Virtual Metrology

Virtual metrology replaces physical metrology tools with data-driven inferential models that estimate wafer or product quality from upstream process variables. These models ingest high-dimensional sensor traces—such as chamber pressure, RF power, gas flow rates, and temperature profiles—to predict critical quality parameters like film thickness, etch depth, or defect density.

• Input space: Time-series data from equipment sensors, often summarized via feature engineering (mean, variance, slope, frequency components)
• Output space: Continuous quality metrics (thickness in angstroms) or categorical classifications (pass/fail)
• Model types: Partial Least Squares (PLS), Gaussian Process Regression, Random Forests, or deep neural networks
• Key advantage: Prediction occurs in seconds post-process, versus hours for physical metrology queue time

The core operational value is the ability to generate immediate quality estimates for every wafer or product unit without waiting for physical measurement. This enables closed-loop process control where deviations are detected and corrected within the same manufacturing lot.

• Inline deployment: Models execute on edge servers or directly within the equipment's computing environment
• Prediction triggers: Automatically invoked at process recipe completion, ingesting the just-generated sensor trace
• Decision integration: Predicted quality values feed directly into Run-to-Run (R2R) controllers to adjust recipe parameters for the next wafer
• Anomaly flagging: Units predicted to be out-of-specification can be immediately routed to rework or hold, preventing further value-add processing on defective material

Raw equipment sensor data is high-volume and noisy. A critical preprocessing step extracts statistically meaningful features that capture the process signature. This transforms a time series of thousands of data points into a compact, informative vector for the predictive model.

• Temporal features: Mean, standard deviation, maximum, minimum, range, and slope over key process steps
• Frequency-domain features: Spectral energy in specific frequency bands via Fast Fourier Transform (FFT) to capture oscillations or plasma instabilities
• Step-specific aggregation: Sensor data is segmented by recipe step (e.g., ramp, stabilization, main etch) and features are computed per step
• Correlation analysis: Features are ranked and selected based on their Pearson correlation with the target quality metric to reduce dimensionality and prevent overfitting

Virtual metrology models degrade over time as equipment ages, consumables wear, and process conditions shift. A robust model lifecycle management strategy is essential to maintain prediction accuracy.

• Drift detection: Continuous monitoring of prediction residuals against periodic physical metrology samples using Cumulative Sum (CUSUM) or Exponentially Weighted Moving Average (EWMA) control charts
• Automatic recalibration: Triggering model retraining or bias-update when drift exceeds a statistical threshold
• Chamber matching: Models must account for subtle differences between nominally identical chambers; transfer learning or chamber-specific offsets are common solutions
• Golden wafer verification: Periodic measurement of a known reference wafer to validate both the physical metrology tool and the virtual metrology model simultaneously

Virtual metrology fundamentally changes the economics of quality control by decoupling inspection frequency from metrology tool capacity. Physical measurements are strategically reduced to a supervisory role.

• 100% virtual inspection: Every unit receives a quality prediction, providing full lot traceability
• Risk-based physical sampling: Physical metrology is reserved for units where the virtual model reports high prediction uncertainty (e.g., via Gaussian Process variance estimates)
• Tool time reclamation: Expensive metrology tools are freed from routine monitoring and redeployed for engineering analysis or complex measurements that cannot be virtualized
• Cost reduction: Direct savings from reduced metrology tool utilization, maintenance, and consumables, often exceeding 30% reduction in metrology cost per wafer

Virtual metrology models are tightly coupled with Fault Detection and Classification (FDC) systems. While FDC monitors equipment health, virtual metrology monitors product health. Together, they provide a complete picture of process capability.

• Dual monitoring: FDC flags equipment excursions; virtual metrology flags product excursions. Cross-correlation enables root cause analysis
• Predictive maintenance trigger: A gradual degradation in predicted quality can signal an impending equipment fault before an FDC limit is breached
• Contextual data enrichment: Virtual metrology predictions are stored alongside FDC summaries in the manufacturing data lake, enabling long-term yield analysis and commonality studies across tools and products
• Run-to-run control synergy: The combination of FDC fault detection and virtual metrology quality prediction provides a robust feedback signal for automated process correction, preventing over-correction based on noisy or faulty data

A discrete feedback methodology that adjusts recipe parameters between processing batches based on post-process metrology to compensate for drifting tool conditions. Virtual metrology provides the critical quality signal for this loop when physical measurements are unavailable.

• Uses an Exponentially Weighted Moving Average (EWMA) filter to update process models
• Corrects for drift and shift disturbances in semiconductor fabrication
• VM enables wafer-to-wafer control instead of lot-to-lot control

A non-parametric Bayesian inference method that models a distribution over possible functions to provide both predictions and calibrated uncertainty estimates for process variables. GPR is a foundational algorithm for building virtual metrology models.

• Provides a confidence interval alongside every quality prediction
• Uses a kernel function to define the smoothness and similarity assumptions
• Enables active learning by identifying predictions with high uncertainty for physical verification

An optimal recursive algorithm that estimates the internal state of a linear dynamic system from a series of noisy sensor measurements by minimizing the mean squared error. In virtual metrology, Kalman filters fuse real-time sensor data with process models.

• Predicts the current state and then updates it with a new measurement
• Handles sensor fusion from multiple heterogeneous data sources
• Extended Kalman Filters (EKF) handle non-linear process dynamics common in etching and deposition

The continuous, bidirectional data flow mechanism that ensures a virtual representation of a physical asset accurately mirrors its real-time state, condition, and behavior. Virtual metrology is the quality dimension of a comprehensive digital twin.

• Combines 3D geometry, physics-based simulation, and data-driven VM models
• Enables what-if analysis for process parameter changes before physical execution
• Requires high-frequency sensor streaming and contextualized data models

A quality control methodology that uses statistical methods to monitor a process, distinguish between common cause and special cause variation, and signal when corrective action is needed. VM predictions feed directly into SPC charts.

• Control charts track VM-predicted quality metrics in real-time
• Western Electric rules detect out-of-control conditions automatically
• Enables fault detection and classification (FDC) without waiting for physical metrology

The engineering discipline of combining data from disparate sensors like RF power monitors, optical emission spectrometers, and pressure transducers to create a unified operational view. Virtual metrology is fundamentally a sensor fusion problem.

• Aligns time-series data from hundreds of equipment sensors to a common timestamp
• Handles missing data and varying sampling rates across sensor types
• Feature extraction identifies the most predictive sensor combinations for quality outcomes