machine learning ocd

**ML-OCD** (Machine Learning-Based Optical Critical Dimension) is a **scatterometry approach that uses machine learning models trained on simulated or measured spectra** — replacing traditional library matching or regression with neural networks, Gaussian processes, or other ML models for faster, more robust CD extraction. **How Does ML-OCD Work?** - **Training Data**: Generate a large synthetic dataset using RCWA simulations (parameter → spectrum pairs). - **Model Training**: Train a neural network (or other ML model) to predict parameters from spectra. - **Inference**: The trained model predicts CD, height, SWA from a measured spectrum in microseconds. - **Uncertainty**: Bayesian ML methods provide prediction confidence intervals. **Why It Matters** - **Speed**: Inference in microseconds — faster than both library matching and regression. - **Robustness**: ML models handle noise, systematic errors, and model imperfections better than exact matching. - **Complex Structures**: Can handle structures too complex for traditional library/regression approaches (GAA, CFET). **ML-OCD** is **AI-powered dimensional metrology** — using machine learning to extract nanoscale dimensions from optical spectra faster and more robustly.