Semiconductor Defect Inspection Using Latent Image Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect inspection methods for semiconductors, such as die-to-database and die-to-die inspections, are limited in their ability to accurately detect defects due to noise and manufacturing errors, especially with the miniaturization of semiconductor circuit patterns, and require design data, which may not always be available.
Innovation Solution
A defect inspection apparatus that calculates feature values from captured images, reduces information quantity using latent variables, estimates image statistics, and detects defects without relying on design data, utilizing a combination of convolutional neural networks for feature extraction and information reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel comparison-based inspection methods (die-to-database or die-to-die) are used, then inspection can be performed, but detection accuracy deteriorates due to noise and manufacturing errors
Solution Approach 1:
The patent introduces a reference image from a different wafer as an intermediary for comparison. Instead of directly comparing the inspection image to design data or the same wafer's reference, the system uses a reference image from another wafer that has undergone the same manufacturing process, thereby accounting for manufacturing variations and noise while still enabling defect detection
Solution Approach 2:
The patent changes the comparison parameter from direct pixel values to difference images and statistical metrics. By calculating the difference between the inspection image and reference image, and then computing statistical parameters (mean, standard deviation, skewness, kurtosis) of the difference image, the system transforms the inspection approach to be more robust against noise and manufacturing errors
2Reliability
If distribution comparison method using design data is used, then robust inspection against manufacturing errors is achieved, but the method cannot be applied when design data is unavailable
Solution Approach 1:
The patent enables the inspection system to be self-sufficient by using only captured images from wafers without requiring external design data. The system generates its own reference image from a different wafer and performs statistical analysis on the difference image, allowing it to adapt to any wafer regardless of design data availability
Solution Approach 2:
The patent creates a reference image by capturing a different wafer that serves as a copy or representative sample of normal manufacturing variations. This reference image acts as a template for comparison, allowing the system to identify defects without needing the original design data, thus copying the essential characteristics of normal variations from one wafer to another
3Adaptability or versatility
If die-to-die inspection is used, then inspection can be performed without design data, but detection accuracy deteriorates due to noise vulnerability
Solution Approach 1:
The patent transitions from direct image comparison to statistical analysis of difference images. By computing statistical parameters (mean, standard deviation, skewness, kurtosis) of the difference image, the system adds a dimensional transformation that enhances defect detectability while maintaining adaptability to wafers without design data
Data Source
AI summary
A defect inspection apparatus includes: a feature value calculation unit calculating a feature value based on a captured image of a sample; an image information reduction unit generating a latent variable by reducing an information quantity of the feature value; a statistic value estimation unit estimating an image statistic value that can be taken by a normal image based on the latent variable; and a defect detection unit detecting a defect in an inspection image based on the image statistic value and the inspection image of the sample.