ML-Guided Curvilinear OPC: Fast, Accurate, and Manufacturable Curve Correction

Abstract

In curvilinear optical proximity correction (OPC), each segment is modeled by a cubic Bézier curve, defined by two endpoints and two intermediate points. Iterative correction of these points is not trivial, and a simple heuristic (Chen et al., 2024) has been used but is not effective. A vertex placement error (VPE) is first introduced to replace edge placement error (EPE) in standard Manhattan OPC. Two machine learning models are applied for accurate curve correction. (1) An MLP is used to locate the new endpoints, while VPE from the previous iteration and a few PFT signals representing local light intensity are provided as inputs. (2) A VPE predictor, constructed with GCNs, is designed to output average (or maximum) VPE over a given layout clip. Once trained, it is used to identify intermediate points after new endpoints are fixed by MLP; this is done through gradient descent optimization such that VPE is minimized and curvature constraints are respected as much as possible. Experimental results demonstrate that the proposed curvilinear OPC reduces OPC iterations from 8 to 5 when average VPE is considered as a target or from 14 to 5 when maximum VPE is a target, with a final VPE reduction of about 5 to 6%.