Comparative Study of AI Models for Wafer Map Defect Pattern Classification Using a Novel Pseudo-Labeling-Based Integrated Dataset
Abstract
Wafer map defect pattern classification is essential for process anomaly diagnosis and yield management in semiconductor manufacturing. However, most existing studies have been conducted using a single labeled wafer map dataset, which limits the diversity of defect patterns. In particular, active learning-based wafer map defect classification has rarely addressed data integration for newly emerging defect patterns while explicitly considering inter-label boundaries and cluster distributions. Therefore, in this study, an integrated dataset with enhanced defect-pattern diversity is constructed by combining labeled WM-811K and unlabeled Mixed38 data through a novel pseudo-labeling-based data integration process that jointly considers cluster distributions and label boundaries to reduce inter-label confusion. Compared with existing pseudo-labeling methods, the proposed method achieves relative improvements of 207.75% in the silhouette coefficient and 16.19% in the Ratkowsky–Lance index, which quantify intra-label compactness and inter-label cluster separation. Meanwhile, based on the integrated dataset with enhanced pattern diversity, the data scale is varied through augmentation, and state-of-the-art (SOTA) AI models based on CNN, transformer, and hybrid architectures are compared under identical conditions to analyze model-specific classification performance and computational efficiency. The validation results show that overall classification performance improves with data scale, while Vision Transformer-based models exhibit competitive accuracy and computational efficiency. By comparing diverse SOTA AI models on an integrated dataset with enhanced defect-pattern diversity and increased data quantity, this study provides a benchmark for AI-based semiconductor wafer map defect pattern classification.