(A) Bayesian deep learning method for predicting highly imbalanced package test results

Abstract

In semiconductor manufacturing, the package test is a process that verifies whether the product specifications are satisfied before finally shipping the products to the customers, classifying the goods or the defects according to the verification results. Since the package test results and customer quality are closely related, it is important to strictly identify and manage defects. In this paper, we propose a method to predict in advance the potential defects using the data of wafer test results which directly related to the package test results. In our method, there are several challenges. First, the package test data is class-imbalanced with a very low defect rate, and the imbalance level may change due to variance of manufacturing processes. Second, there is a complex relationship between package test results and wafer test results. Third, it is more important to increase the detection accuracy for defects, instead of the overall classification accuracy. More specifically, we propose a Bayesian neural network-based model that can adaptively deal with unknown imbalance levels through quantifying uncertainty and flexible adjustment of decision boundary. Using a real semiconductor manufacturing dataset from a global semiconductor company, we show that the proposed model can effectively identify defects even under various imbalance levels.