Using machine learning with optical profilometry for GaN wafer screening.

Area of Science:

• Materials Science
• Semiconductor Manufacturing
• Machine Learning

Background:

• Gallium Nitride (GaN) wafer manufacturing requires cost-effective screening methods to identify defects early.
• Current wafer characterization techniques, like optical profilometry, yield complex data difficult to interpret manually.
• Classical programming models for data interpretation are labor-intensive and require significant human effort.

Purpose of the Study:

• To develop an inexpensive wafer screening technique for GaN manufacturing.
• To leverage machine learning for interpreting optical profilometry data.
• To reduce manufacturing costs by preventing processing of defective GaN wafers.

Main Methods:

• Fabricated over 6000 vertical PiN GaN diodes across 10 wafers.
• Utilized low-resolution, wafer-scale optical profilometry data collected prior to device fabrication.
• Trained four distinct machine learning models to predict device outcomes.

Main Results:

• Machine learning models achieved 70-75% accuracy in predicting device pass/fail status.
• Wafer yield was predicted with an error margin of less than 15% for most wafers.
• Demonstrated the efficacy of using pre-fabrication optical data for quality assessment.

Conclusions:

• Machine learning effectively interprets optical profilometry data for GaN wafer screening.
• This method provides valuable feedback for the manufacturing process.
• The developed technique offers a cost-effective solution for improving GaN wafer quality control.