Virtual Metrology for AI-Enhanced Semiconductor Manufacturing

In semiconductor manufacturing, precision and efficiency are paramount. Machine learning (ML) is increasingly being embedded into production workflows, with Virtual Metrology emerging as a transformative application. By predicting critical process outcomes—such as silicon oxide layer thickness during Chemical Vapor Deposition (CVD)—using sensor data instead of physical measurements, Virtual Metrology reduces costs, increases coverage, and enables faster feedback. However, inference stability remains a key challenge. ML models trained on over 100 input features (e.g., temperature, pressure, gas flow) can degrade over time due to changing environmental conditions or maintenance actions. These shifts can cause discrepancies between predicted and actual values, risking product quality. To address this, a hybrid control strategy is implemented. Partial physical sampling provides ground truth for residual monitoring, while Statistical Process Control (SPC) charts track prediction accuracy in real time. Alerts are triggered when anomalies occur, prompting corrective actions such as increased sampling or model retraining. This approach ensures that the benefits of Virtual Metrology—efficiency, coverage, and speed—are realized without compromising reliability. By embedding intelligent monitoring into the production line, manufacturers move toward smarter, safer semiconductor fabrication.