Machine Learning Edge Detection for Semiconductor Pattern Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The die-to-database method for semiconductor pattern inspection struggles with accurate edge detection due to fixed rule-based methods, leading to erroneous detections or failures, especially with pattern deformations and defects, which can result in missed minute defects outside corners.
Innovation Solution
Applying machine learning for edge detection using training data of inspected pattern images and their edges, with a model generation process that excludes defective patterns and balances data distribution across pattern types to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed rule-based edge detection method is used, then the processing is simple and fast, but the detection accuracy deteriorates due to pattern deformations and defects
Solution Approach 1:
The patent replaces the conventional fixed rule-based edge detection method (mechanical system) with a machine learning-based detection system. The machine learning model learns from training data consisting of pattern images and their corresponding edges, enabling accurate edge detection even when patterns are deformed or contain defects, thus resolving the contradiction between detection accuracy and system complexity.
Solution Approach 2:
The patent changes the detection parameters from fixed rules to learned parameters through machine learning. The model automatically adjusts its internal parameters based on training data, allowing it to adapt to various pattern deformations and defect conditions, thereby improving detection accuracy without requiring complex manual rule adjustments.
2Reliability
If the die-to-database comparison method is used, then repetitive defects can be detected, but corner rounds are erroneously detected as defects due to smoothing filter artifacts
Solution Approach 1:
The patent incorporates feedback mechanisms where the machine learning model continuously learns from training data that includes both perfect patterns and defective patterns. This feedback loop enables the model to distinguish between actual defects and artifacts like smoothing filter-induced corner rounds, improving both reliability and precision simultaneously.
Solution Approach 2:
The patent segments the edge detection task into multiple components by training the machine learning model on diverse datasets that separate different types of patterns and defects. This segmentation allows the model to learn specific features for different pattern types and defect types, enabling accurate differentiation between real defects and artifacts.
3Adaptability or versatility
If training data includes defective patterns, then more comprehensive data coverage is achieved, but the model learns incorrect edge detection patterns
Solution Approach 1:
The patent extracts and removes defective patterns from the training data before using it to train the machine learning model. This extraction ensures that only high-quality training data is used, preventing the model from learning incorrect edge detection patterns while still maintaining comprehensive coverage of valid pattern types through careful selection of representative samples.
Data Source
AI summary
The present invention relates to a pattern edge detection method applicable to a semiconductor inspection apparatus that performs a pattern inspection using pattern design data. This method includes: generating an image of a pattern; detecting an edge of the pattern on the image based on a reference pattern generated from design data for the pattern; repeating generating of an image of a pattern and detecting of an edge of the pattern on the image to produce training-data candidates including a plurality of images and corresponding pattern edges; determining training data by removing pattern edges and corresponding images from the training-data candidates, the pattern edges to be removed being pattern edges satisfying a predetermined disqualification condition; producing an edge detection model by machine learning using the training data; generating an image of other pattern; and detecting an edge of the other pattern on the image using the edge detection model.