In a semiconductor manufacturing operation, it is critical to maintain product yield, cycle time, and quality due to its costly and complex processes. Increasing unit probe (UP) bin fallouts for a qualified device raised delivery concerns and required risk mitigation. Management frequently asks for an estimate of the potential scrap risk from below minimum yields and/or exceeded statistical bin limits. Planning groups need to know if new wafer lot(s) are needed to backfill the die shortage, while product groups are interested in knowing the yield impact at unit probe. As a device owner, I used data mining techniques to identify the shifts of the parametric probe parameters that correspond to UP fallout and made advanced decisions based on predictive models.
In this demonstration, unit probe and class probe data was extracted and cleaned up from the production database. The collected data was split into training, validation, and test data sets. Four models were evaluated and analyzed at wafer level. Classification decision tree models and bootstrap forest classification models were chosen for categorical response analysis (scrap or not), while linear regression models and bootstrap forest regression models were compared for continuous response analysis (UP bin fallout %). Model screening helped reduce parameters from hundreds to a handful. The resulting final decision was made on the wafers' disposition based on the constructed predictive model.
When the UP response was available from the test data set, the accuracy of the two prediction models were compared between the predicted and probed bin fallouts.
