Model-driven optical proximity correction via hypergraph convolutional neural networks and its experimental demonstration

Optical proximity correction (OPC) is a pivotal approach to correct the optical proximity effect in the lithography process by optimizing the mask pattern. However, traditional pixel-based OPC algorithms are computationally intensive because these algorithms use computationally expensive methods (e.g., with gradients or pixel-based analysis) to optimize the mask. This paper proposes a novel OPC method based on model-driven hypergraph convolutional network (MHGCN) to improve the computational efficiency and lithography image fidelity simultaneously. In the MHGCN framework, a dense concentric circular sampling (DCCS) template is first used to extract the feature matrices of the target layout, which are then imported to a hyperedge graph convolutional network (encoder) to predict the OPC result. Then, a lithography-model-based decoder is used to calculate the wafer image of the OPC mask. A low-cost unsupervised training method of the network is developed through the collaboration between the encoder and decoder. To demonstrate the proposed method, a digital-micromirror-device (DMD) maskless lithography testbed is established, and the imaging model of the testbed is calibrated based on a set of training layout patterns. The superiority of the proposed method in both optimization capacity and computational efficiency are proved by the simulation and experimental results.