Efficient deep-learning ResNet architectures for Mueller-matrix Fourier scatterometry

Mueller-matrix Fourier scatterometry is a highly sensitive technique for semiconductor chip inspection. However, traditional scatterometry faces challenges: iterative fitting is slow, and library search methods require vast datasets. Deep learning neural networks offers promising alternative but demands also extensive computational resources for dataset generation, which becomes impractical for large parameter spaces. This study explores the application of deep learning ResNet architectures on Fourier scatterometry. We will demonstrate that when the task of the neural network is divided into several subtasks, the network can efficiently predict six key shape and dimension parameters of silicon grating, including width, height, top and bottom rounding, and sidewall angles. To train the networks, two types of datasets are generated: Mueller matrices in the Fourier plane from the 0th diffraction order and from full diffraction orders. Three models are evaluated: UniNet (a single network predicting all parameters), MonoNet (separate networks for each parameter prediction), and ExpertNet (grouped parameter prediction). Results show that the ExpertNet has similar performance as the MonoNet. Using both models, we achieved superior accuracy for certain parameters with small datasets. Our analysis suggests that this improvement is due to a reduction of the input-output correlations between parameters with the neural network. Furthermore, all parameters benefit from dataset with full diffraction orders, except for the top rounding. These findings provide valuable insights for developing more reliable and accurate neural networks and measurement techniques for Mueller matrix Fourier scatterometry.