fully connected layer.The results show that,compared with the model constructed by the traditional stacked CNN,the proposed model based on the parallel framework can effectively improve the accuracy of lithography hotspot detection and reduce the number of false alarms,achieving more effective lithography hotspot detection. |
PDF Full Text Request