Adaptive reinforcement learning for lithography optimization: a scalable AI-driven solution for next-generation semiconductor manufacturing

Abstract Semiconductor lithography, a pivotal process in integrated circuit (IC) fabrication, accounts for approximately 30% of production costs and faces significant challenges as feature sizes shrink to sub-nanometer scales. Optical diffraction and process-induced distortions complicate precise patterning, necessitating advanced techniques beyond traditional Optical Proximity Correction (OPC). Inverse Lithography Technology (ILT) offers a mathematically robust approach to enhance pattern fidelity, yet its high computational complexity limits scalability. We propose Adaptive Reinforcement Learning for Lithography Optimization (ARLO), a U-Net-based framework integrating self-attention mechanisms and reinforcement learning (RL) to iteratively optimize photomasks using real-time lithographic simulations. Evaluated on the LithoBench benchmark, ARLO achieves a 37.8% reduction in $$L_2$$ Loss and a 74.0% reduction in Process Variation Band (PVB) compared to GAN-OPC, alongside 14.7% and 9.1% $$L_2$$ Loss reductions and 51.3% and 37.1% PVB reductions versus Deep LithoNet (DLN) and RL-ILT, respectively. Despite a higher shot count (181.4% increase vs. GAN-OPC, 59.0% vs. DLN-1, 29.4% vs. RL-ILT), ARLO maintains a competitive runtime of 0.035 seconds per patch. These results position ARLO as a scalable, efficient solution for next-generation semiconductor manufacturing.