Efficient Bilevel Source Mask Optimization

Resolution Enhancement Techniques (RETs) are critical for meeting the stringent requirements of advanced technology nodes in semiconductor manufacturing. Source Mask Optimization (SMO), a prominent RET methodology, simultaneously co-optimizes photomask patterns and illumination sources to maximize the lithographic process window; nevertheless, existing SMO approaches typically separate source and mask optimization into sequential steps, resulting in substantial computational overhead and lacking convergence guarantees. In this work, we present BiSMO, a unified bilevel optimization framework that formulates SMO as a nested optimization problem and leverages three novel gradient-based algorithms for efficient and provably convergent solution search. To further enhance practical performance, we introduce BiSMO-RL, a reinforcement learning–driven extension that dynamically adjusts bilevel optimization hyperparameters to accelerate convergence and improve solution robustness. Experimental results on representative lithographic patterns demonstrate that BiSMO-RL reduces critical lithographic error metrics by up to 40% and achieves more than an 8× improvement in computational efficiency compared to state-of-the-art RET methods. All code and experimental settings are open-sourced at https://github.com/dekura/BiSMO.git.