Cyclegan-Based Mask Diffraction Model

Modeling of photomask nearfield in extreme ultra-violet (EUV) lithography is one of the fundamental tasks for process modeling and physical verification. Rigorous electromagnetic field (EMF) simulation tools are too slow for large area applications, necessitating approximations for computational lithography applications. We duly propose, in this paper, a fast learning-based mask diffraction model with cycle generative adversarial network (CycleGAN) emphasis for EUV mask nearfield simulation. The network parameters are trained in a supervised manner. The proposed CycleGAN network learns the forward and inverse mapping between the photomask and the nearfield in a supervised manner by incorporating cycle consistency loss combinatorically with generative loss functions in both ways. We do not discriminate the learning of the real part and imaginary part of the nearfield by training them simultaneously. Investigation of the weighting the forward and inverse generative loss is also conducted. Simulation results merit fast and accurate inference from photomask to the nearfield and vice versa. It is also shown that different weights, in general, do not undermine the overall performance.