Monotonic Machine Learning (MML) for Curvilinear OPC Retargeting in NA0.33 EUV Single-Patterning 28nm Metal Pitch Curvilinear Logic Technology

Abstract

Retargeting bridges design intent and manufacturable wafer patterns by generating an optimized target prior to Optical Proximity Correction (OPC). In fabs, retargeting is often guided by rule tables calibrated from post-lithography and post-etch metrology. However, for advanced nodes, particularly under EUV patterning, optical and mask-related effects introduce spatially varying and orientation-dependent deviations, and the resulting lithography-to-etch bias becomes highly non-linear. Traditional rule-based retargeting therefore struggles to account for these effects, often leading to unpredictable failures such as micro-bridging or broken features. Moreover, the combination of complex pattern interactions in logic designs and the limited availability of gauge-based CD measurements makes exhaustive manual rule tabulation both impractical and inaccurate, especially for curvilinear retargeting. In representative curvilinear constructs, the lithography-to-etch bias can be strongly anisotropic. For example, line-end regions may shrink while adjacent line-body segments may expand, and curved connections may exhibit localized non-linear bias hotspots. This paper presents a cluster-based, Monotonic Machine Learning retargeting framework for curvilinear patterns. The method groups points of interest into representative clusters and predicts displacement to generate retargeted targets that better reflect spatially varying bias. Results on representative curvilinear patterns show good retargeting fidelity.