Self-supervised multi-resolution learning for label-agnostic morphology representation and clustering of semiconductor thin-film SEM defects

Scanning Electron Microscope (SEM) image analysis plays a vital role in semiconductor thin-film characterisation. In particular, defect detection and classification are performed using SEM images. However, conventional methods rely on labelled datasets or handcrafted features for classification, which limits their generalisation in real-world industrial inspection settings. The present work proposes a self-supervised multi-resolution learning framework for label-agnostic morphology representation learning and clustering of semiconductor thin-film defects. It uses a SEM image Dataset (4591 images) obtained from industrial wafer inspection. The framework starts with image pre-processing to remove acquisition artifacts. It employs a multi-resolution image pyramid for capturing surface morphologies at fine, intermediate, and coarse spatial scales. A shared-weight convolutional encoder that ensures alignment between embeddings across the three resolutions is trained (on an unlabelled dataset) and utilised for unsupervised defect morphology grouping. The framework learns morphological representations without defect labels. However, defect labels are used only for post-hoc evaluation by normalized mutual information (NMI) and visualization. Intrinsic clustering metrics and low-dimensional visualization are used to assess the algorithm's efficacy. Experimental results reveal that the proposed method, gray level co-occurrence matrix (GLCM), local binary patterns (LBP), wavelet-based features, and principal component analysis (PCA) on raw pixels obtained a silhouette score of 0.50, 0.43, 0.31, 0.45, and 0.22, respectively. While normalized mutual information (NMI) values remained moderate across the models. These results reflect the label-agnostic nature of the proposed SSL framework. Further, UMAP and t-SNE visualizations confirm the coherent manifold structure and the effectiveness of morphology-driven grouping. These results demonstrate the robust, scale-invariant quality of the proposed self-supervised multi-resolution learning framework for defect clustering.