Deep Diffractive Neural Network Meta-surface: Smart Optics through Smart Design

I.U. Idehenre
Azimuth Corporation,
United States

Keywords: meta-surface, metasurface, meta-atom, diffractive optical elements, neural networks, RCWA

Summary:

Artificial neural networks (ANNs) have proven incredibly useful across a diverse range of applications including natural language processing and image classification. Such networks typically exist in silico, requiring state-of-the-art hardware to be trained and run. Recent research efforts have resulted in the physical realization of ANNs in the form of Deep Diffractive Neural Networks (D2NN). D2NN are cascaded diffractive optical elements (DOE) that are intended to replicate the behavior of ANN.1,2 These DOE are “programmed” with a phase profile that interacts with incident light to produce the desired response. In much of the literature D2NN are limited to thin plates of dielectric/metallic material such as glass, where the height of the on one surface is varied (see Figure 1). In this presentation we will explore the combination of D2NN with meta-atoms which have the potential to allow sophisticated optical computation/processing/manipulation of polarized polychromatic light. Meta-atoms (see Figure 1) are dielectric/metallic nano-scale structures often etched on the surface of a thin bulk material. Their size and shape allow for the manipulation of light polychromatic and polarized light that is difficult with standard DOEs cannot. D2NN meta-surfaces devices would be passive, thin, light weight, and cheap finding industrial use in areas as diverse as LIDAR, remote sensing applications, virtual/augmented realities devices. While D2NNs meta-surfaces have great potential, their design is not trivial. The approaches popular for modeling meta-atoms require computationally expensive rigorous electromagnetic algorithms (RCWA, FDTD, FEM). The process of identifying shapes that can meet the needs of the (D2NN) is often slow and time consuming. Even the process of constructing the phase profile of the meta-surface can be time consuming. In this presentation, we will discuss the inverse design techniques we use to simulate/design D2NN meta-surfaces. Additionally, we present interesting example simulations for beam steering, spectral signal processing, and polarization imaging (Figure. 2). 1. X. Lin, Y. Rivenson, N. T. Yardimci, M. Veli, Y. Luo, M. Jarrahi, and A. Ozcan, Science 361, 1004–1008 (2018). 2. I. U. Idehenre, E. S. Harper, and M. S. Mills, Opt. Express. 30, 7441-7456 (2022). 3. NIL Technology. (2023, January 06). Diffractive Optical Elements. Nilt.com. https://www.nilt.com/technology/diffractive-optical-elements/ 4. Y.W. Huang, D. Naidoo, Photonics Spectra, 34-36 (2020).