This work develops an expectation-maximization algorithm for jointly estimating posterior distributions and mixed (additive and multiplicative Gaussian) noise parameters in Bayesian inverse problems. The authors apply their method to real-world applications in nanometrology, specifically EUV scatterometry for characterizing nanostructures. JCMsuite was used to simulate the complex optical forward model (solving Maxwell's equations) for a line grating with an oxide layer, generating the data necessary to train and validate their proposed deep learning framework.
P. Hagemann, et al. Mixed noise and posterior estimation with conditional deepGEM. Mach. Learn.: Sci. Technol. 5, 035001 (2024).