Data Additive Residual Learning based Yield Prediction for Semiconductor Manufacturing

In semiconductor manufacturing, yield prediction and yield-related factor detection are critical. Various process data are analyzed to predict the yield and discover the causes of yield lowering and enhancement. Based on the growing field of artificial intelligence, studies have been carried out to improve the efficiency and accuracy of yield analysis by applying machine learning. In empirical eases, engineers continue to add suitable data for the problem, so we suggest modeling that can add the appropriate data sequentially. Based on the ideas of the boosting algorithm, the added data is used to estimate the residual in the existing model. The proposed method is applied to actual industrial data to demonstrate the practical application and improvements in yield prediction. Various optimization procedures are conducted in the modeling process, including those for the base model, the shrink rate, the number of iterations, and the order of dataset addition. The best model of three rounds with a shrink rate of 0.9 showed the minimum mean absolute error value of 0.497. The proposed approach improves prediction accuracy and various attempts for optimization will be the cornerstone of data additive residual learning.