Enhancing CD-SEM accuracy with attention-boosted Noise2Noise model

Abstract

Semiconductor manufacturing relies on Critical Dimension Scanning Electron Microscopy (CD-SEM) for precision in resist pattern measurements. High-resolution CD-SEM images, while desirable, can damage the resist due to increased electron beam exposure with higher frame numbers. To address this, Noise2Noise, a deep-learning noise reduction method, is introduced. Noise2Noise employs multiple noise images for unsupervised noise reduction. However, it struggles with unknown samples and limited training data. This research enhances the Noise2Noise model by introducing Attention and Residual-Recurrent structures to extract high-precision images from low-resolution inputs (1 frame). The Attentionboosted Noise2Noise model in particular exhibits superior accuracy with improved Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) for unseen patterns. Overall, the modeling error characterized by (ΔCD/CD) has been reduced compared to the conventional Noise2Noise method, promising improved CD-SEM accuracy for advanced CMOS manufacturing.