Wafer-Level Die Re-Test Success Prediction Using Machine Learning

For high-volume production, time spent for wafer-level test is a crucial contributor to the overall product cost. At the same time, accurate filtering out of the defective dies at this step is essential to avoid escapes to the later more expensive stages of manufacturing test, e.g. package test, integrated system test, or even test escapes to the customer. The nature of the wafer test setup, in particular the mechanical probing, may imply re-test of the dies failed the test in the first run. This paper proposes a method for applying Machine Learning to efficiently predict whether it is favorable to invest extra time into re-test of a particular failed die. The training relies on supervised learning on a predefined subset of wafers, which include information about the first and re-test runs. The experimental results on actual commercial product data demonstrate efficiency of the method for re-test success prediction and, thus, significant optimization of the manufacturing test time. The demonstrated accuracy of prediction is 78%, subject to selection of suited data features and an efficient neural network model.