Defect Classifier Training Using Representative Subset Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current defect detection methods in semiconductor fabrication struggle with high sensitivity to defects of interest while suppressing noise and false alarms, particularly in the context of ultra-large scale integration where precision and uniformity are critical, and existing methods are inefficient in validating classifiers and determining review budgets for defect classification.
Innovation Solution
A system and method that utilize a processing and memory circuitry to select a representative subset of potential defects based on their distribution in an attribute space, train a classifier using informative attribute values, and apply it to estimate the number of expected defects of interest, thereby improving defect classification and reducing false alarms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If high sensitivity defect detection is applied to detect all potential defects, then defect detection capability is improved, but false alarms and noise increase significantly
Solution Approach 1:
The patent extracts only the most informative and representative defects from the complete defect list to create a reduced training set. By selecting defects based on their representativeness across different attribute spaces (size, shape, location), the system extracts essential training data while excluding redundant and less informative defects, thereby reducing false alarms during classification without sacrificing detection sensitivity
Solution Approach 2:
The patent applies local quality by differentiating defects based on their specific attribute characteristics. Defects are selected and weighted according to their local properties in the attribute space, such as size distribution, shape characteristics, and spatial location patterns. This allows the classifier to learn distinct features from different defect types while maintaining overall high sensitivity and reducing false alarms through localized feature learning
2Measurement precision
If all potential defects are reviewed to validate the classifier, then classification accuracy is improved, but time consumption and processing resources increase
Solution Approach 1:
The patent segments the complete defect list into a representative subset based on attribute space distribution. By dividing defects into segments that represent different regions and characteristics of the attribute space, the system achieves classification accuracy through a fraction of the total defects, significantly reducing validation time and processing resources while maintaining statistical representativeness
Solution Approach 2:
The patent applies partial action by reviewing only a strategically selected portion of defects rather than all potential defects. The representative subset is constructed to provide sufficient training data for accurate classification, avoiding the excessive action of reviewing every defect. This partial review approach achieves adequate classification accuracy with significantly reduced time consumption
3Reliability
If a large number of defects are included in training data, then classifier learning is improved, but the complexity of data processing and storage increases
Solution Approach 1:
The patent extracts only the essential representative defects from the complete dataset, creating a reduced training set that maintains learning quality. By extracting defects based on their representativeness in the attribute space, the system achieves effective classifier learning with a fraction of the original data, significantly reducing data processing and storage complexity while maintaining training reliability
Data Source
AI summary
A system, method and computer readable medium for examining a specimen, the method comprising: obtaining defects of interest (DOIs) and false alarms (FAs) from a review subset selected from a group of potential defects received from an inspection tool, each potential defect is associated with attribute values defining a location of the potential defect in an attribute space; generating a representative subset of the group, comprising potential defects selected in accordance with a distribution of the potential defects within the attribute space, and indicating the potential defects in the representative subset as FA; and training a classifier using data informative of the attribute values of the DOIs, the potential defects of the representative subset, and respective indications thereof as DOIs or FAs, wherein the trained classifier is to be applied to at least some of the potential defects to obtain an estimation of a number of expected DOIs.