Enhancing Confidence and Interpretability of a CNN-Based Wafer Defect Classification Model Using Temperature Scaling...

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Quality Control

Background:

• Accurate wafer defect classification is crucial for semiconductor yield and quality.
• Existing methods primarily focus on classification accuracy, neglecting confidence and interpretability.

Purpose of the Study:

• To develop a model for simultaneous assessment of accuracy, prediction confidence, and interpretability in wafer defect classification.
• To address the class imbalance issue in defect datasets.
• To enhance the practical applicability of AI models in semiconductor production.

Main Methods:

• Utilized a weighted cross-entropy loss function to handle class imbalance.
• Employed a convolutional neural network (CNN)-based model for classification.
• Applied temperature scaling to improve prediction confidence.
• Integrated Local Interpretable Model-Agnostic Explanations (LIME) and Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability.

Main Results:

• Achieved a high classification accuracy of 97.8% on the test dataset.
• Successfully enhanced prediction confidence using temperature scaling.
• Provided visualized explanations for model predictions, enabling user understanding of the decision-making process.

Conclusions:

• The proposed model offers a comprehensive approach to wafer defect classification by integrating accuracy, confidence, and interpretability.
• Explainable predictions enhance the trustworthiness and adoption of AI in semiconductor quality management.
• This research paves the way for next-generation intelligent quality management systems in semiconductor manufacturing.