Fast simulation method of lithography latent image on non-planar wafer for large area mask.

The wafer topography can cause undesirable light scattering during lithographic imaging, affecting the quality of the photoresist profile. Lithography rigorous simulation can predict latent images on non-planar wafers for large area masks, but it is time-consuming. This paper presents a one-to-many deep learning network designed to learn the 3D distribution of latent images distribution from masks and topography patterns. The network incorporates additional height labels into conditional generative adversarial networks (CGANs) to better control latent images at different heights. Initially, the library of the masks, topography patterns, and their corresponding latent images were constructed for network training and subsequent latent image retrieval. A large area layout divided into mask clips and topography clips based on local pattern features. The trained network is then employed to calculate local latent images that do not exist in the training library. Finally, all local latent images are synthesized to simulate the entire latent image. The proposed method is applied to lithography simulations for display panels, demonstrating high accuracy and achieving an acceleration factor of approximately 240 to 340 times compared to the rigorous simulation method. It also shows an improvement of about 2 to 3 times over latent images calculation based on the CGAN (LIC-CGAN) approach.