As feature dimensions shrink, the need for rapid inspection of resist becomes ever more important. This is particularly true in the EUV regime, where concerns over stochastic failures require the measurement of millions of printed features in order to assess defect frequency. Making these measurements on resist-coated wafers represents a distinct challenge, as the optical scattering cross section of thin EUV resist necessitates the use of scanning electron probes, which are intrinsically slow due to the relatively small area probed per unit time. Thus, it is desirable to scan as rapidly as possible. However, owing to electronic and shot noise in the electron microscope, these methods are themselves subject to stochastic effects, producing noisy readouts. These imaging stochastics cloud accurate metrology of the underlying pattern, in particular the ability to accurately and reproducibly extract metrics such as LER, LWR, PSDs, etc.
In this study, we examine several strategies to “denoise” micrographs, with an eye towards recovering the true roughness characteristics of the underlying feature. To that end, we first perform 3D stochastic simulations of photoresist materials using the Multivariate Poisson Propagation Model. Then, scanning electron probe images of these resist “samples” are simulated using a rigorous model. Noise is then be added to these images consistent with assumed electronic and shot noise characteristics of an electron beam scanner. Finally, we apply a neural net approach to denoising the images, with a comparison to standard linear filters. The quality of the denoiser is assessed with respect to accurate recovery of the true LER, LWR, and power spectral characteristics of the original micrograph.