액티브 러닝을 활용한 반도체 웨이퍼 신규 불량 패턴 검출

Abstract

In the semiconductor manufacturing industry, a wafer bin map (WBM) contains defect patterns that provides important clues to identify the root causes of the defect. Traditionally, field engineers classify the pattern types by manually checking WBM. Recently, many studies have been conducted for automatic classification by using deep learning models. To accurately classify defect patterns with convolutional neural network (CNN)-based deep learning models, every WBM must have accurate pattern labels. However, in reality, it takes a lot of time and efforts for engineers to label all the data. In addition, existing CNN-based studies show limitations that cannot detect new defect patterns, frequently occurred in real situations. In this study, we devise a new pattern detection framework based on active learning. Through this, new patterns can be selectively detected even with existing active learning methodologies, and classification performance can be secured by effectively sampling unlabeled data. And we compared the performance and characteristics of each sampling strategy for new pattern detection. The usefulness and applicability of this study was demonstrated by WM-811K, publicly available WBM data.