Self-supervised multi-resolution learning for label-agnostic morphology representation and clustering of...

Area of Science:

• Materials Science
• Computer Vision
• Semiconductor Manufacturing

Background:

• Scanning Electron Microscope (SEM) image analysis is crucial for semiconductor thin-film characterization, particularly for defect detection.
• Conventional methods often require labeled datasets or handcrafted features, limiting their adaptability in real-world industrial settings.
• This restricts the generalization capability of existing defect classification techniques.

Purpose of the Study:

• To propose a novel self-supervised multi-resolution learning framework for label-agnostic morphology representation learning and clustering of semiconductor thin-film defects.
• To enable unsupervised grouping of defect morphologies from SEM images without relying on prior defect labels.
• To evaluate the framework's efficacy using intrinsic clustering metrics and low-dimensional visualization techniques.

Main Methods:

• A self-supervised multi-resolution learning framework was developed using a dataset of 4591 industrial wafer inspection SEM images.
• The framework incorporates image pre-processing, a multi-resolution image pyramid for capturing morphologies at various scales, and a shared-weight convolutional encoder.
• The encoder is trained on an unlabeled dataset for unsupervised defect morphology grouping, with labels used only for post-hoc evaluation.

Main Results:

• The proposed self-supervised framework achieved a silhouette score of 0.50, outperforming traditional methods like GLCM (0.43), LBP (0.31), Wavelet features (0.45), and PCA (0.22).
• Normalized Mutual Information (NMI) values were moderate, reflecting the label-agnostic nature of the approach.
• UMAP and t-SNE visualizations confirmed coherent manifold structures and effective morphology-driven grouping, demonstrating scale-invariant clustering.

Conclusions:

• The proposed self-supervised multi-resolution learning framework offers a robust and label-agnostic approach for semiconductor defect clustering.
• The method effectively learns morphological representations and groups defects based on their characteristics across different spatial scales.
• This framework enhances the potential for automated and generalized defect detection in industrial semiconductor inspection.