A diffusion model enhanced source optimization for fab productivity

In the photolithography process at advanced nodes, source optimization (SO) becomes crucial for enhancing the production yield of each mask. Due to the complexity of physical models as well as the multi-criteria global optimization per full-chip mask, the overall SO procedure for each mask is both time-consuming and resource-intensive in a production. To enhance the fab efficiency and productivity, we propose a new, diffusion-based SO procedure. The framework is established on a latent diffusion model focusing on text-to-image generation. We generated hundreds of layout-source pairs using an in-house physical SO tool as the training dataset, targeting on 28nm nodes. The process parameters and the layout characteristic data are organized as textual inputs to the diffusion model. They are then encoded for conditional image generation, with source images serving as the optimization reference. A complete training-and-inference pipeline is presented, along with visualization of testcases and additional observations.