Fast Optical Proximity Correction Using Graph Convolutional Network With Autoencoders

Abstract

OPC is a very time consuming process for mask synthesis. Quick and accurate OPC using GCN with layout encoder and mask decoder is proposed. (1) GCN performs a series of aggregation with MLP for correction process. A feature of a particular polygon is aggregated with weighted features of neighbor polygons; this is a key motivation of using GCN since one polygon should be corrected while its neighbors are taken into account for more accurate correction. (2) GCN inputs are provided by a layout encoder, which extracts a feature from each layout polygon. GCN outputs, features corresponding to corrected polygons, are processed by a mask decoder to yield the final mask pattern. (3) The encoder and decoder originate from respective autoencoders. High fidelity of decoder is a key for OPC quality. This is achieved by collective training of the two autoencoders with a single loss function while the encoder and decoder are connected. Experiments demonstrate that the proposed OPC achieves 47% smaller EPE than OPC using a simple MLP model.