Adaptive Defect Classifier for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for defect classification in semiconductor manufacturing are cumbersome, slow to respond to dynamic changes, and often produce unbalanced training sets, leading to inefficient classifier creation and monitoring, especially with imbalanced datasets where 90% of the data consists of nuisance defects and only 10% are of interest.
Innovation Solution
An adaptive automatic defect classifier system that includes an energy source, detector, and computer subsystems for real-time defect classification, using clustering methods and user feedback to dynamically update and validate classifiers, ensuring robustness through data redundancy scoring and continuous retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional defect classification methods are used with manual classification of all defects, then classification accuracy can be maintained, but the process becomes cumbersome and time-consuming with slow response to dynamic changes
Solution Approach 1:
The system applies partial automation by automatically classifying only those defects for which sufficient training data exists, while leaving other defects for manual classification. This selective approach reduces the time burden while maintaining accuracy for automatable cases, directly addressing the contradiction between accuracy and time loss.
Solution Approach 2:
The defect classification system performs self-learning by automatically updating its training set with newly classified defects and their ground truth labels. This self-service mechanism enables the system to improve its classification capabilities over time without requiring complete manual retraining, reducing both time loss and maintaining accuracy.
2Reliability
If a large number of lots are collected before classifier creation, then the training set becomes more robust, but the response time to dynamic defect changes increases significantly
Solution Approach 1:
The system dynamically adapts its training set by continuously incorporating new defect data and classifications. Rather than requiring a fixed large number of lots before creation, the system evolves its training data over time, allowing it to respond quickly to new defect types while gradually building robustness through continuous learning.
Solution Approach 2:
The system performs preliminary classification with available data immediately rather than waiting for a predetermined number of lots. This preliminary action enables fast response to dynamic changes, while the training set continues to grow and improve robustness in the background through ongoing data collection and learning.
3Reliability
If all defects are manually classified during training and validation phases, then comprehensive training data is obtained, but the process becomes cumbersome and lacks assisted manual classification
Solution Approach 1:
The system applies partial automation by automatically classifying defects where sufficient training data exists, reducing the manual classification burden. This selective automation maintains training data completeness for automatable cases while significantly easing the operational burden on users.
Solution Approach 2:
The automatic classification system acts as an intermediary that handles routine defect classification tasks, freeing human operators to focus on more complex cases. This intermediary layer maintains comprehensive training data while dramatically improving the ease of operation by eliminating the need for complete manual classification.
4Productivity
If the training set includes mostly nuisance defects with few defects of interest, then the classifier creation is faster, but the resulting classifier lacks robustness for defects of interest
Solution Approach 1:
The system uses feedback from manual classifications of defects of interest to continuously improve the training set composition. By monitoring the balance between nuisance and DOI defects, the system can identify when additional DOI examples are needed and prioritize their collection, ensuring robustness for DOIs while maintaining efficient classifier creation.
Solution Approach 2:
The system dynamically adjusts the composition and weighting of training data based on the detected imbalance. When DOI defects are underrepresented, the system can increase their weight or prioritize their collection, changing the training parameters to ensure robustness for defects of interest while maintaining overall creation efficiency.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach enables faster classifier creation, improved responsiveness to dynamic defect changes, and more balanced training sets, resulting in a more robust and accurate defect classification system that reduces manual review time and costs.
Implementation Method 1
The energy source is configured to generate energy that is directed to a specimen. The detector is configured to detect energy from the specimen and to generate output responsive to the detected energy.
Data Source
AI summary
Methods and systems for classifying defects detected on a specimen with an adaptive automatic defect classifier are provided. One method includes creating a defect classifier based on classifications received from a user for different groups of defects in first lot results and a training set of defects that includes all the defects in the first lot results. The first and additional lot results are combined to create cumulative lot results. Defects in the cumulative lot results are classified with the created defect classifier. If any of the defects are classified with a confidence below a threshold, the defect classifier is modified based on a modified training set that includes the low confidence classified defects and classifications for these defects received from a user. The modified defect classifier is then used to classify defects in additional cumulative lot results.