Physics-Informed Machine Learning for Mask Design in Interference Lithography

Interference lithography is a powerful technique for high-resolution patterning of periodic structures, but its use for non-periodic structures remains limited due to the challenge of mask design. Here, we present a machine learning model based on a U-Net architecture to predict binary mask design from aerial patterns. Due to the loss of phase information in aerial images, this is an ill-posed problem with multiple possible solutions. To address this challenge, we introduce a physics-informed loss function that incorporates angular spectrum wave propagation into training. Furthermore, we show that self-training with iterative bootstrapping enables generalization to artificial patterns with programmed line-breaking defects, which are not explicitly included in the training data. We observe that localized pattern modifications often require global changes in the mask layout due to the underlying Fourier relationship. This study provides a first step toward mask design for complex interference lithography applications.