Optimized Test Error Detection by Probabilistic Retest Recommendation Models

As a result of increasing test costs it becomes more and more attractive to improve the test process using modern machine learning techniques. Consequently, in our paper we focus on an essential part of the test process: efficient test error detection. Our proposed online test error detection approach is based on a novel die probability model, which is able to classify die runs as correct or erroneous solely based on their bins and the bins of adjacent dice on the wafer. To achieve this, an underlying bayesian network represents both, production caused bin relations of adjacent dice and test error influences on die binning. Based on the basic die probability model, we present three wafer retest recommendation models. Since they allow prompt analyses of wafer runs and deliver die or bin specific retest recommendations, they should enable a faster and more efficient test error detection than provided by standard detection methods like static bin limits or regular retests. To evaluate our approach, we used test data from a real semiconductor test process. In the our experiments we studied the basic detection performance of the die probability model and compared our wafer retest recommendation models to the standard strategies in terms of detection ratio, retest ratio and retest efficiency.