半导体封装过程中晶圆缓冲区缺陷检测

Abstract 

Efectively detecting defects in wafer bufer zones is crucial in semiconductor manufacturing processes. If defects are not  detected in the middle of the semiconductor manufacturing process, the semiconductor eventually becomes a defective  product after all processes are completed. To solve this problem, rule-based vision algorithms have been used to identify  defects in the wafer bufer zone. After photographing the wafer bufer zone using a high-speed camera, defects are detected  using the pixel values of the images. However, because of the resin bleed in the wafer, which is an epoxy compound, it is  difcult to detect defects. Therefore, we introduced a deep learning method for semiconductor inspection and created a  high-performance semiconductor inspection algorithm. The defects in the wafer bufer zone should be detected accurately  and quickly. Furthermore, the approximate size of the defects must be extracted. We modifed the Xception model to ft the  wafer data characteristics considering both accuracy and speed. We proposed to extract the size of the defect using class  activation mapping (CAM). We obtained an accuracy of 96.9% from the actual wafer dataset through this framework, and  then managed to extract the size of the defect through CAM.

Introduction 

Semiconductor wafers undergo various packaging processes and unwanted defects must be accurately and quickly  detected during each process. However, epoxy molding compounds (EMC) used to protect semiconductor devices from  external impact, vibration, moisture, and radiation as well as  the protection cover in the bottom of wafers make the visual  and in-process detection of defects difcult (Fig. 1). While  defects in wafers can be checked after all processes are completed, the manufacturing loss due to reduced yield becomes  non-negligible. Fortunately, there is a narrow region not covered with EMC mold called the wafer bufer zone which can  be observed during the process. Hence, real-time inspection  of this zone for detecting wafer defects during the process  is important.

Fig. 1 Schematic representation of inspection area

Images obtained from the inspection of the wafer bufer  zone can be classifed into four types: Normal, Crack, EMC defect, and Notch (Fig. 2). Notch (Fig. 2a) is the  reference point of the wafer and hence is not a defect. We  would like to detect cracks and EMC defects when inspecting the wafer bufer zone during the packaging processes.  It is relatively easy to detect cracks when they exist on a  clean surface (Fig. 2b). However, the detection becomes  difcult when cracks appear on a surface with resin bleeding (Fig. 2c), especially if they are relatively small. EMC  defects occur when they are not properly cover the mold  area where the semiconductor die is located (Fig. 2d, e).  If EMC excessively invades the bufer zone, the protecting cover can stick to it and cause damages to the wafer. If  EMC in the mold area is insufcient, the semiconductor  die cannot be fully protected. These EMC defects are not easy to be detected as they can be confused with notch or  resin bleeding, which is normal.

Fig. 2 Types of images obtained during inspection

For the detection of these wafer defect, rule-based computer vision algorithms can be used. For example, the Sobel  operator can be used to detect edges of images and identify  defects. However, it is difcult to extract cracks using  this operator if the wafer bufer zone is covered with resin  bleeding because too many edges are extracted. Another  popular approach is the Canny algorithm which extracts the  meaningful edges by thresholding after denoising images  using Gaussian flters. While this algorithm generally  provides better results than the Sobel operator, it sufers  from the same difculty in distinguishing cracks from resin  bleeding.

Recently, various deep learning models that use high level  features have been developed in computer vision. For example, object detection models recognize an object from an  image by indicating its location in a bounding box.  Semantic segmentation models classify objects more densely  than the detection model using the pixels within an image. While we can know the location and size of defects  with these models, they require bounding-box annotations or  segmented labels for training. Also, these models are heavy  in terms of the number of model parameters and require a  long time for training and inference. Since the detection of  wafer defects needs to be done quickly during the packaging  processes, these models cannot be used directly in practice.

Alternatively, we can use the classifcation model which  classifes images into the defned types. It is lighter  than detection and semantic segmentation models, and  hence, is advantageous in terms of speed. In addition, there  is no need to annotate labels in images when building the  datasets for training. However, this model is unable to localize the defects within an image and measure their characteristic features such as the size and the area.

In this study, we develop a deep learning-based inspection  method for detecting defects in the wafer bufer zone. The  model was designed to fnd and localize a defect quickly,  and infer its size as well (Fig. 3). We employed a classifcation model, Xception, and modifed it to be suitable for inspecting the wafer bufer zone. To accelerate the  inspection, we changed eight repetitions of middle fow of  the original Xception model to one. In addition, the feature  pyramid network (FPN)  was used to efectively handle various sizes of defects. We utilized the class activation  map (CAM)  to generate a heat map of a specifc class  image, and hence, estimate the size of defects without any  additional supervised learning. The length of cracks and  the area of EMC defects can be approximately obtained  while inspecting the wafer bufer zone. The proposed model  showed higher detection accuracy and faster inference speed  than baseline models.

Fig. 3 Overall workfow. The  training fow and inference  fow are represented with black  solid lines and red dashed lines,  respectively

Inspection model consists of the image classifcation for  identifying defects and the estimation of their size (Fig. 4).  The Xception deep learning model was employed for classifcation with modifcation to enhance the efciency in  detecting defects. The class imbalance problem was alleviated by adopting the focal loss. The identifed region of  defects in the wafer bufer zone was inspected with CAM to  estimate the size of cracks and EMC defects.