Optimization of machine learning guided optical proximity correction

Abstract

Machine learning guided optical proximity correction (ML-OPC) has been proposed to replace computation extensive model-based OPC (MB-OPC) or to provide a good initial OPC solution to work with. Two keys of ML-OPC are the representation of layout segment to be corrected, and the choice of regressors (or classifiers) with its optimization. We propose polar Fourier transform (PFT) signals with initial edge placement error (EPE) as a set of parameters for representation, and random forest regressor (RFR) as our choice of machine learning algorithm. Experimental results demonstrate significant reduction in root mean square (RMS) error in mask bias prediction compared to state-of-the-art ML-OPC approach: reduction from 1.45nm to 0.66nm. Customizing RFR for each group of layout segments that share the similar neighbors allows further reduction of RMS error by 0.10nm.