ResGNN-OPC: a fast optical proximity correction approach based on an interpretable residual graph neural network

Optical proximity correction (OPC) is a pivotal resolution enhancement technique that compensates for the image distortion in the optical lithography process. However, the turn-around time of modern OPC techniques, namely model-based OPC (MBOPC), increases dramatically as the lithography technology node constantly pushes forward. This paper develops a learning-based OPC approach using the residual graph neural network, dubbed ResGNN-OPC, to achieve more than 6000-fold acceleration compared to the conventional MBOPC. The residual network structure is integrated in the graph convolution layer to emulate each iteration of the MBOPC algorithm and to facilitate the increment of the network depth, thereby enhancing the network's interpretability and fitting ability. The superiorities of the proposed method are verified by numerical experiments. Compared to the existing learning-based OPC approach, the proposed method improves the prediction accuracy of OPC corrections by more than 40% and costs comparable runtime. In addition, ResGNN-OPC can output the final OPC solution with fewer or even no subsequent iteration steps but can still achieve similar image fidelity as the MBOPC approach.