Extraction of roughness measurements from thin resists with low signal-to-noise-ratio (SNR) SEM images by applying deep learning denoiser

Abstract

Background: thin resists, below 30 nm, suffer from reduced imaging contrast and signal-to-noise ratio (SNR). This has an impact on the unbiased Line Width Roughness (uLWR) estimation, which becomes artificially low and hence is inaccurate when the SNR drops below certain limit, while reaching an accurate plateau value at higher SNR. Aim: improve the SNR of Scanning Electron Microscope (SEM) images in order to achieve a more reliable and robust roughness measurements on the thin resist, without having to increase the measurement electron dose. Approach: apply an unsupervised machine learning denoising algorithm to SEM raw images of thin resists on two different underlayers. The images were captured using different frame averaging (4, 8, 16, 32, and 64 frames). A systematic analysis is performed to compare the measurements before and after denoising. Results: after denoising, the SNR is improved, the mean CD stayed unchanged, and the roughness using smaller number of frames got closer to the accurate values, obtained using a larger number of frames. Conclusions: we have demonstrated that the use of a machine learning-based denoising algorithm enhances the SNR of SEM images without changing the mean CD and is beneficial for accurate and robust roughness measurements of thin resist.

e-beam metrology of thin resist for high NA EUVL

“…However, this approach would lead to extensive measurement times, reduced throughput, as well as possible damage to the material caused by the high dose. We investigated an alternative approach using artificial intelligence (AI) denoising to improve the SNR without increasing the dose, [16][17][18][19] using a U-Net architecture-based unsupervised denoising machine learning algorithm. In Fig.…”

“…• Yuta Kawamoto, Yuzuru Mizuhara, Makoto Suzuki, Wataru Mori, Toru Ishimoto, Takumichi Sutani, Shunsuke Koshihara (Hitachi) • Peter de Schepper (Inpria) • Andrew Cockburn, Aviram Tam, Jens Van Hoof, Yaniv Abramovitz, (AMAT) Some of the material in this paper has been previously published in the proceedings of SPIE Advanced Lithography and SPIE Photomask Technology and EUV Lithography in 2022. 4,9,11,16)…”

Universal denoiser to improve metrology throughput

“…SEM images are a critical input source for metrology and inspection in semiconductor processes, and efforts 1,2 to improve the quality of SEM images have been ongoing for an extended period. Moreover, recent efforts [3][4][5] have seen the emergence of deep learning-based methods as well.…”

“…Another approach 4 to enhancing the quality of SEM images for thin resists involves the application of an unsupervised denoising algorithm. This method aims to improve the signal-to-noise ratio (SNR) without increasing the measurement electron dose, addressing the issue of low SNR and its impact on accurate Line Width Roughness (LWR) estimation in thin resists.…”

“…Several studies [1][2][3][4][5] have been introduced to enhance images obtained from E-beam tools in the semiconductor field, most of which are inspired by general image denoisers designed for the RGB space and standard cameras in Figure 1: Overview of our method; the proposed universal denoiser takes low-framed images (Raw-N l F) as input and outputs high-framed images (Predicted-N h F), where N l is the number of frames in the low-framed input and N h is the number of frames in the high-framed output, designed to match the ground truth high framed images (GT-N h F). The input images can come from a variety of multiple domains, and our method effectively and robustly processes them through our novel conditioning scheme.…”