Synthetic Die-to-Database (D2DB) Lithography Layout Dataset

This dataset contains 20,000 synthetic 128\u00d7128 layout-style images for D2DB inspection research. Each sample includes a binary CAD image (lines, rectangles, vias), a physics-rendered lithography-style image, and a SEM-like noisy image. Three classes are provided: clean (no defect), geom_var (geometric\/process variation only), and topo_defect (true topological defects: bridge, open, missing_via). Metadata (metadata.csv) lists ID, split (train\/val\/test), class, defect_type, and paths to cad\/physics\/sem PNGs. Generated with the accompanying Python script using only common libraries (numpy, OpenCV, Pillow, pandas). Suitable for benchmarking die-to-database and defect detection methods without proprietary layout data.