Research On TSV Defect Detection Method Based On Hybrid Extreme Learning Machine

The emergence of three-dimensional stack integration technology has opened a new path for the development of the semiconductor industry.As a key technology and an important part of 3D stack integration,the reliability problem of Through Silicon Via(TSV)cannot be ignored.TSV is essentially a plurality of conductive copper pillars wrapped in an insulating layer and embedded in different substrates.This technology is widely used in vertical interconnections between 3D integrated circuit chips and chips.However,because TSV itself is prone to structural defects,and the location is uncertain and the types are complex,these small structural defects will have a greater impact on the signal transmission performance of TSV,which seriously affected the yield of 3D integrated circuit products.Therefore,it is necessary to study TSV modeling and defect detection technology.As the physical size of TSV gradually decreases,the manufacturing materials become more diverse,and the spatial layout becomes more and more complicated.It is difficult for traditional TSV inspection methods to accurately detect TSV defects.This paper proposes a defect detection method for TSV based on PSO-ELM hybrid limit learning machine.The main research work is as follows:1.Ground-signal-signal-ground TSV(GSSG-TSV)as the research object,use 3D simulation modeling software HFSS for trouble-free TSV modeling,and use ADS software to build equivalent The circuit analyzes the effectiveness of the fault-free GSSG-TSV.At the same time,the GSSG-TSV void defect model,pinhole defect model and micro-substrate misalignment defect model were established using HFSS.Based on the built defect model,the S-parameters of different defects were extracted,and the analysis of GSSG-TSV signal transmission was analyzed based on different size void defects,different height void defects,different size pinhole defects,different height pinhole defects and micro-substrate misalignment defects impact on performance.2.For the GSSG-TSV void defect,pinhole defect,and micro-substrate misalignment defect,extreme learning machine(ELM),artificial neural network(ANN),K nearest neighbor(KNN),random forest(RF),and support vector machine(SVM)and other common machine learning classification algorithms explore TSV defect detection and analyze the reasons why the above algorithms have different classification effects on sample data.Based on the sample data obtained from the GSSG-TSV model,Particle swarm optimization(PSO)was used to optimize the weights,biases,and optimal nodes of the extreme learning machine.The supervised machine learning method is adopted to usethe optimized hybrid extreme learning machine model.It uses different frequency excitation signals and S parameters to predict the types of defects that occur in TSV,and locates specific TSV defects after determining the type of defects.A new non-destructive TSV defect detection method is proposed,and a hybrid defect learning machine is used to study the composite defect detection.The experimental results show that the accuracy of the non-destructive TSV defect detection method for classification of GSSG-TSV structural defects is 89.62%,the accuracy of defect location is 92.37%,and the accuracy of defect classification for composite defects is 85.41%.3.Signal-Ground TSV(SG-TSV)and Ground-Signal-Ground TSV(GSG-TSV)defect models are established using HFSS.For void defects,pinhole defects,and misaligned micro-substrate defects,the PSO-ELM hybrid limit learning machine model was used to detect the defects above.The experimental results show that the hybrid limit learning machine model has an accuracy rate of 88.47% for SG-TSV structural defects classification,90.25% for defect location accuracy,91.87% for GSG-TSV structural defect classification accuracy,and 92.18% for defect localization accuracy.Through the TSV defect detection of different models,the universality of the new non-destructive TSV defect detection method proposed in this paper in the application of TSV defect detection is verified.