Deep Learning Hybrid Architecture Based on Vision Transformer for Phase Analysis of Moiré Fringes

Overlay accuracy is a fundamental indicator of a photolithography machine performance. Misalignment between the mask and wafer is the main factor affecting overlay accuracy. The photolithographic alignment method, which uses Moiré fringes, is notable for its straightforward optical path and high precision. However, the alignment accuracy is significantly influenced by the Moiré fringe phase analysis algorithm. This paper proposes a hybrid deep learning architecture based on a Vision Transformer for Moiré fringe phase analysis. By training on various types of Moiré fringe datasets, the model can predict the fringe wrapping phase, allowing for the analysis of elements within the wrapping phase, including displacement information. This method combines the multi-head attention mechanism of a Vision Transformer with deep learning feature extraction to build a hybrid deep learning architecture. This model effectively learns the mathematical mapping between the Moiré fringe phase information and actual offset, accurately outputting true Moiré fringe phase data. Results show that despite the presence of Gaussian noise and tilted states, the hybrid architecture maintains a Root Mean Square Error (RMSE) within the range of 6–7 nm, and a Structural Similarity Index (SSIM) above 0.70, and Peak Signal-to-Noise Ratio (PSNR) is consistently maintained above 36. Consequently, the proposed model demonstrates superior robustness in handling noisy data compared to existing phase retrieval techniques. Additionally, the model has been optimized in structure to more efficiently extract phase information from complex Moiré fringe patterns. This study offers valuable insights for expanding Moiré fringe imaging applications.