Wafer Region Yield Prediction: Employing Majority Under-Sampling and Output Binarization on Low-Yield Threshold

The Wafer Acceptance Test (WAT) is a vital process that measures the manufacturing quality of the wafer using a series of parametric measurements to ensure that the devices manufactured are within design specification. The current framework when using WAT in Yield Prediction (YP) is to use the average value of all WAT measured on the wafer to predict the average wafer yield. This framework causes significant loss of information and does not provide the correct yield information that is more often used in practice, which is the wafer yield loss pattern. Knowledge on the wafer yield loss pattern will provide important insights such as possible process issues and the yield loss nature. To address this, a wafer region yield prediction model is proposed by using the WAT site data found within each region to predict whether the region is high or low yield. Each wafer region WAT site data will be pre-processed with majority under-sampling and the yield of each site is determined based on the shortest distance between the die to the site location. To change it to a classification problem, the yield output for each site will be binarized by finding the low yield threshold which is determined by domain experts using k-means clustering with gap statistic to find the optimal number of clusters for each site. Each region model will be trained using an eXtreme Gradient Boost (XGB) model and Support Vector Machine (SVM) using f-beta scoring to determine the best prediction threshold across different precision emphasis. Each of the five region’s prediction model can achieve a higher precision score while achieving a relatively high f-beta score.