Machine-learning-empowered multispectral metafilm with reduced radar cross section, low infrared emissivity, and visible transparency

For camouflage applications, the performance requirements for metamaterials in different electromagnetic spectra are usually contradictory, which makes it difficult to develop satisfactory design schemes with multispectral compatibility. Fortunately, empowered by machine learning, metamaterial design is no longer limited to directly solving Maxwell’s equations. The design schemes and experiences of metamaterials can be analyzed, summarized, and learned by computers, which will significantly improve the design efficiency for the sake of practical engineering applications. Here, we resort to the machine learning to solve the multispectral compatibility problem of metamaterials and demonstrate the design of a new metafilm with multiple mechanisms that can realize small microwave scattering, low infrared emissivity, and visible transparency simultaneously using a multilayer backpropagation neural network. The rapid evolution of structural design is realized by establishing a mapping between spectral curves and structural parameters. By training the network with different materials, the designed network is more adaptable. Through simulations and experimental verifications, the designed architecture has good accuracy and robustness. This paper provides a facile method for fast designs of multispectral metafilms that can find wide applications in satellite solar panels, aircraft windows, and others.