Fast and Accurate Curvilinear OPC with ML-Guided Curve Correction

In curvilinear OPC, each pattern segment is assumed as a cubic Bézier curve, which is defined by two endpoints and two intermediate points. Iterative correction of the curves is very time consuming even with simple heuristic [1]. Two ML models are introduced for fast and accurate curve correction. An MLP is used to locate the new endpoints, while correction error (called VPE) from the previous iteration and a few PFT signals representing local light intensity are provided as inputs. VPE predictor is constructed with both target and mask patterns being modeled with nested GCNs; the output is average VPE. Intermediate points are identified while the VPE predictor is optimized with a goal of minimizing average VPE and violations of curvature constraints and geometric continuity constraints. The proposed OPC is compared to the reference [1]: the number of OPC iterations is reduced from 12 to 3 while final average VPE is reduced from 0.49nm to 0.44nm.