Active Learning Defect Classifier Using Bayesian Uncertainty Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for defect classification in semiconductor manufacturing face challenges such as the difficulty in finding suitable defect examples for training, inability to support deep learning models, lack of theoretical foundation for heuristic acquisition functions, and inability to provide uncertainty estimations, leading to inefficient defect discovery and classification.
Innovation Solution
A system and method for active learning that includes an imaging subsystem and computer subsystems for selecting data points based on uncertainty estimations, acquiring labels, and training a defect classifier using Bayesian deep learning to enable adaptive sampling and joint setup across multiple tools.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional heuristic acquisition functions are used for defect sampling, then the process is simple to implement, but there is no theoretical foundation and no lower bound performance guarantee
Solution Approach 1:
The patent introduces Bayesian deep learning as an intermediary framework that bridges the gap between simple heuristic methods and rigorous theoretical guarantees. The Bayesian framework provides probabilistic predictions with uncertainty quantification, serving as a mediator that transforms arbitrary data points into statistically sound sampling decisions with theoretical performance bounds.
Solution Approach 2:
The patent changes the fundamental parameters of the acquisition function by moving from deterministic heuristic rules to probabilistic Bayesian predictions. This parameter transformation enables the system to provide both theoretical performance guarantees through posterior uncertainty bounds and practical effectiveness through adaptive sampling based on predicted defect locations.
2Measurement precision
If deep learning models are used for defect classification, then the model accuracy is improved, but the training data requirement increases significantly
Solution Approach 1:
The patent applies preliminary action by using Bayesian deep learning to pre-identify high-probability defect locations and generate synthetic training data before actual defect classification. The Bayesian model performs preliminary uncertainty analysis to guide where training data should be collected, reducing the overall quantity needed for deep learning model training.
Solution Approach 2:
The system implements self-service through active learning loops where the Bayesian model automatically identifies its own training needs and guides data collection. The model serves itself by generating acquisition functions that prioritize samples most beneficial for its own improvement, reducing dependency on large pre-labeled datasets.
3Ease of manufacture
If manual defect example collection is used for training, then the training process is simple to setup, but the time and effort required increases significantly
Solution Approach 1:
The patent implements feedback mechanisms where the Bayesian deep learning model continuously learns from collected defect data and automatically refines its acquisition function. This feedback loop replaces manual defect example collection with automated iterative learning, maintaining setup simplicity while dramatically reducing the time required for defect discovery through adaptive sampling.
Solution Approach 2:
The patent substitutes the mechanical process of manual defect hunting with an automated Bayesian optimization system. Instead of manually searching and collecting defect examples, the system uses probabilistic models and acquisition functions to automatically guide data collection, replacing human effort with computational intelligence while maintaining ease of setup.
4Reliability
If uncertainty estimations are provided in defect prediction, then the reliability of predictions is improved, but the computational complexity increases
Solution Approach 1:
The patent segments the computational workload by separating uncertainty estimation from final defect classification. The Bayesian deep learning model divides predictions into discrete uncertainty levels that guide active learning sampling, breaking down the complex task of simultaneous classification and uncertainty quantification into manageable segments that reduce overall computational burden.
Data Source
AI summary
Methods and systems for performing active learning for defect classifiers are provided. One system includes one or more computer subsystems configured for performing active learning for training a defect classifier. The active learning includes applying an acquisition function to data points for the specimen. The acquisition function selects one or more of the data points based on uncertainty estimations associated with the data points. The active learning also includes acquiring labels for the selected one or more data points and generating a set of labeled data that includes the selected one or more data points and the acquired labels. The computer subsystem(s) are also configured for training the defect classifier using the set of labeled data. The defect classifier is configured for classifying defects detected on the specimen using the images generated by the imaging subsystem.