Keywords: Inverse design, conditional variational autoencoder, functional response, metamaterial
TL;DR: We extend conditional variational autoencoder for discrete responses to a more general version that can handle functional responses. A series of experiments confirm that the proposed method is effective to achieve inverse design for metamaterials.
Abstract: Metamaterials are emerging as a new paradigmatic material system, providing unprecedented and customizable properties for various engineering applications. However, the inverse design of metamaterials, which aims to retrieve the metamaterial microstructure according to a given electromagnetic response, is very challenging as it is non-trivial to unveil the nonintuitive and intricate relationship between the microstructures, and their functional responses. In this study, we resolve this critical problem by extending the classic conditional variational autoencoder for discrete responses to a more general version that can handle functional responses. By encoding microstructures and their electromagnetic response curves into common latent spaces via deep neural networks and aligning them via a specific loss function, the proposed functional response conditional variational autoencoder can unveil the implicit relationship between microstructures and their electromagnetic responses efficiently. The proposed novel learning framework not only facilitates metamaterial design greatly by avoiding the time-consuming case-by-case numerical simulations in the traditional forward design, but also has the potential to resolve other problems with similar structures.
Conference Poster: pdf