Combining Full Wafer Inspection with Deep Learning to Recognize Wafers with Critical Defects

The paper will present an approach to automate and accelerate detection of defective wafers in high-volume manufacturing by innovatively complementing fast, non-contact and non-destructive photoluminescence wafer mapping with smart image classification using deep learning methods. Specific focus is on distinguishing wafers with regularly occurring defect patterns in the measurement from those revealing atypical patterns from buried, electrically active defects, which are hard to detect by visual inspection of the photoluminescence map. The latter should be reliably and automatically identified after the measurement, thus preventing decrease of production reliability or even product failure through the introduction of further measures on the affected wafers.This task has been treated as an image classification problem for which neural networks-based algorithms were developed. On the one hand, these algorithms deliver promising classification results for the use-case specific wafer maps that are ready for critical assessment by process experts and for comparison to other defect identification techniques. On the other hand, the developed algorithms were verified to perform well on commonly available reference data.