Fast Simulation of Electromagnetic Fields in Doubly Periodic Structures With a Deep Fully Convolutional Network

Abstract

The simulation of electromagnetic fields in many applications with double periodicity can be viewed as an image-to-image translation problem, where the structure is known and the fields need to be predicted. In this work, we take extreme ultraviolet (EUV) lithography as an example and propose a very efficient technique based on the U-Net architecture, a deep fully convolutional network (FCN). The U-Net is trained with data of some typical patterns in a 128 nm $\times\,\,128$ nm unit cell, and good accuracy can be obtained in the prediction of the near field on much larger unit cells with complex patterns. Furthermore, numerical experiments demonstrate that the proposed method is three orders of magnitude faster than conventional state-of-the-art methods, such as the spectral-element spectral-integral method.