Deep Learning-Based Fast Full-Wave Electromagnetic Forward Solver Using Physics-Induced Loss

Xingyue Guo

doi:10.1109/LAWP.2024.3424940

2024-09-01

PRE

DOI：10.1109/LAWP.2024.3424940

原文链接

原文求助

摘要

En 中文

In this letter, a new deep learning (DL) solver is introduced, offering a rapid and effective full-wave computational method for simulating the electromagnetic forward (EMF) process. The core of this solver is the deep residual convolutional neural network (DRCNN) architecture. The DRCNN ingests the incident EM field along with the permittivity contrast distribution within the specified region of interest (ROI), and in turn, it generates the resultant EM field that is influenced by the incident EM wave. During its offline training, an EM scattering simulation tool has been developed to calculate the EM scattered fields. The loss function includes two distinct components: the data-induced loss, which quantifies the discrepancy between the DRCNN's predicted total EM fields and the actual labeled EM fields and the physics-induced loss, which measures the variance between the EM scattered fields derived from the actual labeled total EM fields and the scattered fields generated from the DRCNN's predictions. The training of this proposed EMF solver only relies on a simple synthetic dataset. Compared with traditional approaches, the proposed DL solver addresses EMF issues with notable precision while significantly cutting down on computational duration. Numerical benchmarks have demonstrated the viability of this DL solver.

Keyword:

Training

Scattering

Mathematical models

Loss measurement

Decoding

Convolution

Method of moments

Convolutional neural network

deep learning (DL)

electromagnetic forward process

real time