Defect Classification via Confidence Thresholds and Manual Review
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic defect classification systems in the semiconductor industry face challenges in achieving high purity of classification while managing rejection rates, particularly due to classification uncertainties and overlap regions in multi-dimensional feature spaces, leading to inefficiencies in defect identification and classification.
Innovation Solution
The system employs a combination of multi-class and single-class classifiers, with adjustable confidence thresholds to maximize classification purity while maintaining a predefined rejection rate. Defects with low confidence are further analyzed by additional inspection modalities, including human inspectors or other data sources, to refine classification results and integrate them for comprehensive reporting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If automatic classifier is applied to all defects, then classification purity is improved, but rejection rate increases due to classification uncertainties
Solution Approach 1:
The system applies different classification strategies to different defect subsets: high-confidence defects are classified automatically with strict purity criteria, while low-confidence defects are handled with more lenient criteria or manual review. This local differentiation allows the system to maximize overall purity without unnecessarily rejecting all defects.
Solution Approach 2:
The system dynamically adjusts confidence thresholds and classification parameters based on defect characteristics, inspection conditions, and performance metrics. By changing parameters such as confidence levels, kernel widths, and decision boundaries, the system optimizes the balance between purity and rejection rate for different operating conditions.
2Productivity
If confidence threshold is increased to improve classification purity, then fewer defects are rejected, but classification accuracy decreases
Solution Approach 1:
The system employs dynamic confidence thresholds that adapt based on defect characteristics, inspection conditions, and accumulated performance data. Rather than using a fixed threshold, the system adjusts thresholds dynamically to maintain optimal balance between classification volume and accuracy for different defect types and inspection scenarios.
Solution Approach 2:
The system incorporates feedback loops where classification results, including manual reviews and corrections, are used to refine and update confidence thresholds and classifier parameters. This continuous learning process allows the system to improve accuracy while maintaining high classification throughput over time.
3Reliability
If multiple inspection modalities are applied to low confidence defects, then classification completeness is improved, but system complexity increases
Solution Approach 1:
The system segments defects into confidence-based groups and applies different inspection modalities to each group. Only low-confidence defects undergo additional inspection modalities or manual review, while high-confidence defects are processed automatically. This segmentation reduces overall system complexity by limiting complex processing to only the necessary subset of defects.
Solution Approach 2:
The system uses a unified multi-class classifier framework that can handle multiple defect types and inspection modalities through a common architecture. This universal approach allows the system to manage complexity by providing a single integrated platform that can adapt to different inspection needs rather than requiring separate specialized systems for each modality.
Data Source
AI summary
A method for defect classification includes storing definitions of defect classes in terms of a classification rules in a multi-dimensional feature space. Inspection data associated with defects detected in one or more samples under inspection is received. A plurality of first classification results is generated by applying an automatic classifier to the inspection data based on the definitions, the plurality of first classification results comprising a class label and a corresponding confidence level for a defect. Upon determining that a confidence level for a defect is below a predetermined confidence threshold, a plurality of second classification results are generated by applying at least one inspection modality to the defect. A report is generated comprising a distribution of the defects among the defect classes by combining the plurality of first classification results and the plurality of second classification results.