Improved manufacturability in curvilinear OPC for random DRAM contacts

Guaranteeing high pattern fidelity and good pattern shape for random DRAM contacts has been one of the most challenging tasks for OPC (Optical Proximity Correction), and, as patterning pitch shrinks, curvilinear OPC is regarded as the required solution for this layer. One of the challenges of the approach is the long runtime of curvilinear OPC, and another issue is to avoid any SRAF(Sub Resolution Assist Feature) printing due to aggressive curvi SRAF placement. In this paper, curvilinear OPC with machine learning (ML) approach is demonstrated to show affordable OPC runtime for random DRAM contacts. A new pattern selection method is developed to identify unique pattern sets from full-chip. After the ML OPC model is applied, a simpler OPC engine is applied to ensure the good convergence and MRC(Mask Rule Check) resolution. This flow showed improved litho performance over regular OPC. For major critical patterns, the DOF (Depth Of Focus) improved by more than 33% by the curvilinear OPC. To avoid any SRAF printing, SRAF patterns are added during the resist modeling process. This resist model shows no SRAF print defects compared to conventional method.