Integrated curvilinear OPC and SRAF optimization through reinforcement learning

Sub-resolution assist features (SRAFs) are small extra patterns placed on the mask to enhance the main pattern fidelity. SRAFs help reduce edge placement error (EPE) and improve the process window by making the printed patterns less sensitive to variations in focus and exposure. While SRAFs are typically inserted before optical proximity correction (OPC) and left unmodified to reduce correction complexity, simultaneous adjustment of SRAFs and main patterns may yield superior accuracy and process windows. We propose a reinforcement learning (RL)-based method to optimize both curvilinear SRAFs and curvilinear main patterns. Initial SRAFs are identified using a gradient-based method. Both SRAFs and main patterns are modeled as B´ezier curves (defined by two endpoints and two intermediate points). At each RL iteration, an actor updates curve endpoints, followed by heuristic relocation of intermediate points. The critic evaluates resulting patterns, providing expected rewards to guide actor training. If the maximum vertex placement error (VPE) or process variation band (PVB) width remains high, additional SRAFs may be inserted with new gradient map; RL iterations then resume. Once the actor has been trained, it is applied to actual curvilinear OPC and SRAF optimization. Experimental results show that the proposed method reduces maximum VPE by 7.6% and maximum PVB width by 23.0% compared to curvilinear OPC with fixed SRAFs.