Deep learning based computational imaging and optical metrology

Abstract

Manufacturing computer chips involves creating incredibly small and intricate patterns on silicon wafers. Ensuring the accuracy of these microscopic features is critical to the performance and reliability of the chips. Optical scatterometry, a technique that uses light scattering to measure these features, is essential for this quality control process. However, traditional scatterometry methods face challenges: iterative fitting methods are slow, while library search methods require enormous amounts of data. This project investigated the use of deep learning to overcome these limitations and significantly accelerate optical scatterometry. Deep learning, a powerful form of artificial intelligence, can learn complex relationships from data. In this project, a specific type of deep learning network called ResNet was trained using simulated optical measurements. These simulations mimic how light scatters off the microscopic features on a chip. By analyzing this scattered light, the ResNet learns to predict crucial parameters describing the features' shape and dimensions, such as their width, height, and the angles of their sidewalls. The research compared various ResNet architectures and strategies for handling the prediction process. One strategy, called UniNet, uses a single network to predict all parameters simultaneously. Another, called MonoNet, uses separate networks for each parameter, predicting them one at a time. A third approach, ExpertNet, uses a combination of networks to predict groups of related parameters. The findings showed that MonoNet offers the best performance overall, successfully decoupling the parameters and achieving high accuracy with smaller datasets. Further more, real-world measurements are always affected by noise and uncertainties. This research explicitly examined the impact of noise on the deep learning models' performance. By adding simulated noise to the training data, the researchers could assess the robustness of the different network architectures. The results provide valuable insights into developing more reliable and accurate measurement techniques, even in the presence of real-world imperfections.