Investigating a Machine Learning Approach to Predicting White Pixel Defects in Wafers-A Case Study of Wafer...

Area of Science:

• Semiconductor Manufacturing
• Materials Science
• Machine Learning Applications

Background:

• CMOS image sensors (CIS) are crucial for mobile and photographic devices, with white pixels indicating performance issues.
• White pixels in CIS products stem from metal impurities or grinding blade defects during wafer production.
• Addressing grinding blade defects offers a practical approach to mitigating white pixel issues.

Purpose of the Study:

• To analyze white pixel defects in silicon wafer manufacturing in Taiwan.
• To utilize machine learning for identifying and predicting factors causing white pixel defects from grinding operations.
• To provide practical insights for reducing white pixel defects in CIS products.

Main Methods:

• Customer-reported white pixel defect data analysis.
• Application of machine learning models, including Classification and Regression Trees (CART) and Random Forest (RF).
• Identification of critical grinding parameters influencing white pixel defect formation.

Main Results:

• CART and RF models achieved 95.18% accuracy in predicting white pixel defects from grinding.
• Grinding load and torque were identified as critical factors in white pixel defect genesis.
• The study provides a rare, practical analysis of white pixel defects in a real-world manufacturing setting.

Conclusions:

• Machine learning models effectively predict white pixel defects in CIS manufacturing.
• Controlling grinding load and torque is essential for minimizing white pixel defects.
• Proactive measures based on these findings can reduce customer complaints and enhance CIS product quality.