Unsupervised Defect Segmentation for Wafer Inspection Noise Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Defect segmentation in semiconductor wafers is inefficient due to the vast number of data points and the shrinking differences between noise floors associated with defects and nuisance inspection datapoints, making supervised segmentation impractical and ineffective.
Innovation Solution
An unsupervised segmentation system that uses a controller to receive inspection datasets with signal and context attributes, labels them based on positions in a signal space, and trains a classifier to segment defects into groups without human intervention, utilizing both signal and context attributes to identify overlapping distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised segmentation is used with human experts, then defect classification accuracy is improved, but processing time and operational complexity increase significantly
Solution Approach 1:
The system performs unsupervised segmentation automatically without human expert intervention. The algorithm independently processes inspection datasets, labels datapoints with class labels based on signal space positions, and segments defects into groups using trained classifiers, enabling the system to serve itself rather than requiring external human expertise for each segmentation task
Solution Approach 2:
The system transforms the segmentation approach by changing from supervised to unsupervised methodology. It uses signal attributes and context attributes to define signal space, then applies parameter-based clustering and classification algorithms that automatically adapt to the data characteristics without human-guided parameter tuning, achieving both speed and accuracy
2Measurement precision
If supervised segmentation with multiple attributes is used, then segmentation accuracy is improved, but system complexity and practical implementability worsen
Solution Approach 1:
The system segments the complex segmentation task into distinct automated components: (1) defining signal space using signal attributes, (2) labeling datapoints with class labels based on their positions in signal space, (3) training classifiers with context attributes and class labels, and (4) segmenting defects into groups using the trained classifiers. This modular approach reduces overall system complexity while maintaining high segmentation accuracy
Solution Approach 2:
The system replaces the mechanical process of human expert supervision with automated computational algorithms. Instead of human experts manually analyzing multiple attributes, the system uses computer-based signal space mapping, automated labeling, and machine learning classifiers to perform segmentation, significantly reducing operational complexity while maintaining or improving accuracy
3Productivity
If segmentation is based solely on signal attributes, then processing speed is improved, but segmentation effectiveness deteriorates due to shrinking noise floor differences
Solution Approach 1:
The system enhances segmentation effectiveness by adding another dimension to the analysis. It combines signal attributes (which provide processing speed) with context attributes (which provide segmentation discrimination power). The context attributes add a new dimension to the feature space, enabling the system to distinguish between defects with similar signal characteristics but different contextual meanings, thereby maintaining both speed and effectiveness
Data Source
AI summary
An inspection system may receive inspection datasets from a defect inspection system associated with inspection of one or more samples, where an inspection dataset of the plurality of inspection datasets associated with a defect includes values of two or more signal attributes and values of one or more context attributes. An inspection system may further label each of the inspection datasets with a class label based on respective positions of each of the inspection datasets in a signal space defined by the two or more signal attributes, where each class label corresponds to a region of the signal space. An inspection system may further segment the inspection datasets into two or more defect groups by training a classifier with the values of the context attributes and corresponding class labels for the inspection datasets, where the two or more defect groups are identified based on the trained classifier.