Deep Learning Ensemble for Semiconductor Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting defects in semiconductor manufacturing are inadequate for high volume manufacturing (HVM) due to high false positive rates and sensitivity to optimizer parameters, lacking theoretical foundations for performance improvement with increased computational resources.
Innovation Solution
A system utilizing an ensemble of deep learning (DL) models with a pseudo-loss function is employed for defect detection. The ensemble is trained with a training dataset to minimize the pseudo-loss function, achieving a balance between true defect capture and false positive reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current defect detection methods are used in high volume manufacturing, then throughput can be maintained, but false positive rates remain high and reliability deteriorates
Solution Approach 1:
The patent divides the defect detection task into multiple independent deep learning models that process different aspects or features of the inspection data. Each model in the ensemble specializes in detecting specific defect patterns, and their predictions are combined to achieve high accuracy while maintaining throughput through parallel processing.
Solution Approach 2:
The patent replaces traditional mechanical or rule-based defect detection systems with deep learning-based automated detection. This substitution enables the system to achieve both high throughput through efficient computational processing and high reliability through learned patterns from training data, eliminating the trade-off between speed and accuracy.
2Reliability
If traditional optimization methods are used for defect detection, then computational resources can be limited, but performance improvement with increased resources lacks theoretical foundation
Solution Approach 1:
The patent changes the parameters of the deep learning models systematically, including network architecture, learning rates, and ensemble sizes, to optimize performance. The theoretical framework guides these parameter changes to ensure that increased computational resources translate into predictable and measurable improvements in detection accuracy.
Solution Approach 2:
The patent implements feedback mechanisms where the performance of the ensemble model is continuously evaluated against ground truth data, and the results are used to refine the model parameters and training processes. This feedback loop ensures that computational resources are efficiently utilized to achieve optimal detection accuracy.
3Reliability
If deep learning models are trained with extensive hyperparameter tuning, then detection accuracy can be improved, but manufacturing cost and time increase
Solution Approach 1:
The patent performs preliminary actions by pre-training the deep learning models on large datasets and pre-tuning the hyperparameters using systematic search methods before deployment. This preliminary preparation reduces the need for extensive hyperparameter tuning during actual manufacturing operations, thereby reducing both time and cost while maintaining high detection accuracy.
Solution Approach 2:
The patent discards poorly performing models or hyperparameter configurations from the ensemble during the selection process, and recovers computational resources by focusing training efforts only on the most promising configurations. This selective approach reduces overall training time while achieving the desired false positive rate.
Data Source
AI summary
Methods and systems for detecting defects in images of a specimen are provided. One system includes a computer subsystem configured for training an ensemble of deep learning models by altering one or more parameters of the ensemble until a pseudo-loss function determined based on output of the ensemble is approximately equal to but not greater than 0.5. The computer subsystem is also configured for detecting defects in runtime specimen images by inputting the runtime specimen images into the trained ensemble and generating runtime labels for the runtime specimen images indicating if a defect has been detected in the runtime specimen images based on outputs of the deep learning models in the trained ensemble.