Inverse deep learning methods and benchmarks for artificial electromagnetic material design | JoVE Visualize

Area of Science:

Computational electromagnetics
Materials science
Artificial intelligence

Background:

Artificial electromagnetic materials (AEMs) like metamaterials, photonic crystals, and plasmonics are crucial for controlling electromagnetic wave interactions.
Deep inverse models (DIMs), a type of deep learning, are increasingly used for designing AEMs to achieve specific scattering properties (e.g., transmission/reflection spectra).
A comprehensive comparison of existing DIMs for AEM design is lacking.

Purpose of the Study:

To compare the performance of eight state-of-the-art deep inverse models (DIMs) for designing artificial electromagnetic materials (AEMs).
To introduce and evaluate two novel DIMs for AEM design.
To determine the optimal approach for AEM design problems, considering both ill-posed and well-posed scenarios.

Main Methods:

Evaluated eight state-of-the-art deep inverse models (DIMs) on three distinct artificial electromagnetic material (AEM) design challenges.
Introduced and assessed two novel DIMs within the context of AEM design.
Compared DIM performance against conventional deep neural networks for well-posed design problems.

Main Results:

Deep inverse models (DIMs) demonstrated the capability to rapidly generate accurate designs for achieving custom scattering properties in artificial electromagnetic materials (AEMs) across all tested problems.
The Neural-Adjoint DIM approach exhibited the strongest overall performance across the evaluated AEM design scenarios.
For well-posed AEM design problems, conventional deep neural networks outperformed specialized DIMs.

Conclusions:

Deep inverse models (DIMs) are effective tools for accelerating the design of artificial electromagnetic materials (AEMs) with tailored scattering characteristics.
The Neural-Adjoint model presents a robust option for AEM design, though problem type dictates the optimal deep learning strategy.
It is recommended to utilize a conventional deep neural network as a baseline for AEM design, especially when problem well-posedness is uncertain.