SwinT-ILT: Swin Transformer embedding end-to-end mask optimization model

BackgroundResolution enhancement techniques (RETs) are widely used to improve the quality of masks in lithography flow. Optical proximity correction (OPC), such as inverse lithography technology (ILT), improves mask printability, but conventional ILT suffers from computational overhead. Advanced learning-based methods accelerate the optimization process, but the quality of its mask could be more satisfactory to academia and industry.AimTo improve the quality of masks while accelerating the optimization process, we propose SwinT-ILT, an end-to-end ILT optimization framework embedded in the Swin Transformer.ApproachThe framework consists of a feature extraction module and a feature construction module. Leveraging the Swin Transformer, the feature extraction module extracts deep features, and the feature construction module reconstruct masks based on in-depth features. To enhance the resolution of feature maps without introducing noise, we incorporate a pixel shuffle layer into the feature construction module. Furthermore, we set a specific training objective that introduces the domain knowledge of the imaging system, thereby accelerating the convergence.ResultsQuantitative results show that our SwinT-ILT achieves exceptional mask printability with relative mask manufacturability in only 0.5 s turnaround time.ConclusionsOur work contributes to the industry by effectively reducing semiconductor manufacturing cycles and enhancing manufacturing quality.