Artificial Wafer Defect Map Generation for CNN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models, specifically CNNs, face challenges in accurately classifying wafer defect maps due to small and imbalanced data sets, leading to overfitting and inadequate coverage of defect classes, particularly when dealing with unclear SEM images and diverse defect patterns.
Innovation Solution
The generation of an artificial data set based on the characteristics of wafer defect classes, using techniques like Python or GUI tools, which is then augmented to create a more robust training set for CNNs, allowing for better prediction accuracy across a high number of classes and improving defect classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If actual wafer defect map data is used for training, then the model learns from real defect patterns, but the data set is small and imbalanced leading to overfitting and inadequate coverage of defect classes
Solution Approach 1:
The patent generates artificial wafer defect map images that copy the essential characteristics and patterns of real defect maps. These synthetic images replicate various defect types, patterns, and visual features observed in actual semiconductor wafer inspection data, enabling the training of deep learning models with a large, balanced data set without relying solely on limited real defect samples
Solution Approach 2:
The patent employs data augmentation techniques that transform existing defect map images by applying various parameter changes including rotations, flips, scaling, cropping, and other geometric transformations. These parameter modifications create diverse variations of original defect patterns, effectively expanding the training data set size and improving model generalization while preserving the underlying defect characteristics
2Quantity of substance
If data augmentation is applied to increase data set size, then more defect classes are covered, but the data remains imbalanced and overfitting persists
Solution Approach 1:
The patent performs preliminary classification and analysis of real defect data to identify the distribution, types, and characteristics of various defect classes before generating artificial data. This preliminary action enables targeted generation of synthetic defect maps that specifically address imbalanced classes, ensuring that artificial data is created to complement and balance the real data distribution rather than uniformly augmenting all classes
Solution Approach 2:
The patent applies sophisticated data augmentation transformations including geometric operations (rotation, flipping, scaling), photometric transformations (brightness, contrast adjustments), and advanced techniques like mixup and cutmix. These parameter changes create diverse and realistic variations of defect patterns that maintain class balance and prevent overfitting by exposing the model to broader variations of each defect type
3Adaptability or versatility
If a high number of defect classes are classified, then comprehensive defect identification is achieved, but model accuracy decreases due to limited data per class
Solution Approach 1:
The patent generates artificial defect map images for each defect class in the multi-class classification problem, creating synthetic samples that replicate the visual characteristics and patterns of rare or underrepresented defect types. This copying approach enables the model to learn from sufficient examples across all defect classes without requiring a proportional increase in real defect data for each class
Solution Approach 2:
The patent employs class-specific data augmentation strategies that apply parameter transformations tailored to each defect class characteristics. This includes using appropriate geometric and photometric transformations that preserve the distinctive features of each defect type while creating sufficient training samples to maintain high classification accuracy across all classes
Data Source
AI summary
A device includes image generation circuitry and convolutional-neural-network circuitry. The image generation circuitry, in operation, generates a digital image representation of a wafer defect map (WDM). The convolutional-neural-network circuitry, in operation, generates a defect classification associated with the WDM based on the digital image representation of the WDM and a data-driven model generated using an artificial wafer defect digital image (AWDI) data set and associating AWDIs with classes of a defined set of classes of wafer defects. A wafer manufacturing process may be controlled based on the classifications of WDMs.