Fast source mask optimization adopting mask prediction and feedback method with similarity penalty.

Source mask optimization (SMO) based on gradient descent is a widely utilized technique in computational lithography. However, it is crucial to enhance the optimization efficiency of SMO, especially at advanced nodes. A mask predictive feedback with similarity penalty term SMO method is developed in this paper. This method incorporates a similarity penalty term into the loss function. It also employs a mask prediction feedback (MPF) method to more effectively utilize the mask and its gradient information during the iterative process. The simulation results indicate that the proposed method can reduce running time by at least 30% compared to using the Adam optimizer while ensuring target optimization conditions. Additionally, ablation experiments reveal that the proposed method improves algorithm efficiency more effectively than using the MPF method or similarity penalty term individually.