Good Die Prediction Modelling from Limited Test Items

This paper proposes a test cost reduction method using machine learning techniques. The proposed method tries to predict good dies among the manufactured dies on the way of test process. If a die is predicted as good before completing all of the test process, the die will be allowed to be shipped without going through the remaining test process which contains costly burn-in test and final test. By a SVM-based procedure together with K-fold cross validation, a prediction model to judge certainly good dies is created from known results of the selected test items. In order to evaluate the method in terms of the business effectiveness, we also propose new evaluation measures, "cost reduction rate" and "bad die escape rate", which enable to confirm zero-defect oriented test cost reduction. Experimental results obtained through test data for industrial dies requiring zero-defect show that the proposed method has significant predictability with high test cost reduction capability.