Deep Inverse Design of Metamaterials and Metasurfaces

Metamaterials and metasurfaces - here termed artificial electromagnetic materials (AEMs) - offer unprecedented control over the scattering of electromagnetic waves. Often the design of AEMs is cast as a challenging inverse problem, where an electromagnetic response is desired and the AEM which gives rise to it is sought. Inverse design problems are difficult because they are often ill-posed. Over the last several years deep learning has successfully been applied to solving AEM inverse problems and significant advance has been made. We overview various deep learning methods used to study inverse problems - termed deep inverse models (DIMs) - and compare their performance on benchmark datasets. It is found that the neural adjoint (NA) method is the most accurate DIM on problems investigated, whereas the popular Tandem and conditional variational autoencoder (cVAE) DIMs are often the fastest. We also summarize the different DIM approaches and explain the reasons for their varied success, and give an outlook of this exciting research area.