Corner-focused evaluation of AI-based design-to-real prediction for FPD photomasks

Abstract

We developed an AI-based photomask simulator that predicts lithographic deformation from design GDS data for Flat Panel Display (FPD) photomasks. The simulator converts design layouts into image form and reproduces their optical appearance consistent with 100× measured real images. A deep-learning U-Net with SPADE residual blocks enables spatially adaptive modulation that preserves fine geometry around corners and thin features. Evaluation uses 3 μm × 3 μm Regions of Interest (ROIs) centered at vertices whose incident edges fall within the 0.4–1.5 μm effective length range. The ROIs include diverse corner geometries—rectangular, concave (C-shaped), threshold-like, diagonal, and staircase forms—representing common FPD mask patterns. Some diagonal corners appear visually <0.4 μm due to pixel stepping yet retain equivalent edge lengths and are evaluated under the same criteria. With 34 nm/px resolution, a 0.01 μm deviation corresponds to ~0.3 px, providing a practical sc ale for interpreting errors. Corner fidelity is quantified using Δr (corner-rounding difference) and ΔEPE95 = measured − predicted. Results show Δr median 0.011 μm (IQR 0.005–0.023) and ΔEPE95 median 0.000 μm (IQR −0.005–0.016), indicating sub-pixel-level agreement well within ±0.02 μm tolerance. This demonstrates geometry-agnostic consistency across varied corner types and pattern sizes. The proposed corner-focused evaluation and AI-based simulator provide a reproducible foundation for more accurate OPC correction, inspection calibration, and design–manufacturing correlation in next-generation FPD photomasks.