In the relentless pursuit of efficiency and precision, semiconductor manufacturing has become a proving ground for cutting-edge technologies. Among these, machine learning (ML) is increasingly being embedded into production. One promising application for ML is Virtual Metrology—a method that infers critical process outcomes using sensor data instead of physical measurements. But while the benefits are clear, ensuring the stability of these predictions over time remains a challenge.
Semiconductor Manufacturing Meets Machine Learning
Semiconductor fabrication is a complex, multi-step process involving extreme precision. One such step is Chemical Vapor Deposition (CVD), where thin layers of materials—like silicon oxide—are deposited onto wafers. Traditionally, the thickness of these layers is measured using physical metrology tools like ellipsometers, which is expensive and time-consuming. Moreover, only a small fraction of wafers (often just one per lot) is sampled, leaving gaps in process visibility.
Fig. 1: Conventional (physical) metrology steps in semiconductor manufacturing: Many processing steps need a physical metrology step afterwards – requiring significant resources and increasing production time as well as reaction time.
This is where ML enters the scene. By training models on historical sensor data—such as temperature, pressure, and gas flow rates—engineers can predict layer thicknesses. This technique, known as Virtual Metrology, leverages the rich sensor data collected during CVD processing to estimate the silicon oxide layer thickness. The ML models generate predictions based on over 100 input features including:
- Chamber temperature readings
- Pressure variations
- Gas flow rates and variations
The advantages are compelling:
- Cost Reduction: Fewer physical measurements mean lower operational expenses.
- Increased Coverage: Predictions are available for every wafer, enabling more complete process and quality control.
- Faster Feedback: Engineers can react to process deviations in near real-time.
However, this shift from physical to virtual measurement introduces a new set of challenges—chief among them, inference stability.
Why Inference Can Become Unstable
Despite the sophistication of ML models, they are not infallible. Even with 100+ input features, the model may not capture all variables influencing layer thickness. We identified two factors that can cause inference accuracy to degrade over time:
- Changing environmental conditions: An unmonitored change in environmental conditions, or possibly equipment aging, can slowly increase the difference between predicted and actual thickness—known as the residual.
- Maintenance actions: Maintenance actions, such as replacing or cleaning parts in the CVD chamber, can abruptly alter process dynamics, leading to sudden, unexpected jump in the residuals.
These instabilities pose a risk to product quality and process reliability, making it essential to monitor and control the model’s performance continuously.
Keeping the Residuals in Check
To mitigate inference instability, a robust stability control strategy is embedded into the production workflow. This includes:
- Partial Physical Sampling: A fraction of wafers is still measured using traditional metrology. These measurements serve as ground truth to compute prediction residuals.
- Statistical Process Control (SPC): SPC techniques are a well-established standard for quality control in semiconductor manufacturing. These techniques are now applied to inference residuals: The residuals are displayed in real-time in an SPC chart and monitored for large individual residuals or systematic trends or shifts over time.
- Real-Time Alerts: If anomalies are detected—such as a sudden jump in residuals—engineers are immediately notified to investigate and take corrective action, which might include increasing the sampling rate of traditional metrology or retraining the model on more recent data.
Fig. 2: ML prediction vs physical measurement: The SPC Chart shows mean, range and raw values (recorded from multiple wafer locations) of the residuals (given in Å) over time. Each dot indicates a wafer; each color a different CVD equipment.
This hybrid approach ensures that while the system benefits from the efficiency of Virtual Metrology, it doesn’t compromise on reliability or quality.
Toward Smarter, Safer Manufacturing
Embedding ML inference into semiconductor production lines is a significant step toward smarter manufacturing. Virtual Metrology offers transformative benefits, but its success hinges on maintaining prediction stability. By combining intelligent monitoring with traditional safeguards, manufacturers can harness the power of ML without losing control.
Blog signed by: ams-OSRAM team