Investigating the Efficiency of Using U-Net, Erf-Net and DeepLabV3 Architectures in Inverse Lithography-based 90-nm Photomask Generation

The paper deals with the inverse problem of computational lithography. We turn to deep neural network algorithms to compute photomask topologies. The chief goal of the research is to understand how efficient the neural net architectures such as U-net, Erf-Net and Deep Lab v.3, as well as built-in Calibre Workbench algorithms, can be in tackling inverse lithography problems. Specially generated and marked data sets are used to train the artificial neural nets. Calibre EDA software is used to generate haphazard patterns for a 90 nm transistor gate mask. The accuracy and speed parameters are used for the comparison. The edge placement error (EPE) and intersection over union (IOU) are used as metrics. The use of the neural nets allows two orders of magnitude reduction of the mask computation time, with accuracy keeping to 92% for the IOU metric.