Trans-UNet for fast and accurate photoresist modeling with variable development time

Optical projection lithography is widely used for manufacturing integrated circuits and producing advanced optoelectronic components, micro-electromechanical systems (MEMS), and emerging flat optics. Lithography simulation serves as a critical tool to verify and improve the patterning fidelity and final production yield of the lithography process. Conventional photoresist models fail to achieve both high speed and high accuracy in simulation, while most of the existing deep-learning methods are limited to predicting the photoresist contours at a constant development time. This work develops, to our knowledge, a novel deep-learning model called TransUNet to rapidly generate the photoresist contours with varying development times based on the input aerial images. Numerical experiments demonstrate that the proposed method improves the computational efficiency by more than 1000-fold over the rigorous model. Additionally, the proposed method can accurately calculate the resist contours at multiple development moments and outperforms other comparative methods in prediction performance.