Comprehensive Overview of Deep Inverse Models in Metamaterials Design
The design of electromagnetic metamaterials to achieve desired functionalities often presents a challenging inverse problem, where the required structural parameters are sought for a given electromagnetic response. Traditional computational electromagnetic simulation (CEMS) methods can be computationally expensive and inefficient for exploring the vast design space. Over the past several years, deep learning has emerged as a powerful tool to address these challenges, leading to the development of various deep inverse models (DIMs). These models aim to learn the inverse mapping from a desired spectral response to the corresponding metamaterial geometry. This paper provides a comprehensive overview of the different types of DIMs applied in metamaterials design, discusses the inherent difficulties of the inverse problem, including its ill-posed nature, compares the performance of various DIMs based on accuracy and inference speed, and highlights future directions in this rapidly advancing field. We explore the strengths and limitations of different DIM architectures, such as Invertible Neural Networks, Neural Adjoint methods, and Tandem Networks, in tackling the complexities of metamaterial inverse design, particularly the issue of non-uniqueness in solutions.
