Machine Learning Enhanced Optical Proximity Correction Modeling for High-NA EUV Lithography

Abstract

High-NA EUV lithography presents challenges not only due to anamorphic optics but also because of the demand for highly accurate Optical Proximity Correction (OPC) models. In this work, we investigate OPC Modeling for a high-NA EUV lithography process with an 18 nm minimum pitch line-space array structure with a minimum 9 nm after-development critical dimension (ADI-CD). To ensure robust pattern fidelity control, a CD deviation of less than 10% is typically required within the manufacturing process window. This translates to an ADI-CD error budget of under 1 nm, which must be allocated across all sources of variation, including OPC model inaccuracies. Achieving this precision requires an accurate topographical mask model that accounts for background reflections from different EUV absorbers. In addition, the novel metal-oxide photoresist (MOR) and development process introduces further complexity in metrology, imaging, and modeling. For example, long-range chemical flare, extending beyond the optical influence of EUV, must also be incorporated into the modeling framework as well. To address these challenges, machine learning (ML)-assisted OPC modeling provides a promising solution for achieving the required ADI-CD prediction accuracy. A proprietary convolutional neural network (CNN) primarily captures local effects within its receptive field, making it well-suited for modeling spatially localized proximity effects. Additionally, by integrating CNN-based proximity modeling with long-range effect information—such as flare maps—the ML model can effectively capture both local and global influences, resulting in improved overall accuracy. We demonstrate the importance of carefully optimizing the ML architecture and show that our approach can meet the stringent precision requirements for OPC in advanced technology nodes using high- NA EUV.