On Uses of Noise Analysis for the Uncertainty Quantification of Line Edge Roughness Estimation

The high-volume manufacturing of the next generation of semiconductor devices requires state-of-the-art statistical analyses to reduce measurement uncertainties. Prediction intervals can characterize the reliability of the predictive performance of regression models. Normalized conformal prediction is a mathematically motivated approach to find distribution-free prediction intervals which depends on skillful modeling to evaluate a regression model fit to training data, and quantile regression is a classical and widely applied guideline to devise prediction intervals. Noise levels impact the reliability of line edge roughness estimates from noisy scanning electron microscope images. We propose procedures to generate prediction intervals based on image denoising and other image processing techniques and demonstrate significant improvements over earlier methods developed for the deep convolutional neural network EDGENet.