Unsupervised Representation Learning and Explainable Clustering for Wafer Map Pattern Analysis

As wafer maps become increasingly complex and high-dimensional, conventional clustering methods often fail to uncover subtle but meaningful defect patterns critical for yield enhancement and fault diagnosis in semiconductor manufacturing. We present an unsupervised clustering framework tailored to wafer map analysis, combining a convolutional autoencoder for automated feature extraction with principal component analysis for dimensionality refinement. Additionally, we incorporate improved deep embedded clustering, which augments the autoencoder with a clustering-oriented Kullback-Leibler divergence loss to learn compact and confident latent representations. Using standard clustering metrics and extensive visualization, our method is evaluated on two private industrial micro-electromechanical systems datasets and the public MIR-WM811K dataset. Unlike prior approaches, we introduce a comprehensive evaluation strategy that includes (i) cluster confidence and entropy distributions to assess prediction determinism, (ii) semi-supervised scoring for structure-aware validation, and (iii) interpretable visual tools, such as SHapley Additive exPlanations maps, gradient-weighted class activation mapping overlays, and average cluster profiles to support human-in-the-loop decision-making. Results show that our framework consistently outperforms baseline methods, including pretrained visual models like DINOv2 and TIMM-ResNet, in both clustering quality and interpretability. By aligning unsupervised representations with domain-specific failure semantics, the proposed pipeline enables more transparent and actionable analysis of wafer maps. Integrating automated feature learning, probabilistic confidence modelling, and visual attribution offers a robust path toward root-cause identification and process optimization in modern semiconductor fabrication.