AI-based analysis algorithm incorporating nanoscale structural variations and measurement-angle misalignment in spectroscopic ellipsometry

Spectroscopic ellipsometry (SE) is a powerful, non-destructive technique for nanoscale structural characterization. However, conventional SE data analysis typically assumes perfectly periodic specimen structures, overlooking fabrication-induced structural variations and thereby reducing the accuracy of predicted structural parameters. We have developed an enhanced analysis framework that explicitly accounts for both nanoscale structural variations and measurement-angle misalignment by introducing the concept of an average Mueller matrix (MM), which represents statistical distributions of nanoscale structures. In addition, we introduce a high-throughput MM-generation neural network that enables rapid data preparation by approximating rigorous coupled-wave analysis (RCWA) simulations for large numbers of specimens across a broad range of structural parameters. The model achieves a mean-squared error of 9.99 × 10−8 MSE when validated against RCWA-simulated MM data for one-dimensional SiO2 nanogratings. Finally, we apply our analysis framework to experimentally measured MM data, achieving highly accurate dimensional predictions with errors below 0.4 nm when compared with structural parameters measured by scanning electron microscopy (SEM). We believe that this analysis algorithm significantly advances the potential for high-precision SE-based metrology in semiconductor, photonic, and display manufacturing.