Rare Defect Training Data for Semiconductor Yield Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning (ML) binning and classification methods for semiconductor defects fail to accurately recognize rare events due to insufficient representation in training sets, leading to misclassification and increased operator workload without guaranteed accuracy.
Innovation Solution
A system and method for generating a rare defect type training data set by collecting and labeling images of known rare defects on specimens, which are then combined with specimen-specific training data to enhance the accuracy of defect classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more data is accumulated for training to improve model performance, then the model accuracy may improve, but the operator workload and time-to-results significantly increase
Solution Approach 1:
The system performs preliminary action by proactively detecting and collecting rare defect events during normal operation before they become critical issues. The automated rare event detection system continuously monitors inspection data, identifies rare defects, and accumulates training data in advance, eliminating the need for operators to manually collect extensive training data later.
Solution Approach 2:
The system implements self-service through automated rare event detection and automatic training data generation. The inspection system itself identifies rare defects, extracts relevant images, and creates training data without operator intervention. This self-accumulating approach reduces operator workload while maintaining model accuracy.
2Reliability
If manual SEM review and classification of inspection data is performed to ensure all rare events are identified, then training data completeness improves, but operator workload increases significantly
Solution Approach 1:
The system replaces the mechanical manual SEM review process with an automated computational approach. The rare event detection system uses algorithms to automatically identify rare defects in inspection data, substituting operator manual review with automated image processing and pattern recognition, thereby reducing operator workload while maintaining detection completeness.
Solution Approach 2:
The system introduces an intermediary automated detection layer between the inspection process and training data generation. This intermediary system processes inspection data, identifies rare events, and prepares training data automatically, serving as a mediator that eliminates the need for direct operator involvement in rare event identification.
3Measurement precision
If current ML methods are used for defect classification, then common defects are classified accurately, but rare defects are misclassified due to insufficient representation in training sets
Solution Approach 1:
The system performs preliminary action by proactively detecting and collecting rare defect events during normal operation before they become critical issues. The automated rare event detection system continuously monitors inspection data, identifies rare defects, and accumulates training data in advance, eliminating the need for operators to manually collect extensive training data later.
Solution Approach 2:
The system changes the parameter of training data composition by dynamically incorporating detected rare events into the training set. As new rare defects are identified, the system automatically updates the training data to include these new patterns, making the model adaptable to changing defect types without retraining from scratch.
Data Source
AI summary
Methods and systems for determining information for a specimen are provided. One method includes collecting images of known rare defect types previously detected on one or more other specimens and assigning training labels to the images responsive to the known rare defect types in the images. The method also includes storing the collected images and the assigned training labels as a rare defect type training data set. The rare defect type training data set is unsuitable for use in training a component configured for determining information for the specimen from runtime images generated for the specimen by an imaging system until the rare defect type training data set is combined with training images and corresponding training labels generated for the specimen. The method may also or alternatively perform these steps for known tool excursion events previously detected on one or more additional specimens.