Analytics-statistics mixed training and its fitness to semisupervised manufacturing

Area of Science:

• Semiconductor Manufacturing
• Machine Learning
• Computational Physics

Background:

• Machine learning (ML) models are increasingly used for predicting semiconductor manufacturing outcomes.
• Traditional ML models often struggle with anomalous effects influenced by specific fabrication environments.
• Technology computer-aided design (TCAD) offers robust physical modeling but has limitations in capturing all real-world variations.

Purpose of the Study:

• To integrate TCAD physical models into the ML training procedure for improved prediction accuracy.
• To develop an analytics-statistics mixed training (ASMT) approach for semiconductor process optimization.
• To demonstrate the effectiveness of ASMT in reducing prediction errors compared to standard ML methods.

Main Methods:

• Developed an analytics-statistics mixed training (ASMT) approach incorporating TCAD models into ML training.
• Created a mixed dataset combining experimental results with TCAD simulation data.
• Applied the ASMT scheme to the BOSCH process for performance evaluation.

Main Results:

• The ASMT approach, using a mixed dataset, significantly improved prediction accuracy.
• Compared to a classic neural network (NN) baseline, ASMT effectively decreased the mean square error (MSE).
• The integration of TCAD physical models enhanced the ML model's ability to predict process outcomes.

Conclusions:

• The ASMT method offers a powerful hybrid approach for intelligent semiconductor manufacturing.
• Combining statistical induction (ML) with analytical deduction (TCAD) leads to more accurate predictions.
• This integrated methodology is crucial for advancing the precision and efficiency of future semiconductor fabrication.