Multi-Engine Substrate Defect Classification to Reduce Escape Rates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual classification of semiconductor substrate defects leads to human errors and inefficiencies, resulting in underkill or overkill of dies, and existing automated methods suffer from high escape rates and inaccuracies.
Innovation Solution
A multi-classification engine using multiple machine learning models, such as CNN and KNN, for substrate defect classification, with continuous learning to adapt to new defect patterns and reduce false positives and negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual classification is used, then operators can review and classify substrate defects, but human errors occur leading to underkill or overkill of dies
Solution Approach 1:
The patent replaces the manual mechanical classification process with an automated machine learning-based classification system. Multiple ML models (CNN, KNN, SVM) automatically classify substrate defects without human intervention, eliminating human errors while maintaining high accuracy. The system processes defect images through multiple algorithms that output classification results, defect scores, and confidence levels without requiring operator workload.
Solution Approach 2:
The classification system performs self-service by automatically reviewing and classifying substrate defects without requiring operator intervention. The multiple ML models independently analyze defect images, generate classifications, and provide defect scores autonomously. The system serves itself by continuously learning from new data and improving its own classification accuracy over time through model updates.
2Productivity
If existing automated classification methods are used, then operator workload is reduced, but escape rates remain high due to inaccuracies
Solution Approach 1:
The patent merges multiple machine learning models (CNN, KNN, SVM) into a unified classification system. Each model contributes its strengths to the overall classification process, with CNN extracting features, KNN comparing patterns, and SVM making final classifications. This combination of multiple models reduces escape rates by cross-validating results and improving overall accuracy while maintaining high classification throughput.
Solution Approach 2:
The classification system implements feedback mechanisms where defect classification results and defect scores are continuously analyzed. Misclassified defects are identified through confidence level thresholds and fed back into the system for model retraining and improvement. This feedback loop continuously reduces escape rates by learning from errors and improving classification accuracy over time while maintaining productivity.
3Extent of automation
If a single machine learning model is used, then automation is achieved, but accuracy is insufficient leading to misclassification
Solution Approach 1:
The patent combines multiple machine learning models (CNN, KNN, SVM) to achieve both full automation and high accuracy. Each model performs a specific function: CNN for feature extraction, KNN for pattern recognition, and SVM for classification. The merged system processes defect images through all models and aggregates results, achieving automated classification with superior accuracy that reduces misclassification compared to single-model systems.
Solution Approach 2:
The classification system uses a composite approach by integrating multiple ML models with different algorithms and strengths. Like composite materials combine properties of individual materials, this composite ML system combines the pattern recognition capabilities of CNN, the similarity-based classification of KNN, and the decision boundary strength of SVM to achieve high accuracy automated classification that overcomes the limitations of any single model.
Data Source
AI summary
Methods and systems are provided for improving substrate defect classification in semiconductor manufacturing by using more than one machine learning model to classify substrate defect data. The method comprises a defect inspection module that captures substrate defect image data and a defect classification part that processes the data using more than one machine learning model. The output from each model is used to produce the final classified data. The defect score is calculated based on the classification results and this defect score is used to identify the misclassified substrate defect. The model can be updated after each inspection run cycle leading to increased accuracy and a lower escape rate.