Machine Learning-Based Soft Sensor for Real-Time Wire Bow Prediction in Diamond Multi-Wire Sawing.

Area of Science:

• Materials Science and Engineering
• Manufacturing Process Optimization
• Artificial Intelligence in Manufacturing

Background:

• Real-time wire bow monitoring is crucial for wafer quality and preventing wire breakage in diamond multi-wire sawing (MWS).
• Physical sensors are impractical in industrial MWS due to sludge, space constraints, and high maintenance costs.
• Existing monitoring methods lack reliability and cost-effectiveness in harsh MWS environments.

Purpose of the Study:

• To develop a novel data-driven soft sensor framework for predicting wire bow in MWS using machine learning.
• To address the limitations of physical sensors in industrial MWS environments.
• To provide a reliable and low-cost solution for real-time wire bow monitoring.

Main Methods:

• A feature engineering pipeline using variance thresholding and correlation analysis to select key process variables.
• Systematic evaluation of six machine learning algorithms, including eXtreme Gradient Boosting (XGBoost).
• Two-stage hyperparameter optimization for XGBoost and SHAP (SHapley Additive exPlanations) for model interpretability.

Main Results:

• The optimized XGBoost model achieved a high coefficient of determination (R2) of 0.992 and a mean absolute error (MAE) of 0.116 mm.
• Accurate wire bow prediction was demonstrated across spatially distributed positions (head, middle, tail) of the wire web.
• SHAP analysis provided insights into the mechanical dependencies influencing wire bow.

Conclusions:

• The proposed data-driven soft sensor framework offers a reliable and cost-effective alternative to physical sensors for wire bow monitoring in MWS.
• This approach significantly enhances the potential for real-time quality control and defect prevention in wafer manufacturing.
• The findings pave the way for improved automation and process stability in the semiconductor industry.