Active Learning Wafer Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor inspection systems using multiple electron detectors struggle with low combined detection efficiencies for secondary and backscattered electrons, leading to inadequate image quality for high accuracy and high throughput defect inspection and metrology of two-dimensional and three-dimensional structures.
Innovation Solution
A machine learning model is trained using a pool of labeled measurement images to determine a utility function value for unclassified measurement images. Based on this value, unclassified images are classified without using the machine learning model and added to the pool of labeled images, improving the model's training and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple electron detectors are used to detect secondary and backscattered electrons, then surface information collection and buried layer information collection are improved, but combined detection efficiency remains low
Solution Approach 1:
A machine learning model serves as an intermediary to classify defects in measurement images. The model processes images containing both secondary electron and backscattered electron information, automatically identifying and categorizing defects without requiring manual intervention or complex combined detection systems. This intermediary approach resolves the contradiction by achieving high measurement precision through intelligent image analysis while maintaining productivity through automated processing.
Solution Approach 2:
The patent replaces the mechanical/physical approach of using multiple electron detectors with an information processing approach using machine learning. Instead of physically combining multiple detection systems to improve image quality and efficiency, the invention uses computational methods to analyze and classify defects from measurement images, substituting the complex physical detection system with an intelligent software-based classification system.
2Measurement precision
If manual classification of measurement images is performed to improve training data quality, then machine learning model accuracy is improved, but processing time and cost increase
Solution Approach 1:
The machine learning model performs self-service by automatically classifying defects in measurement images without requiring manual intervention. The system uses the trained model to process new images, extract defect information, and categorize them autonomously. This self-service capability resolves the contradiction by maintaining high classification accuracy through intelligent algorithms while eliminating the time loss associated with manual classification processes.
Solution Approach 2:
The invention extracts the classification function from the manual process and embeds it within the machine learning model. By taking out the time-consuming manual classification step and replacing it with automated model-based classification, the system preserves measurement precision while removing the bottleneck of manual processing time.
3Measurement precision
If the machine learning model is iteratively trained with more labeled data to improve accuracy, then defect classification precision is improved, but training time and computational resources increase
Solution Approach 1:
The system performs preliminary action by pre-processing measurement images and pre-training the machine learning model with initial labeled data before actual defect classification tasks. This preliminary preparation allows the model to be ready for deployment with baseline accuracy, and enables incremental learning where the model can be efficiently retrained with new data when needed, rather than requiring extensive retraining for every improvement.
Solution Approach 2:
The invention applies partial action by training the machine learning model with a sufficient but not excessive amount of labeled data to achieve the required classification precision. Rather than continuously increasing training data and training time indefinitely, the system identifies the point of diminishing returns where additional training provides minimal improvement, thus optimizing the balance between classification precision and training time investment.
Data Source
AI summary
Systems and methods for training a machine learning model to classify defects with utility-function-based active learning are described. In one embodiment, one or more non-transitory, machine-readable mediums are configured to cause a processor to at least determine a utility function value for unclassified measurement images, based on a machine learning model, wherein the machine learning model is trained using a pool of labeled measurement images. Based on a determination that the utility function value for a given unclassified measurement image is less than a threshold value, the unclassified measurement image is output for classification without the use of the machine learning model. The unclassified measurement images classified via the classification without the use of the machine learning model are added to the pool of labeled measurement images. The machine learning model is trained based on the measurement images classified via the classification without the use of the machine learning model.