GLAU-Net: a fast lithography imaging model for chemically amplified resists

Lithography simulation based on rigorous models is essential but time-consuming for developing and optimizing the chip manufacturing process. Exposure and post-exposure bake (PEB) are crucial steps to generate the PEB latent images, which directly determine the final photoresist profile. This paper proposes a novel, to our knowledge, deep-learning-based lithography imaging model, dubbed GLAU-Net, to accelerate the simulation efficiency of exposure and PEB steps for the extensively used chemically amplified resist. The network incorporates a mask attention gate module that dynamically assigns higher weights to the critical mask regions. A composite loss function combining the mean-squared error with a gradient loss is proposed to improve the prediction accuracy on the latent image boundaries. Experiments show that the GLAU-Net accelerates the simulation efficiency by 235 times and four times compared to the CPU-based and GPU-accelerated rigorous models, respectively. In addition, the proposed method outperforms the other popular deep-learning models in terms of prediction accuracy on the latent image and post-development photoresist profile.