Neural network assisted multi-parameter global sensitivity analysis for nanostructure scatterometry

Abstract Optical scatterometry has been widely used for measuring periodic thin-film structures in a fast and non-invasive way. However, shrinking structure dimensions, along with increasing structural complexity, give rise to challenging such nanoscale scatterometry. For in-line nano-fabrication process control, it is critical to enhance the measurability of small feature change. Accordingly, there is a strong need to evaluate the global sensitivity performance of different measurement strategies for identifying the best scheme that exhibits a remarkable change of optical responses under small dimension fluctuation. Such analysis requires not only an appropriate sensitivity indicator but also time-intensive computation. This paper presents an efficient approach to multi-parameter global sensitivity analysis (GSA) for scatterometry. A neural network (NN) assisted, moment-independent, and adaptive GSA method is proposed. A computationally cost-effective deep neural network is developed as a surrogate model for circumventing the simulation-based forward modeling. A scatterometry configuration problem for typical grating nanostructures is studied to demonstrate the effectiveness and the generalization potential of our approach. Our results show that the proposed approach is a powerful tool for not only examining sensitivity performance but also expediting the GSA optimization process for fast scatterometry.