Automated SRAF generation for 75-nm contact holes patterning using generative AI in 193-nm immersion lithography

Abstract

Contact holes are essential patterns in semiconductor manufacturing but exhibit smaller process windows and lower image contrast than line-and-space features, making their lithography particularly challenging. To avoid costly mask rewrites after AIMS inspection, we investigate a machine-learning framework that automatically detects and repairs problematic contact-hole layouts early in the design flow. We present an automated framework for contact-hole mask optimization under 193-nm ArF immersion lithography using PROLITH™ simulations. Mask layouts are procedurally defined in Python as numerical coordinates, enabling hole- only masks pattern synthesis with reduced human bias. A total of 4,941 75-nm contact-hole masks were generated and labeled with a 4:6 ratio of compliant to non-compliant patterns, based on rule-based lithographic criteria. Using the latest YOLOv12 detector as a binary classifier, we achieved approximately 94% accuracy. When a layout is classified as non-compliant, we invoke the inpainting component of a Latent Diffusion Model (LDM) trained only on compliant masks. Stage 1 optimizes a variational autoencoder (VAE) to learn a compact latent space, and Stage 2 trains a U-Net diffusion model in that space. After 1,000 epochs, conditioning on separately prepared hole-only masks, the model synthesizes high-fidelity SRAF-enhanced layouts, which are re-simulated in PROLITH™ to evaluate process window and confirm low MEEF. This combination of machine learning and generative inpainting enables efficient optimization of contact-hole assist features, reducing design time and mask-revision cost. The framework is extendable to high-NA EUV and multipatterning, and helps reduce human-induced mask errors while improving design-technology co-optimization and reliability in high-volume manufacturing.