Machine learning for the free-form inverse design in nanophotonics

Doctoral thesis, 2023

In the fast developing field of nanotechnology, nanophotonics has emerged as a revolutionary technology, controlling the complex interplay between light and matter at the nanoscale. Crucial to the advancement of this field is the ability to manipulate the propagation of electromagnetic waves by shaping the transmitted or reflected wavefront through the design of structures that are only a fraction of the wavelength in size. The design of these subwavelength structures forming a quasicontinuous material is complex and usually exceeds human intuition.

In this thesis, we present our approach utilizing deep neural networks for both the prediction of optical properties of nanostructures and their inverse design. First, we demonstrate the feasibility of employing neural networks to accurately predict the optical properties of nanostructured materials, with a particular emphasis on metasurfaces. Given the increasing need for miniaturization together with smaller optical losses, the emergence of metasurfaces---single layers of phase-modifying nanostructures---has heralded a significant advancement, enabling unprecedented manipulation of light beyond the capabilities of traditional materials.

Subsequently, we present methodologies to achieve inverse design of metasurfaces, guided by specific desired optical attributes and fabrication constraints. This is achieved through neural network design and the generation of training data, labeled with corresponding optical characteristics and manufacturability constraints. We implemented a conditional generative adversarial network comprised of five synergistic neural networks. This approach effectively mitigates challenges related to design non-uniqueness, mode collapse, and experimental feasibility.
Additionally, our research explores various techniques to optimize and stabilize neural network training and introduces novel network graph compositions, contributing to a versatile generator model capable of conceiving multiple metasurface unit cells with predefined, interconnected properties. This thesis thus provides insights and methodologies for inverse design, contributing to the continued evolution of nanophotonic design and applications.