STROBOSCOPIC DATA-DRIVEN, INTEGRATED, AND INTELLIGENT MACHINE LEARNING-BASED ALGORITHMS FOR SEMICONDUCTOR WAFER INSPECTION

In the manufacturing process, it is important to improve the yield rate, which significantly affects manufacturing costs. Especially, even a very minor flaw can ruin the entire production of semiconductors on a wafer. The semiconductor industry treats an inspection process significantly; however, the inspection process mostly depends on human inspectors. Therefore, it is a time-consuming process and affected by human factors that cause serious challenges during the inspection process. Moreover, skilled laborers are required to ensure the integrity of inspections. To achieve an effective inspection method for reducing cost and improving yield rate, this paper introduces a data-driven, integrated, and intelligent framework that enables an inspection of particle defects on a semiconductor wafer surface and the location of the particles by machine learning-based algorithms. There are two processes in the framework; scanning a semiconductor wafer surface to generate raw image sets and investigating images to predict their status with or without defects. The stroboscopic imaging technique was utilized to in-line scan the wafer while the wafer is rotating on the spindle. Thus, instead of taking static images one by one, sequence of the wafer surface images can be obtained while the wafer is rotating which can reduce time for taking images. Once the surface images of a wafer are collected, in the second stage, machine-learning techniques are utilized to investigate the status of a semiconductor wafer. There are two subdivided steps in this stage; determining whether a semiconductor wafer has a defect and locating the position of a defect (a particle). As images by visible light from a microscope are used which has low contrast, it is not easy to distinguish the features of defects from backgrounds. Hence, image-processing techniques are applied to emphasize defects in images. Moreover, the number of images for training is amplified by augmenting images with varying contrast and rotating. For the first step among two subdivided steps, an image sets have two categories (with particles and without particles) are used to train, validate, and predict. CNN-based algorithms are applied to classify images with particles and without particles. The next step also has two categories (with particles and without particles) but cropped images are used in this step. Images are selected and cropped by computer vision, and machine-learning model is developed by using CNN-based algorithms to find a particle and locate its position on a wafer. The prediction is performed by striding a full-size image, and the location of a particle is saved and printed when the prediction determines that there is a particle. Thus, by implementing a data-driven, integrated, and intelligent framework machine learning-based algorithm on the inspection of a semiconductor wafer, it is expected that these studies will contribute to developing rapid, efficient, robust, and accurate semiconductor wafer inspection methods.