Inverse Design of Multifunctional Metasurface Based on Multipole Decomposition and the Adjoint Method

Abstract

The functional and optimal design of optical metasurfaces poses great challenges, particularly in situations requiring multifunctionalities. The conventional forward design relies on a priori knowledge, which limits the development of metasurfaces with customized functions. Inverse design has achieved major breakthroughs in developing new optical functionalities but has been hindered by inevitably invoking a full-wave electromagnetic solver, subjected to low efficiency, slow convergence, and being hard to implement. Here, we propose a general and computationally efficient inverse-design framework based on multipole decomposition and adjoint optimization. By this method, we demonstrate an achromatic metalens working across the visible and near-infrared band with a bandwidth over 400 nm. We also show the design of the metasurface for multiplexed focusing at the RGB wavelength, and the metasurface for stimulated emission depletion (STED) microscopy, which generates a tightly focusing point at the excitation beam wavelength and doughnut-shaped focus for the STED beam wavelength. Our scheme could be further improved to include more geometrical degrees of freedom and show great applications in free-space diffractive optics and deep diffractive neural network.

Spectro-polarimetric-depth Imaging by Inverse-Designed Single-Cell Metasurface

“…We set the distance between the phase mask and the object planes as 2 z = cm, and that between the phase mask and the sensor as 62 μm d = (hereafter referred them as object-image distance). It takes around two hundred iterations for the loss function L to converge, using the moving method of asymptotes (MMA) to update the gradient [4].…”

“…ETASURFACES that exhibit unique optical properties have ushered in a new era of meta-optics with extraordinary ultra-compact optical devices [1]. Metasurface enables spatially varying phase, amplitude and polarizations control on an incoming optical beam, for example, as for high-numerical-aperture focusing or imaging [2], beam steering over large angles [3], and computer-generated holography [4], [5], [6]. In particular, metasurface-based multi-channel computational imaging has attracted great attention due to the compact size and strong light manipulation ability of metasurface [7], [8], [9], [10].…”

Multi‐Dimensional Multiplexed Metasurface Holography by Inverse Design

“…The derivatives of the loss function corresponding to multiple images with respect to the length and width of each meta-atom are calculated and input into the method of moving asymptotes (MMA) in the NLopt toolbox so as to quickly update and optimize the metasurface parameters. [8,45,46] Thus, a direct relationship between the structural parameters of the metasurface and the output reconstructed images is built, which enables the end-to-end design of miscellaneous meta-devices. With the help of the proposed algorithm, the exact metasurface parameters can be quickly obtained by simply changing the target performance requirements of meta-devices.…”

Deep Learning and Adjoint Method Accelerated Inverse Design in Photonics: A Review

“…As a result, except for topological optimization, the AM has also been widely applied to other types of photonic inverse design tasks, more or less, along with different acceleration techniques. In [28], large metasurfaces were optimized, where the AM is used to calculate the gradient of electric field strength relative to all design parameters. In the optimization process, multipole expansion theory (MET) was employed to accelerate the computation, which offered orders-of-magnitude-higher efficiency than the FDTD-based optimization method did.…”

“…Because the AM can calculate the derivatives over all the design parameters and because it modifies the parameters in proportion to the figure of merit (FOM) gradient using only forward and adjoint simulations at each iteration, regardless of the number of design parameters [26,27], it has been successfully applied to various photonic systems [27,28]. This is one of the major reasons why the AM has been widely adopted for the topology or shape optimization of photonics devices [28], where the number of design parameters can be very large to describe complex free-form geometries. The traditional AM was recently extended to nonlinear device modeling in the frequency domain [29].…”