Noise Reduction in SEM Images using Deep Learning

Abstract

Measurement of patterns formed on wafers is required for defect inspection in mass production and for pattern quality evaluation in research and development. Scanning electron microscope (SEM) images are used for pattern measurement. The number of SEM scans must be reduced because of the incidents such as reduced throughput and damage to the resist. However, frame average images from fewer SEM images are noisy, and the noise makes it difficult to measure the pattern. In our proposed method, a deep learning was trained to perform noise reduction to measure patterns from noisy SEM images. Denoised images using the proposed method were evaluated with a 256-frame average image as a pseudo-correction image. The evaluation was made with PSNR and SSIM image quality evaluation, and with RMSE and power spectral density (PSD) of edge positions estimated using the tool. The results of noise reduction of single-frame image with proposed method were PSNR 32dB, SSIM 0.91, and RMSE 0.43nm, and showed high image quality and high accuracy in edge position estimation. With proposed method, an unbiased PSD-like graph with no noise floor was obtained. In addition, there is no significant difference between PSD graphs using single-frame images and 16-frame average images. These results indicate that proposed method can effectively remove noise from a few-frame average images, and that the denoised images can be used for pattern measurement and roughness evaluation using PSD.