ML-OCD (Machine Learning Optical Critical Dimension) is the application of machine learning to scatterometry data analysis — using neural networks, random forests, or other ML models to replace or augment traditional RCWA-based library matching for faster, more robust extraction of structural parameters from optical spectra.

ML-OCD Approaches

• Direct Regression: Train a neural network to directly map spectra → geometric parameters — bypass library search.
• Hybrid: Use ML for initial parameter estimation, then refine with physics-based regression.
• Virtual Metrology: Train ML models to predict reference measurements (CD-SEM, TEM) from OCD spectra.
• Transfer Learning: Pre-train on simulation data, fine-tune on real measurement data for domain adaptation.

Why It Matters

• Speed: ML inference is orders of magnitude faster than RCWA library computation — real-time parameter extraction.
• Complex Structures: ML can handle structures too complex for tractable RCWA libraries — high-dimensional parameter spaces.
• Robustness: ML can learn to ignore systematic errors that confuse physics-based models — data-driven robustness.

ML-OCD is AI-powered scatterometry — using machine learning for faster, more robust extraction of critical dimensions from optical measurements.