Beyond the Golden Die: A Hybrid Die-to-Database Deep Learning Framework for Systematic Defect Detection

Abstract

Traditional Die-to-Die (D2D) inspection methods struggle to detect systematic and die-repeating defects that commonly occur in advanced semiconductor manufacturing nodes. To address these limitations, this work proposes a hybrid Die-to-Database (D2DB) inspection framework that bridges the substantial domain gap between ideal GDSII/CAD layouts and noisy, proximity-distorted SEM imagery. The proposed three-stage Render-and-Compare pipeline integrates (i) global CAD–SEM alignment, (ii) hybrid physics-based rendering and GAN-based SEM synthesis, and (iii) a topology-aware Siamese comparator that distinguishes connectivity-related defects from benign process variation. Experiments conducted on a 20,000-pattern synthetic dataset demonstrate that the hybrid approach achieves a false-alarm rate below 4% and a topological defect recall of 93%, significantly outperforming physics-only, GAN-only, and direct CAD–SEM subtraction methods (J. Ma et al., 2023, Y. Nam et al., 2022, P. De Bisschop & E. Hendrickx, 2019, Mack, 2007, Isola et al., 2017, Koch et al., 2015). These findings indicate that the hybrid D2DB methodology provides a scalable and accurate alternative to D2D inspection for identifying systematic and electrically critical defects in next-generation semiconductor processes.