A Physics-Constrained, Design-Driven Methodology for Defect Dataset Generation in Optical Lithography

초록

Artificial intelligence (AI) based defect inspection in micro/ nano manufacturing is limited by the scarcity of high quality, physically grounded training data. To address this bottleneck, this study proposes a design-driven and fabrication-based methodology that generates lithographic defects through mathematically defined morphology operations and physically replicates them using digital micromirror device (DMD)-based maskless lithography. Bridge, burr, and pinch defects were systematically generated from defect-free layouts using mathematical morphology and physically replicated using DMD-based maskless lithography. The fabricated defect patterns were compared with their defect-free references to obtain pixel-level annotations. The resulting dataset contains 3,530 optical micrographs and 13,404 annotated defect instances, including morphology-generated bridge, burr, and pinch defects and naturally occurring contamination. Quantitative analysis of 3,000 matched defect pairs showed positive correspondence between designed and fabricated defect sizes, with a Pearson correlation coefficient of 0.760 and a Spearman rank correlation coefficient of 0.728. Five instance segmentation models were evaluated using a base-layout-disjoint split, achieving AP@0.50 values of 0.704 to 0.990 across the morphology-generated defect classes, while contamination remained more challenging. Qualitative evaluation on experimentally fabricated ICCAD-derived layouts showed that models trained on the constructed dataset could localize defect regions in previously unseen and more complex patterns. These results demonstrate that mathematically controlled design perturbations are measurably reflected in fabricated patterns and support the proposed methodology as a physically grounded approach to lithography inspection dataset construction. © 1988 IEEE.

키워드