Research On Automatic Detection Technology Of Wafer Surface Defects

In the process of semiconductor production,the wafer processing process is complex.In order to ensure the quality,almost every step of the process needs to be tested.Otherwise,the defective products flow into the next process,which not only has a great impact on the overall quality of the product,but also causes a huge waste of production costs.In wafer processing,many defects will appear on the surface of the wafer.In view of these surface defects,rapid automatic detection equipment is almost monopolized by foreign enterprises,and there are still many manual detection methods that are still inefficient in domestic semiconductor enterprises.Therefore,it is necessary to independently realize rapid and accurate automatic detection equipment for wafers.In this paper,the automatic detection technology of wafer surface defects is studied,and the visual detection scheme of wafer surface defects is designed based on the characteristics of the wafer in the project.In order to achieve high-speed automatic detection,the rapid acquisition of clear image of wafer surface and fast and accurate image defect detection algorithm are mainly studied and designed.First of all,due to the fine defects of wafer detection,it is necessary to use the micro-optical system for image acquisition.The depth of field of the imaging system is small,so in the high-speed visual inspection system,in order to ensure the clear image of the collected wafer surface,the automatic focusing system of the flying beat was designed.Based on the rapid acquisition of the wafer surface image by the flying beat,the real-time automatic focusing method is used to ensure the clarity of the image.In this paper,an automatic focusing system based on ranging method was designed.By planning the flying path of the wafer surface,the ranging focusing of each shooting point can be completed in advance in the flying process to ensure that the image is clear.The range-focusing system uses the spectral confocal sensor as the ranging equipment and the image processing algorithm to calibrate the range-focusing benchmark automatically.The deviation between focus results and artificial focus is less than 1um,which can obtain clear images.Secondly,due to the complex defect characteristics of wafer surface to be detected,in order to quickly and accurately detect the defect detection of wafer surface in the high-speed and high-precision detection system,a visual defect detection algorithm based on deep learning is designed.Firstly,the GDS file information of wafer processing is proposed,and the collected wafer image is first segmented and extracted to shield the interference of the background information of the wafer,reduce the amount of image data to be processed,and improve the defect detection speed,and this method has faster speed than the traditional grain extraction algorithm.Secondly,the grain image is divided into four parts,and the proportion of grain defects in the image after cutting is 4 times higher than that before cutting,which improves the accuracy of deep learning defect detection.Finally,YOLOX-Darknet53 is determined as a neural network for deep learning defect detection.The speed and accuracy of the traditional target detection network and the new YOLOX-Darknet53 network are tested by the wafer defect dataset.The superiority of YOLOX-Darknet53 network in speed and accuracy is determined,which can realize fast and accurate wafer surface defect detection.Finally,the automatic detection software for wafer surface defects was developed,and automatic detection experiments were carried out on some wafer products.The results show that the time for the system to complete the detection of an eight-inch wafer defect is 107 minutes.The clarity of the obtained wafer surface image can reach99.9 %,and the recall rate of surface defect grains is not less than 95 %,which meets the needs of factory production.