Deep Image Prior-Based Super Resolution for Fast Electromagnetic Forward Modeling

Abstract

Training data and generalization capability are two of the major obstacles hindering the widespread application of deep learning in computational electromagnetics. To alleviate these challenges, this communication proposes a deep image prior (DIP)-based modeling framework to facilitate fast-forward modeling of large-scale problems. This framework leverages an unsupervised learning network with super-resolution capability, requiring only the low-resolution field distribution portrait in the model as input. The low-resolution field is initially generated using the conventional finite-difference frequency-domain (FDFD) method with a coarse mesh. Through iterative updates, the network produces a refined field distribution that achieves an accuracy comparable to that obtained with a mesh twice as dense as the original one. In addition, an automated termination criterion is introduced for DIP iteration. Extensive numerical experiments validate the super-resolution capability of our method in handling 2-D scattering problems with diverse shapes and material properties. In addition, a large-scale model is deployed to demonstrate the generalization capability of the proposed modeling framework for realistic applications. Notably, the proposed method exhibits superior computational efficiency compared to conventional FDFD for large-scale modeling, as it only requires forward modeling on a coarse mesh. By eliminating the need for training data and demonstrating good generalization capability, this method shows significant potential in addressing the challenges posed by conventional modeling techniques, particularly in terms of the intensive computational overhead associated with large-scale problems.