패턴 불균형에 강건한 자가 지도학습을 활용한 웨이퍼 불량 패턴 클러스터링 방법 제안

Abstract

This study proposes a wafer defect pattern clustering model that can recognize defect patterns without the class label of the defect patterns. In the first step, noise defects are removed from each wafer bin map (WBM) image using the Depth-First Search (DFS) algorithm to clarify the defect pattern. Next, the defect patterns are clustered using the Dirichlet process, and the clustering results are adjusted by tuning the extracted features based on self-supervised learning. By employing a weighted cross-entropy loss that considers the cluster size, the model becomes robust to the imbalance of cluster sizes during the fine-tuning process. The proposed method can facilitate the identification and resolution of the causes of defects that occur during semiconductor processing.