A fast and manufacture-friendly optical proximity correction based on machine learning

Pixel-based optical proximity correction (PBOPC) is currently a key resolution enhancement technique to push the resolution limit of optical lithography. However, the increasing scale, density and complexity of modern integrated circuits pose new challenges to both of the OPC computational intensity and mask manufacturability. This paper aims at developing a practical OPC algorithm based on a machine learning technique to effectively reduce the PBOPC runtime and mask complexity. We first divide the target layout into small regions around corners and edge fragments. Using a nonparametric kernel regression technique, these small regions are then filled in by the weighted linear combination of a subset of training OPC examples selected from the pre-calculated libraries. To keep balance between the image fidelity and mask complexity, we use an edge-based OPC (EBOPC) library to synthesize the OPC patterns in non-critical areas, while use another PBOPC library for hotspots. In addition, a post-processing method is developed to refine the regressed OPC pattern so as to guarantee the final image fidelity and mask manufacturability. Experimental results show that, compared to a currently professional PBOPC software, the proposed algorithm can achieve approximately two-fold speedup and more manufacture-friendly OPC patterns.