virtual metrology

Virtual metrology predicts wafer measurement results from process tool sensor data without physical measurement, enabling faster feedback and reduced metrology cost. Concept: process sensor data (trace data) contains information about wafer outcomes—build regression models to predict metrology values. Applications: (1) CD prediction—predict critical dimension from etch tool sensors; (2) Film thickness—predict thickness from CVD/PVD sensor data; (3) Sheet resistance—predict Rs from implant or anneal data; (4) Overlay—predict alignment from scanner sensor data. Model types: (1) Linear models—PLS (partial least squares) widely used for interpretability; (2) Nonlinear—neural networks, random forests for complex relationships; (3) Hybrid—physics-informed models using process knowledge. Implementation steps: (1) Collect paired data—sensor traces + metrology measurements; (2) Feature extraction—summarize traces into model inputs; (3) Model training—regression model development; (4) Validation—test on held-out data, production validation; (5) Deployment—real-time prediction, health monitoring. Benefits: (1) 100% wafer prediction (vs. sampled metrology); (2) Faster feedback—predictions available immediately; (3) Reduced metrology tool load; (4) Enable tighter APC—every wafer adjustment. Challenges: model drift requiring recalibration, chamber-to-chamber differences, handling process changes. Adoption growing in advanced fabs as key enabler for APC and yield improvement with reduced cycle time.