Core Concepts
Machine learning algorithms are crucial for efficient wafer defect identification in semiconductor manufacturing.
Abstract
This survey paper delves into methodologies using machine learning (ML) classification techniques to identify defects on wafers in semiconductor manufacturing. It provides a taxonomy of methodologies, observational and experimental evaluations, and discusses the future prospects of ML classification techniques for wafer defect identification. The content is structured around various classification methods like Convolutional Neural Networks (CNN), Residual Neural Networks (ResNet), Adversarial Training, XGBoost, Decision Trees, Support Vector Machines (SVM), Logistic Regression, K-Nearest Neighbor (KNN), Learning Vector Quantization (LVQ), Network with Self-Calibrated convolutions, Hopfield Neural Network (HNN), Adaptive Boosting (AdaBoost), Generative Adversarial Network (GAN), and SVM for multi-label classification. Each technique's components, rationale, conditions for optimal performance, research papers employing the technique, case studies, and applications are discussed comprehensively.
Methodologies:
- Convolutional Neural Networks: Efficiently extract features from wafer images.
- Residual Neural Networks: Hierarchical feature extraction process.
- Adversarial Training: Enhances robustness and accuracy through adversarially generated examples.
- XGBoost: Leverages gradient boosting framework for efficient defect identification.
- Decision Trees: Hierarchical model simplifies complex decision-making processes.
- Support Vector Machines: Supervised learning model differentiates between defect types.
- Logistic Regression: Models probability of specific defects based on features.
- K-Nearest Neighbor: Classifies based on similarity to previously classified examples.
- Learning Vector Quantization: Trains prototype vectors to represent defect categories.
- Network with Self Calibrated convolutions: Dynamically adjusts filters based on defect characteristics.
- Hopfield Neural Network: Content-addressable memory system for pattern recognition.
- Adaptive Boosting: Ensemble technique combining weak classifiers into a strong classifier.
Multi-Label Techniques:
- Generative Adversarial Network: Dual-network architecture enhances multi-label defect classification.
- Support Vector Machine: Categorizes multiple labels simultaneously using SVM's margin-based classification.
Stats
Chen et al [19] introduced a predefined CNN model along with transfer learning for wafer map defect pattern recognition.
Shen and Zheng [20] developed JFLAN using CNNs to extract transferable features of wafer maps.
A.R and James [21] automated wafer defect classification using CNN combined with a memristor crossbar structure.
Quotes
"Effective defect monitoring is vital for production yield in chip fabrication."
"Deep learning reduces the need for manual feature extraction."
"By utilizing advanced ML techniques, systems can discern imperceptible defects."