A hybrid system for detecting semiconductor wafer defects using modified MobileNet with multi-head attention

Area of Science:

• Semiconductor Manufacturing
• Computer Vision
• Machine Learning

Background:

• Semiconductor wafer defect identification is complex, challenging current systems with intricate patterns and long-range defect dependencies.
• Existing methods struggle to detect small defects and handle imbalanced datasets common in wafer manufacturing.

Purpose of the Study:

• To develop a hybrid system for improved semiconductor wafer defect detection.
• To address limitations in capturing complex defect patterns, long-range dependencies, and minor defects.
• To overcome class imbalance issues in defect datasets.

Main Methods:

• A modified MobileNet architecture utilizing Swish activation and multi-head attention for feature extraction.
• Integration of an error-correcting output code (ECOC)-based support vector machine (SVM) classifier to handle dataset imbalance.
• Application of histogram equalization to enhance the visibility of minor defects in wafer images.

Main Results:

• The proposed hybrid system achieved a testing accuracy of 98.55% on the WM-811K dataset.
• Superior performance metrics including AUC (99.74%), recall (93.34%), precision (95.64%), and F1-score (94.42%) were obtained.
• Ablation studies confirmed the effectiveness of each proposed modification.

Conclusions:

• The hybrid MobileNet-ECOC-SVM system significantly enhances semiconductor wafer defect identification accuracy and robustness.
• The modifications effectively address challenges like complex patterns, long-range dependencies, minor defects, and class imbalance.
• This approach offers a promising solution for quality control in semiconductor manufacturing.