Machine Learning Reticle Inspection Without Design Database
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current reticle inspection techniques are inadequate for detecting defects in non-repeating patterns without a design database, particularly in semiconductor manufacturing, where defects are becoming increasingly smaller and harder to identify due to the decreasing size of semiconductor devices.
Innovation Solution
A machine learning method using multidimensional classifiers formed from reflected and transmitted images or signals from various inspection modes, which maps feature vectors into a higher-dimensional space to locate defects in reticles without relying on a design database, employing techniques such as voxel-based mapping, healing of holes, and distance transformation to enhance defect detection sensitivity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional reticle inspection techniques are used, then the inspection process is simple, but the sensitivity of defect detection is insufficient for small defects in non-repeating patterns
Solution Approach 1:
The patent transforms the inspection problem from 2D image space to higher-dimensional feature space by extracting multiple features (intensity, gradient, texture, frequency) and mapping them into a multidimensional space. This dimensional transformation enables the classifier to distinguish defects from pattern variations more effectively, achieving higher detection sensitivity without proportionally increasing system complexity.
Solution Approach 2:
The patent introduces a machine learning classifier as an intermediary between the raw inspection images and the defect detection decision. This classifier learns the complex mapping between image features and defect presence from training data, serving as a mediator that handles the complexity of pattern recognition while keeping the overall system architecture manageable.
2Reliability
If design database is used for inspection, then defect detection accuracy improves, but the method cannot inspect non-repeating patterns or when design database is unavailable
Solution Approach 1:
The patent enables the inspection system to learn from the training data and become self-sufficient in detecting defects without requiring external design database information. The classifier learns the characteristics of defects and normal variations directly from the training images, making the system adaptable to non-repeating patterns and independent of design database availability.
Solution Approach 2:
The patent changes the inspection approach from database-dependent pattern matching to database-independent feature-based classification. By transforming the inspection parameters from direct pattern comparison to learned feature space distance measurement, the system achieves both high reliability and broad adaptability to different pattern types.
3Measurement precision
If inspection sensitivity is increased to detect smaller defects, then false positives and negatives increase
Solution Approach 1:
The patent segments the defect detection task into multiple independent feature extractions (intensity, gradient, texture, frequency) and subsequent classification in feature space. This segmentation allows each feature to capture different aspects of defect characteristics, improving sensitivity while the combined feature space classification maintains reliability by considering multiple indicators simultaneously.
Solution Approach 2:
By mapping inspection data into higher-dimensional feature space, the patent creates additional separation between defect and non-defect cases. The multidimensional distance measurement in feature space provides more discriminative power, enabling the system to detect smaller defects with higher confidence and reduce false positives compared to traditional 2D image analysis.
Data Source
AI summary
Apparatus and methods for inspecting a specimen are disclosed. An inspection tool is used at one or more operating modes to obtain images of a plurality of training regions of a specimen, and the training regions are identified as defect-free. Three or more basis training images are derived from the images of the training regions. A classifier is formed based on the three or more basis training images. The inspection system is used at the one or more operating modes to obtain images of a plurality of test regions of a specimen. Three or more basis test images are derived from to the test regions. The classifier is applied to the three or more basis test images to find defects in the test regions.