Virtual Metrology: Predicting Wafer Quality Without Physical Measurement
Virtual Metrology: Making Every Wafer “Inspected”
In semiconductor manufacturing, physical metrology is a critical step in ensuring wafer quality — using optical or electrical methods to measure wafers and verify whether process indicators such as film thickness, critical dimension (CD), and etch depth meet specifications. However, physical metrology has several unavoidable pain points:
- Time-consuming: A single measurement can take several minutes to tens of minutes, severely impacting production line throughput
- Expensive: Metrology equipment is costly to purchase and maintain
- Sampling-only: Limited by capacity, typically only 5%-10% of wafers are measured, leaving the majority in a “blind zone”
Virtual Metrology (VM) was developed to address these challenges. It leverages sensor data generated during the equipment processing cycle (such as temperature, pressure, gas flow rates, and RF power), combined with AI predictive models, to infer wafer process quality indicators in real time — providing a quality assessment for every wafer without the need for actual physical measurement.
How VM Works
The core approach of virtual metrology can be summarized in three steps:
- Data acquisition: Real-time collection of process parameters (FDC data) from equipment sensors during wafer processing, typically encompassing tens to hundreds of dimensions
- Feature engineering and modeling: Using historical data combined with a small number of actual physical metrology results as labels to train machine learning or deep learning models
- Online prediction: After a new wafer completes processing, the model outputs quality predictions at millisecond speed based on real-time sensor data
This means the production line can upgrade from “sampling inspection” to “full inspection” — every wafer has a corresponding quality prediction, providing the data foundation for subsequent R2R automatic tuning and anomaly detection.
Core Value of VM
1. Significantly Reduce Metrology Costs
By using VM to replace a portion of physical metrology, fabs can reduce the frequency and number of metrology tools required, alleviating the metrology capacity bottleneck.
2. Earlier Detection of Process Anomalies
VM delivers quality predictions the instant processing is complete, without waiting in the metrology queue. This means anomalous wafers can be intercepted earlier, reducing waste in downstream process steps.
3. Enabling Continuous Yield Optimization
Full-coverage quality prediction data gives process engineers richer analytical dimensions, helping to identify the root causes of yield fluctuations and accelerating process optimization.
Key Challenges in Deploying VM
While VM’s value is widely recognized, several challenges remain in practical deployment:
- Model accuracy: Sensor data noise and drift can affect prediction accuracy, requiring continuous model updates
- Equipment drift: As equipment ages or undergoes preventive maintenance (PM), process conditions change, and models must have adaptive capabilities
- Small sample problem: During the early stages of new process or new equipment ramp-up, there is insufficient labeled data to train high-quality models
Addressing these challenges requires introducing transfer learning and online learning mechanisms at the algorithm level, while at the engineering level, standardized equipment data collection must be achieved through the SECS/GEM protocol.
VM Best Practice: Edge Deployment
Deploying VM models at the equipment edge, rather than in a centralized cloud or MES system, is the current industry best practice. The advantages of edge deployment include: low-latency inference, on-premises data retention, and no disruption to existing IT architecture.
MST Semiconductor’s NeuroBox Edge Intelligence Platform is designed precisely for this scenario — supporting VM inference directly at the equipment level, delivering millisecond-speed full-wafer quality prediction.
