Position-Aware Self-Supervised Learning for Wafer Map Defect Pattern Recognition

Abstract

Wafer map defect pattern recognition is an indispensable component of semiconductor manufacturing, providing crucial information for identifying the root causes of defects in semiconductor production. In recent years, to address the overreliance on labeled data in supervised learning approaches, some efforts have introduced the concept of self-supervised learning into wafer map defect pattern recognition. However, these studies often ignore the significant data characteristics related to the spatial location of defect clusters on the wafer map itself. To address this issue, we designed an RingDistanceConv (RDConv) module to consider the impact of two types of position information—coordinates and distances—on wafer map defect recognition and proposed the position-aware self-supervised learning framework. Our framework achieved an accuracy of 96.41% on the WM-811K dataset with eight defect classes.