Efficient source and mask optimization based on interpretable hypergraph auto-encoding network

Source and mask co-optimization (SMO) is an important technique to improve the image fidelity and process window of the advanced lithography process. Modern SMO methods adopt pixelated representations of source and mask to improve the degrees of optimization freedom, but those approaches are computationally intensive and time consuming. To our knowledge, this paper is the first to develop an efficient SMO method based on the advanced hypergraph deep learning framework. Firstly, a novel mask clip selection method based on sparse signal reconstruction is developed to rapidly determine the representative mask clips for source optimization. Based on the selected mask clips, the source pattern can be optimized using the fast gradient-based algorithm. Then, an advanced deep learning approach, dubbed hypergraph auto-encoding network, is used to accelerate the mask optimization under the optimal illumination condition. In the proposed network, a hypergraph model is used to predict mask optimization solutions directly based on layout features. A reverse decoder based on the lithography physical imaging model is used to self-supervise the network training. Simulation results demonstrate the superiorities of the proposed methods in terms of lithography image fidelity, process robustness, and computational efficiency.