Wafer Defect Classification Using Autoencoder Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional wafer defect detection methods face performance issues due to high-dimensional data and high computational complexity in large-scale datasets, leading to inefficiencies in semiconductor fabrication.
Innovation Solution
An inspection system utilizing a classification convolutional neural network (CNN) with an encoder and decoder, combined with clustering, to capture and classify wafer images, automatically extracting features and generating soft labels for improved defect detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image classification methods are used for wafer defect detection, then defect types can be identified, but performance deteriorates due to high-dimensional data and high computational complexity on large-scale datasets
Solution Approach 1:
The patent extracts only the most salient features from wafer images using a trained neural network model, rather than processing all high-dimensional pixel data. The feature extraction network identifies and outputs only the critical defect-related features, significantly reducing the data dimensionality while maintaining detection accuracy.
Solution Approach 2:
The patent segments the defect detection task into two distinct stages: (1) feature extraction using a trained neural network to identify salient defect features, and (2) classification using a simpler classifier on the extracted features. This segmentation allows each stage to be optimized independently, reducing overall computational complexity.
2Reliability
If conventional image classification methods with multiple training images are used, then classifier performance can be improved, but processing time and computational resources increase significantly
Solution Approach 1:
The patent performs preliminary feature extraction and selection during the training phase, creating a compact representation of defect features that can be quickly processed during inspection. The neural network is pre-trained to recognize salient features, so during actual defect detection, only the extracted features need to be classified, not the entire high-dimensional images.
Solution Approach 2:
The patent creates a simplified copy of the defect detection problem by extracting essential features that capture the critical information needed for classification. Instead of working with full-resolution images, the system works with a compressed feature representation that preserves defect characteristics while reducing computational burden.
3Measurement precision
If high-dimensional wafer image data is processed in detail, then defect detection precision can be maintained, but productivity decreases due to high computational complexity
Solution Approach 1:
The system extracts only the salient defect features from full wafer images, discarding redundant information. This feature extraction process maintains detection precision by focusing on critical defect characteristics while reducing the data volume that requires detailed processing, thereby increasing inspection throughput.
Solution Approach 2:
The patent transforms the problem from processing high-dimensional pixel data to processing low-dimensional feature vectors. By changing the parameter representation from raw image pixels to extracted feature values, the system achieves the same detection precision with significantly reduced computational requirements, enabling higher productivity.
Data Source
AI summary
An inspection system for determining wafer defects in semiconductor fabrication may include an image capturing device to capture a wafer image and a classification convolutional neural network (CNN) to determine a classification from a plurality of classes for the captured image. Each of the plurality of classes indicates a type of a defect in the wafer. The system may also include an encoder to encode to convert a training image into a feature vector; a cluster system to cluster the feature vector to generate soft labels for the training image; and a decoder to decode the feature vector into a re-generated image. The system may also include a classification system to determine a classification from the plurality of classes for the training image. The encoder and decoder may be formed from a CNN autoencoder. The classification CNN and the CNN autoencoder may each be a deep neural network.