Defect Classifier Using Hybrid Rule and Example-Based Connection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional defect classification systems in semiconductor manufacturing face challenges in accurately classifying a large number of defect classes, particularly when complex defects occur, leading to cumbersome recipe setting operations and lengthy setup times, especially when integrating rule-based and example-based classifiers.
Innovation Solution
A defect reviewing apparatus that automatically connects rule-based and example-based classifiers using training samples and user-defined attribute information, allowing for dynamic adjustment of classification classes and enabling high-precision classification even with numerous classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of classification classes are defined to accurately categorize complex defects, then classification precision is improved, but the complexity of recipe setting operations increases
Solution Approach 1:
The system automatically generates classification recipes by having the classifier itself learn from training data and define class connections without manual intervention. The classifier autonomously determines feature importance, class hierarchies, and decision boundaries, eliminating the need for operators to manually configure complex recipe parameters for each classification class.
Solution Approach 2:
The system dynamically adjusts classification parameters based on the specific defect data being analyzed. Instead of using fixed pre-defined recipes, the classifier adapts feature weights, threshold values, and class relationships in real-time based on the training data characteristics, allowing accurate classification across diverse defect types without manual recipe reconfiguration.
2Measurement precision
If manual recipe setting operations are performed for each classification class, then classification accuracy can be optimized, but setup time becomes lengthy
Solution Approach 1:
The system performs preliminary learning from training data before actual classification operations. By pre-processing and storing class relationships, feature importance rankings, and decision models during the training phase, the system eliminates the need for time-consuming manual recipe setting during operational setup, achieving both accuracy and speed.
Solution Approach 2:
The classifier automatically generates and optimizes its own classification recipes by analyzing training data patterns, eliminating the need for manual operator intervention in recipe creation. This self-configuration capability reduces setup time while maintaining high classification accuracy through data-driven parameter optimization.
3Adaptability or versatility
If rule-based and example-based classifiers are integrated to handle complex defects, then classification capability is improved, but the complexity of connecting different classifier types increases
Solution Approach 1:
The system merges rule-based classification (using feature thresholds and logical conditions) with example-based classification (using learned patterns from training data) into a unified classifier architecture. The integration automatically combines both approaches by having the classifier learn from training data and generate hybrid decision models that leverage the strengths of both rule-based and example-based methods without manual integration complexity.
Solution Approach 2:
The integrated classifier autonomously manages the coordination between rule-based and example-based components by automatically learning from training data which methods to apply and how to weigh them. The system self-determines the optimal combination strategy based on the specific classification task and data characteristics, eliminating manual configuration complexity.
Data Source
AI summary
In apparatuses for automatically acquiring and also for automatically classifying images of defects present on a sample such as a semiconductor wafer, a classifying system is provided which are capable of readily accepting even such a case that a large number of classification classes are produced based upon a request issued by a user, and also even such a case that a basis of the classification class is changed in a high frequency. When the user defines the classification classes, a device for designating attributes owned by the respective classification classes is provided. The classifying system automatically changes a connecting mode between an internally-provided rule-based classifier and an example-based classifier, so that such a classifying system which is fitted to the classification basis of the user is automatically constructed.