Universal Deoxidation of Semiconductor Substrates Assisted by Machine Learning and Real-Time Feedback Control.

Area of Science:

• Semiconductor Manufacturing
• Materials Science
• Artificial Intelligence

Background:

• Substrate oxidation during semiconductor manufacturing degrades device performance.
• Optimizing deoxidation in molecular beam epitaxy (MBE) is challenging and inconsistent due to substrate and process variations.
• Current deoxidation methods rely heavily on expert experience, leading to variable outcomes.

Purpose of the Study:

• To develop an automated, accurate, and standardized deoxidation process for semiconductor manufacturing.
• To overcome the limitations of traditional, expertise-dependent deoxidation methods.
• To enable consistent high-quality fabrication of state-of-the-art devices.

Main Methods:

• Utilized a machine learning model integrating interpolation and vision transformer (Interpolation-ViT) techniques.
• Employed reflection high-energy electron diffraction (RHEED) videos as input for the model.
• Developed an automated deoxidation system within a controlled architecture.

Main Results:

• The Interpolation-ViT model accurately predicts substrate status for automated deoxidation.
• Demonstrated successful deployment of models trained on one MBE system to others with high accuracy.
• Standardized deoxidation temperatures across various equipment and substrates, improving process consistency.

Conclusions:

• The developed AI approach offers a standardized and reliable method for substrate deoxidation in semiconductor fabrication.
• This technology has the potential to revolutionize optoelectronic and microelectronic manufacturing processes.
• The findings pave the way for more consistent and efficient semiconductor device production.