Machine learning-assisted imaging and design of metamaterials

Abstract

Over the past two decades, metamaterials have been developed to provide unconventional constitutive parameters in various regimes, including optics, microwaves, acoustics, and elastic waves. Recent advances in the field have led to the development of metasurfaces, which consist of a single layer of metamaterial atoms with spatially dependent structural profiles. These metasurfaces store a vast amount of information about light-matter interactions through their structural parameters, leading to various applications, including the generation of vortex beams, holograms, and, more recently, imaging with polarization capabilities. As these functionalities become more and more complex, emerging machine learning techniques become useful either for inverse design of these metasurfaces or for extracting information from wave propagation data when the generated wavefront interferes with an object for imaging purposes. In this thesis, I focus on exploiting the rich degrees of freedom of metamaterials or metasurfaces for imaging and inverse design applications, with the assistance of machine learning techniques. Regarding imaging applications, from a data-centric perspective, we can define imaging as the process of seeking a minimal representation of wave propagation data, which includes material or geometric parameter profiles of the object to be imaged. These material parameter profiles can be the spatially varying PDE coefficients in the underlying dynamical equations or the scattering matrices of the constituent elements of an object. First, we applied this methodology to elastic wave imaging to resolve the spatial profiles of density, modulus (spring constants), or structural parameters using the β-VAE machine learning model. The β-VAE, a type of variational autoencoder, determines the minimally required latent variables for data representation. Second, we employed the same β-VAE scheme for low-light coincidence imaging. Here, the β-VAE was used to learn the interference between wavefronts generated by a metasurface and an unknown object, enabling the extraction of the Jones matrix profile of the unknown object. Interestingly, the β-VAE was applied to learn physical process that is intrinsically probabilistic in nature for photon coincidence measurements. The β-VAE-based approach demonstrates higher accuracy, requiring fewer photons, and outperforming previously semianalytically derived algorithms. Additionally, it offers the advantage of assessing the sufficiency of information obtained from a designed experimental procedure. For the inverse design applications in generating tailor-made wavefronts for metasurfaces, as the first example, we establish a single deep neural network that integrates both the inverse design of metamaterial structures and optical hologram generation. This allows us to design bianisotropic metasurfaces to generate metasurface holograms with four distinct co- and cross-polarization channels. The integrated approach also allows us to achieve better accuracy in the resultant holograms than the conventional iterative Gerchberg-Saxton algorithm, particularly when more constraints are applied to the metamaterial structures due to fabrication limitations. As the second example, we focus on the inverse design of metasurfaces for microwave power allocation. Rather than using simulation data, a programmable metasurface is adopted to generate arbitrary reflection phase profiles, in which the measured intensities for a significant amount of the experimental data set are used for network training. Such an approach facilitates the combination of the programmable metasurface with machine learning, which can be a cost-effective substitute for reconfigurable intelligent surfaces. Overall, this thesis highlights the potential of machine learning leveraged by the vast number of engineering degrees of freedom offered by metamaterials. Assisted by machine learning techniques, metasurfaces can aid in imaging by discovering minimal representations and also in the inverse design of metasurfaces for generating on-demand wavefronts and holograms. In both applications, the adoption of machine learning results in higher accuracy and automated algorithm discovery, compared to conventional approaches.

Date of Award	2023
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Jensen Tsan Hang LI (Supervisor)