Definition
Virtual metrology (VM) is a data‑driven methodology that predicts the outcomes of physical measurement processes (metrology) by using statistical or machine‑learning models, thereby providing estimated metrology values without performing the actual measurement on the product or component.
Overview
In semiconductor manufacturing, lithography, and other high‑precision fabrication environments, metrology is essential for assessing critical dimensions, layer thicknesses, and other parameters that affect device performance. However, direct metrology can be time‑consuming, costly, and may introduce contamination or damage. Virtual metrology addresses these limitations by constructing predictive models from historical process data, sensor readings, equipment settings, and prior metrology results. Once validated, the models generate “virtual” metrology data in real time, enabling faster process control, yield prediction, and equipment optimization. VM is often integrated into advanced process control (APC) loops, statistical process control (SPC) frameworks, and closed‑loop manufacturing architectures.
Etymology / Origin
The term combines “virtual,” meaning simulated or occurring in a non‑physical domain, with “metrology,” the science of measurement. The concept emerged in the early 2000s as semiconductor fabs sought to increase throughput and reduce the cost of in‑line metrology. Academic publications and industry white papers from that period first described VM as a predictive approach leveraging multivariate regression, neural networks, and other machine‑learning techniques.
Characteristics
| Characteristic | Description |
|---|---|
| Data‑centric | Relies on large volumes of historical process and metrology data collected from sensors, equipment logs, and prior measurement results. |
| Model types | Includes linear and nonlinear regression, partial least squares (PLS), support vector machines (SVM), deep neural networks, and ensemble methods. |
| Real‑time capability | Designed to produce predictions at the same rate as the manufacturing line, often within seconds or less. |
| Validation | Models are routinely validated against a subset of actual metrology data to monitor prediction accuracy and drift. |
| Integration | Typically embedded within APC or SPC software platforms and may feed downstream control actions (e.g., recipe adjustments). |
| Benefits | Reduces metrology tool usage, shortens cycle time, lowers consumable costs, minimizes wafer handling, and improves overall equipment effectiveness (OEE). |
| Limitations | Prediction accuracy depends on the quality and representativeness of training data; models may require periodic retraining to accommodate equipment aging or process changes. |
Related Topics
- Metrology (manufacturing) – The practice of measuring physical dimensions and properties of manufactured parts.
- Advanced Process Control (APC) – Automated control strategies that use real‑time data to optimize manufacturing processes.
- Statistical Process Control (SPC) – The use of statistical methods to monitor and control a process.
- Machine learning in manufacturing – Application of algorithms that learn patterns from data to improve process decisions.
- Digital Twin – A virtual replica of a physical system used for simulation, analysis, and optimization, of which VM can be considered a specific implementation focused on metrology.
Note: The description above reflects the current understanding of virtual metrology as documented in scholarly articles, industry technical reports, and standards from semiconductor manufacturing and related fields.