Unsupervised feature learning and optimised clustering with fine-tuning for automatic semiconductor wafer defect recognition

Abstract Automated identification of wafer defect maps (WDMs) is critical for yield enhancement and quality control in semiconductor manufacturing. However, as defect patterns become increasingly diverse and dynamic, conventional supervised methods remain constrained by their dependence on fully labelled data and their limited adaptability to changing defect distributions. This study proposes a two-stage framework for automatic wafer defect map (WDM) recognition using predominantly unlabelled data with limited prior knowledge. First, a deep autoencoder-based representation learning approach is developed to automate feature extraction, followed by Elbow-guided K-means clustering to determine the number of clusters and capture the underlying defect structure. This stage provides an initial basis for automatic labelling without requiring exhaustive manual annotation. In the second stage, because these initial labels may become less reliable under complex data distributions and evolving defect patterns, XGBoost-based test-time training and semi-supervised refinement (TTT-SSR) is used to refine predictions by exploiting limited knowledge from the training phase. Experimental results on a benchmark mixed-WDM dataset show that the proposed framework achieves 95% classification accuracy, together with competitive precision, recall and F1-score relative to state-of-the-art methods, while also improving clustering quality across key evaluation metrics.