CNN-Based Defect Classification for Semiconductor Wafer Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As semiconductor device dimensions shrink, defect detection becomes increasingly challenging due to smaller defects causing unwanted aberrations, and existing attribute-based image fusion methods are inefficient in differentiating between defects of interest and nuisance data, often requiring perfect image alignment.
Innovation Solution
The method involves identifying a primary scanning mode through a hot scan, performing additional scans with secondary modes that deviate in focus, aperture, spectrum, or polarization, and using a convolutional neural network (CNN) to classify defects by training on image sets and determining an ideal secondary scanning mode for improved separation between defects and nuisance data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If attribute-based image fusion is used to differentiate defects from nuisance data, then defect detection capability is improved, but the method requires perfect image alignment and is time-consuming for mode selection
Solution Approach 1:
The patent transforms the mode selection problem from a manual attribute-based process to an automated parameter-driven process. The CNN model takes image data as input and automatically outputs the optimal mode pair selection, changing the decision parameters from manual inspection to algorithmic determination. This resolves the contradiction by eliminating the time-consuming manual mode selection while maintaining defect detection capability through automated image analysis.
Solution Approach 2:
The patent replaces the manual mechanical process of mode selection with an automated computational system. Instead of manually selecting modes based on attribute analysis, the system uses a CNN-based automated workflow that processes images and determines optimal mode pairs algorithmically. This substitution eliminates the time loss associated with manual mode selection while preserving the defect differentiation capability.
2Measurement precision
If multiple scanning modes are used for image fusion, then separation between defects and nuisance data is improved, but image alignment requirements become more stringent
Solution Approach 1:
The patent performs preliminary image processing and alignment corrections before the fusion process. The automated CNN-based workflow includes preprocessing steps that prepare images for fusion, handling alignment issues in advance. This preliminary action resolves the contradiction by preparing images to meet alignment requirements before fusion, enabling multi-mode fusion to achieve better defect separation without requiring manual intervention for alignment.
Solution Approach 2:
The patent introduces an automated CNN-based intermediary system that mediates between multiple scanning modes and the final defect classification. This intermediary automatically handles the complexity of aligning and fusing multiple modes, transforming the raw multi-mode images into properly aligned and fused results. The intermediary resolves the alignment precision requirement by automatically managing the alignment process throughout the fusion workflow.
3Productivity
If simple attribute-based fusion is used, then processing speed is maintained, but defect classification accuracy deteriorates
Solution Approach 1:
The patent implements a dynamic processing workflow that adapts to each defect case. Rather than using a static simple fusion method, the CNN-based automated workflow dynamically selects and applies appropriate processing steps based on the input images. This dynamic approach maintains processing speed by automating decisions while improving classification accuracy through adaptive image analysis and intelligent mode pair selection.
Data Source
AI summary
Systems and methods for classifying defects using hot scans and convolutional neural networks (CNNs) are disclosed. Primary scanning modes are identified by a processor and a hot scan of a wafer is performed. Defects of interest and nuisance data are selected and images of those areas are captured usa7ing one or more secondary scanning modes. Image sets are collected and divided into subsets. CNNs are trained using the image subsets. An ideal secondary scanning mode is determined and a final hot scan is performed. Defects are filtered and classified according to the final hot scan and the ideal secondary scanning mode. Disclosed systems for classifying defects utilize image data acquisition subsystems such as a scanning electron microscope as well as processors and electronic databases.