Abstract
Good-die-in-bad-neighborhoods (GDBN) has long been used as a technique to identify die test escapes in wafer processing, effectively lowering DPPM (defect parts per million) by finding dice that test good initially but fail further reliability tests. Traditionally, GDBN identifies suspicious dice by investigating the amount of die failures in the neighbourhood of a die under investigation (DUI). In this work, we propose using epitaxial growth defects during wafer manufacturing as an input to directly predict where these die failures occur. Using the same GDBN approach, we can exploit the spatial patterns and density of various types of epitaxial defects around a die to predict its probability of failure. Convolutional Neural Networks (CNN) will be used to realize this, which has been proven to be effective in capturing complex spatial relationships in computer vision tasks. In this seminar, methods of how epitaxial defects can be encoded as CNN inputs will be discussed, as well as the effect specific types of epitaxial defects have on die failures.
Anyone interested is welcome to attend.