Virtual Metrology

VM models ingest high-frequency equipment sensor data—such as chamber pressure, RF power, gas flow rates, and temperature profiles—collected during processing. A trained machine learning model (often a neural network or Gaussian process) maps these multivariate sensor traces to post-process quality metrics like film thickness, etch depth, or surface roughness. The prediction is generated immediately after processing, eliminating the latency of physical metrology and enabling real-time quality decisions.

In semiconductor manufacturing, physical metrology is typically performed on only a sampling of wafers due to cost and throughput constraints. VM provides predicted quality values for every wafer, feeding directly into Run-to-Run (R2R) controllers. This transforms sparse, lot-level feedback into dense, wafer-level feedback, allowing the control system to detect and correct process drift before it produces out-of-spec material.

VM models degrade over time as equipment ages, consumables wear, and process conditions shift. A robust VM deployment includes automated drift detection that compares VM predictions against occasional physical measurements. When prediction error exceeds a threshold, the system triggers model recalibration or retraining. Techniques like moving-window partial least squares and just-in-time learning adapt models without full retraining, maintaining accuracy in dynamic production environments.

VM is often deployed alongside Fault Detection and Classification (FDC) systems. While FDC monitors equipment health by detecting anomalous sensor patterns, VM translates those patterns into predicted quality outcomes. Together, they form a comprehensive monitoring framework: FDC flags the equipment fault, and VM quantifies the impact on product quality, enabling operators to make informed disposition decisions without waiting for physical metrology results.

Modern VM systems leverage several advanced techniques:

• Deep autoencoders for unsupervised feature extraction from high-dimensional sensor data
• Transfer learning to adapt models across similar tool types with minimal new training data
• Ensemble methods that combine multiple model types to improve prediction robustness
• Physics-informed neural networks that incorporate known process physics to constrain predictions
• Edge deployment on factory-floor compute for sub-second inference latency

VM delivers measurable value across the manufacturing enterprise:

• Reduced metrology cost: Fewer physical measurements required, lowering tool capital and labor expenses
• Increased throughput: Eliminates queue time waiting for metrology results before subsequent processing
• Scrap reduction: Early detection of quality excursions prevents processing of additional defective material
• Improved process control: Dense, real-time quality data enables tighter control limits and higher CpK values
• Enables advanced process control: Provides the quality feedback signal required for R2R and MPC strategies

A form of adaptive process control where recipe parameters are modified between processing runs based on post-process metrology. Virtual metrology provides the wafer-level quality estimates that feed the R2R controller, enabling corrections without waiting for physical measurements.

• Compensates for drift in tools like etchers and CMP
• Uses EWMA or model-based algorithms to update recipes
• Reduces reliance on send-ahead wafers

The practice of measuring workpieces or process conditions directly within the manufacturing equipment during or immediately after processing. Virtual metrology complements in-situ sensors by inferring quality parameters that cannot be directly measured, such as film thickness or etch depth, using available equipment data.

• Provides immediate data for closed-loop control
• Eliminates wafer breakage risk from handling
• Enables 100% inspection paradigms

A quality control methodology using control charts to distinguish between common-cause and special-cause variation. Virtual metrology feeds SPC systems with high-frequency, virtual measurements, enabling real-time process monitoring at a granularity impossible with sparse physical sampling.

• Western Electric rules detect out-of-control conditions
• VM enables site-level SPC across the wafer
• Reduces inspection sampling risk

A system that monitors equipment sensor traces in real-time to detect anomalies and faults. While FDC identifies equipment health issues, virtual metrology translates those same sensor traces into predicted quality outcomes, bridging the gap between tool state and product quality.

• Uses time-series pattern matching on sensor data
• VM adds the quality impact dimension to FDC alarms
• Enables predictive scrapping of affected wafers

A high-fidelity virtual representation of a physical asset that synchronizes with real-time data. A virtual metrology model is a specialized, quality-focused digital twin of a process chamber, predicting the physical outcome of a recipe based on the equipment's current state and sensor signatures.

• VM is the quality dimension of the process twin
• Enables what-if simulation for recipe changes
• Supports virtual design of experiments

The computational process of combining data from multiple disparate sensors to produce a more accurate and reliable estimate than any single sensor could provide. Virtual metrology is inherently a sensor fusion problem, integrating optical emission spectroscopy, RF match data, pressure, and temperature signals to predict a single quality metric.

• Handles heterogeneous data types and rates
• Kalman filters and neural networks are common approaches
• Improves prediction robustness against single-sensor failure