A Synthetic CAD-to-SEM Pipeline for Design-Based Semiconductor Defect Detection Using Structural Similarity Analysis | Chapter 3 | New Horizons of Science, Technology and Culture Vol. 9

 

The growing complexity of semiconductor manufacturing at advanced technology nodes has intensified the need for robust, automated defect inspection methods. Design-based inspection, which compares scanning electron microscope (SEM) images of fabricated wafers against their original computer-aided design (CAD) layouts, offers a powerful approach for detecting both systematic and random defects. However, the development and benchmarking of such inspection algorithms is severely constrained by the scarcity of labelled SEM defect datasets, which are expensive to produce and tightly guarded as proprietary by fabrication facilities. This chapter presents an end-to-end, open-source pipeline for generating synthetic paired CAD and SEM image datasets with controllable, ground-truth-labelled defects, and for performing automated defect detection through structural similarity (SSIM) analysis. The pipeline comprises five modular stages: (1) synthetic layout generation in industry-standard OASIS format, (2) paired CAD and SEM-like image rendering with physically motivated degradation models, (3) configurable synthetic defect injection, (4) phase-correlation-based image alignment followed by local SSIM computation and morphological post-processing for defect mask extraction, and (5) aggregate scoring and ranking of inspection sites by defect severity. The paired image rendering generates, for each inspection site, a clean binary CAD image and a SEM-like image with sequential physically motivated degradations simulating realistic electron microscopy effects. The defect injection module selectively modifies a configurable fraction of SEM images (default 10%) to introduce synthetic defects. The detection stage processes each matched pair of CAD and SEM images through four substeps: alignment, similarity mapping, thresholding, and morphological cleanup. Experimental results on a generated dataset of 200 image pairs demonstrate that the pipeline achieves a detection F1-score of 0.93 under moderate noise conditions and degrades gracefully as imaging noise increases. The fully reproducible, configurable nature of this toolkit makes it suitable for algorithm benchmarking, machine learning model pre-training, threshold optimisation studies, and educational demonstrations of semiconductor inspection concepts.

 

 

Author(s) Details

Balachandar Jeganathan
ASML, San Jose, CA, USA.

 

Please see the book here :- https://doi.org/10.9734/bpi/nhstc/v9/7212