A self-training end-to-end mask optimization framework based on semantic segmentation network*

As the process advances and the minimum linewidth gets closer to the physical limit, inverse lithography technique (ILT) is widely used for optical proximity correction (OPC). However, the computational overhead of the ILT method is high, and the printability of the mask is poor. In response to these limitations, we proposed ERFNet-ILT, a self-training method for an end-to-end learning framework for generating optimized masks directly from layout patterns, which introduces a feature fusion module at the end of the encoder and uses dilated convolution to expand the receptive field, thereby extracting layout pattern information such as edges, vertices and corners of the layout pattern from the feature map. The framework has shorter model building time and higher mask printability. Compared with the state-of-the-art methods, experimental quantitative results show that the proposed framework achieves 2.5 % squared L2 error and 7.9 % process variation band reduction within a comparable mask correction time.