Deep Learning And Terahertz Imaging-based Defect Detection Method For Integrated Circuit

The set of quality management process of semiconductor manufacturing and quality control to guide the operation of the enterprise,mainly includes the quality of design,manufacture,test design,reliability design,incoming material control,production process,the statistical process control,defects reliability verification and final testing,screening,and to achieve the requirements of the product’s high quality.To meet stringent quality standards,statistical process control,defect screening and reliability verification are carried out at each step of production to ensure that the output of each process is within the normal range before proceeding to the next process.In all aspects of the semiconductor manufacturing and packaging process,the medium of the IC packaging will inevitably have defects,and these defects will seriously affect the operating efficiency of semiconductor components and integrated circuits,and directly affect the life and performance of electronic products.Some serious defects can lead to a large number of product recalls,resulting in significant financial losses.Therefore,in the entire IC manufacturing process,it is necessary to perform quality inspection on each process to avoid large-scale defective products.Therefore,improving the defect detection accuracy of semiconductor ic products has important academic value and industrial application value for improving product production efficiency and reducing production costs.However,existing defect detection technology cannot perform non-contact nondestructive testing and detect porosity defects in semiconductor ICs,bottom voids,uneven thickness of wafers,chip substrate cracks,internal circuit contamination,internal bonding wires falling off,and packaging materials.The Terahertz time domain spectral imaging technology can effectively obtain the time domain and frequency domain information of the internal structure of the semiconductor ICs,thereby providing the possibility for the detection of semiconductor IC internal defects.This research is divided into two parts.In the first part,the terahertz(THz)imaging accuracy cannot meet the requirement that the high imaging accuracy of the fine features of semiconductor ICs.Extracting the subtle defects use the related algorithms,including image fusion algorithm,image reconstruction algorithm and image enhancement algorithm.In the second part,it is low detection efficiency of the IC internal packaging defects in in actual application.The high resolution terahertz imaging results combined with deep learning will be used to carry out the key technical research and experimental verification of terahertz imaging methods in IC defect detection.In practical application,it provides an effective method for high quality management and quality control in IC production process.The innovative research results in this thesis are as follows:(1)Research of the image restoration and fusionThis thesis studied the point diffusion function(PSF)of image restoration algorithm.The THz source images of different frequency points are obtained by deconvolution terahertz images with PSF.Aiming at the feature representation deficiency of terahertz multi-source image,this thesis propose the multi-focus image fusion algorithm based on wavelet transform and utlize the different fusion rule to handle the low-frequency component and highfrequency component.The wavelet transform fusion algorithm has a good effect on noise suppression in image fusion task.The image fusion algorithm based on sparse representation model is proposed to solve the problem of dictionary expression ability of traditional sparse representation fusion methods.It effectively solves the problems with noise interference and hidden image feature of the terahertz image fusion problem.Aiming at the situation that the wavelet fusion method causes the loss of image edge information and part of the details and reduces the image quality,and the sparse representation fusion method is easy to cause the image grayscale discontinuity and the details are blurred,this thsis proposes the fusion algorithm based on multiscale transformation and sparse representation.The fusion algorithm utlize the image multi-scale transformation and low-frequency/high-frequency component separation techniques.The experimental results show that the terahertz image fusion algorithm based on multi-scale transformation and sparse representation has great potential in improving the resolution of terahertz semiconductor IC images.(2)Research of the image reconstruction and enhancementAiming at the problems that the blurred details of the original THz image are resulted from the attenuation of high-frequency components in the far-field imaging and the Fraunhofer diffraction,the proposed reconstruction model based on the absorption coefficient is fully considering the Gaussian beam theoretical model.The proposed model solves the image blurring effect caused by the attenuation of highfrequency components and Fraunhofer diffraction.Without deconvolution,the proposed reconstruction model based on the absorption coefficient provides a theoretical method for terahertz image reconstruction in multiple scenes and real-time detection in industrial sites.To characterize more detailed features inside the IC terahertz images,this thesis propose two image enhancement algorithms,including the optimized multi-scale Retinex algorithm(OP-MSR)and double multi-scale latent low-rank representation algorithm(DM-LatLRR).The OP-MSR combines the multiscale Retinex algorithm with the atmospheric scattering model,not only restore the radiance of the image scene and improve the image blur,but also effectively improve the gray dynamic range of the image and enhance the texture characteristics of the image.The DM-LatLRR applies the latent low-rank representation to multiple representation levels to extract multi-scale detail matrices,and applies the multi-scale Gaussian function to each matrix component pre-removal blurry.The experimental results show that the DM-LatLRR method can obtain IC features such as bonding wire breakage,dielectric layer cracks,substrate dielectric layering and silver paste coating range,which overcomes the shortcomings of existing methods and achieves better results than other terahertz image enhancement methods.(3)Research of IC defect detection algorithm based on deep learningIn view of the low efficiency of deep learning applied to defect detection in industrial production,this thesis constructs the dataset of semiconductor IC defect detection and proposes the algorithms of semiconductor IC defect detection based on deep learning in small sample dataset.By using the constructed data set of semiconductor IC defect detection,the parameters of the network model are optimized to achieve a more simplified network structure and parameter scale,which improves the network’s processing ability for small sample data and the detection performance of the model.① This thesis proposes the transfer learning based on MobileNet V2 network.The transfer learning network based on the MobileNet V2 network is proposed,the structure and parameter scale of the network are simplified.The network uses the public dataset ImageNet for pre-training.Combined with the characteristics of the terahertz image,a fine-tuning scheme is adopted to train only the last two convolutional modules and the subsequent full connection layer,and freeze the remaining layers.The experimental results indicate that the network performance based on MobileNet V2 training is superior to other CNN in terms of accuracy and operation speed.However,the training model based on MobileNet V2 algorithm needs a large amount of computing resources,and the network computing speed is low,it cannot be applied to the real-time online detection of industrial defects.②This thesis proposes the lightweight convolutional neural network model(LiCNN).Because the defect characteristics of semiconductor products are very subtle or highly similar,the classic CNN model has many network layers.The classic CNNs are difficult to extract image feature,which the later network layers can easily discard the shallow features extracted by the previous network layers,or the earlier network layers cannot completely extracting defect features.So,the classic CNNs are unable to meet the requirements for IC defect detection accuracy.The parameters of LiCNN are optimized to achieve a more streamlined network structure and parameter scale,which improves the network’s processing capability for small sample data and the detection performance of the model.Experimental results show that the LiCNN achieves better results than other methods in the task of terahertz image semiconductor IC defect detection.The above work integrates terahertz time-domain spectroscopy imaging technology with deep learning algorithm,and proposes the high-precision IC defect detection algorithm.The work provides an effective method for semiconductor IC internal defect detection and product quality control and explores the new path,which solves the problem of internal IC defect on-line real-time detection.