AI-initialized level-set inversion for lithographic mask reconstruction

As feature sizes in semiconductor manufacturing continue to shrink, accurate mask inspection and wafer-level prediction have become increasingly challenging. This paper presents a lithography-driven mask reconstruction framework that infers physically meaningful mask patterns from aerial images captured by mask reviewers. The proposed approach is grounded in an image formation model based on stacked pupil shift matrices and ensures physical interpretability and alignment with real lithography processes. The framework integrates a level-set-based inverse modeling approach with adaptive time-step optimization methods, including Barzilai–Borwein method and Golden Section Search, to ensure convergence efficiency and stability. To address the sensitivity of level-set methods to initialization, a deep learning-based model trained on lithography-aware data is introduced to generate accurate initial level-set functions. Additionally, an upsampling technique is employed to overcome pixel resolution limitations and to refine mask edge smoothness without increasing runtime. Experimental results demonstrate that the reconstructed masks generate aerial images that closely match those from mask reviewers. Compared with the sidelobe search, our AI-initialized method substantially improves reconstruction accuracy and convergence, especially in cases involving subresolution assist features. Furthermore, wafer-level evaluations exhibit strong alignment between simulated and actual CD variations, and matching slopes are consistently above 0.8. The proposed framework effectively bridges the gap between aerial image analysis and wafer behavior prediction, and offers a robust, scalable solution for advanced mask review and verification workflows.