Accelerating optical proximity correction through deep learning

Abstract

Optical proximity correction (OPC) is necessary process for mask synthesis. We address three methods to adopt deep learning into OPC flow for faster OPC process. First, a bidirectional recurrent network (BRNN) based OPC is presented. Based on that mask correction during OPC has causality between mask biases, we map each mask bias into single instance in RNN, so that prediction can consider causality between segments. Next, we propose to optimize resist kernels using convolutional neural network (CNN).The structure of resist model and CNN are similar, we can use resist kernels as filters in CNN, so that kernels are optimized through CNN training process. As a result, we can derive complex kernel shapes, rather than function-based kernels, so that the number of convolutions in resist model is reduced and reduces the lithography simulation runtime. Finally, assist feature (AF) print fix using CNN is presented. By using gradient map from CNN, we can identify which part of AF cause AF printing and give guides to fix AF print while minimizing the effect in process window.