Neural Adjoint Method for Meta-optics: Accelerating Volumetric Inverse Design via Fourier Neural Operators

Chanik Kang, Hyewon Suk, Haejun Chung · 2026

Read Neural Adjoint Method for Meta-optics: Accelerating Volumetric Inverse Design via Fourier Neural Operators by Chanik Kang, Hyewon Suk, Haejun Chung (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-17425.txt.

Introduction

Kang, Suk, and Chung speed up the design of meta-optics, the tiny structures that steer light for imaging and color routing. Designing them means solving Maxwell's equations over and over, and broadband designs can demand thousands of such solves. Their Neural Adjoint Method trains a Fourier neural operator to predict the sensitivity field for each voxel, replacing costly adjoint solves with fast predictions. A staged design sharpens high-frequency detail, and tests on color routers, metalenses, and waveguide converters cut design time from hours to seconds.

Catalog record

CREATORChanik Kang, Hyewon Suk, Haejun Chung
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35718
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.17425

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neural Adjoint Method for Meta-optics: Accelerating Volumetric Inverse Design via Fourier Neural Operators Chanik Kang, Hyewon Suk, Haejun Chung Abstract . Meta-optics promises compact, high-performance imaging and color routing. However, designing high-performance structures is a high-dimensional optimization problem: mapping a desired optical output back to a physical 3D structure requires solving computationally expensive Maxwell’s equations iteratively. Even with adjoint optimization, broadband design can require thousands of Maxwell solves, making industrial-scale optimization slow and costly. To overcome this challenge, we propose the Neural Adjoint Method, a solver-supervised surrogate that predicts 3D adjoint gradient fields from a voxelized permittivity volume using a Fourier Neural Operator (FNO). By learning the dense, per-voxel sensitivity field that drives gradient-based updates, our method can replace per-iteration adjoint solves with fast predictions, greatly reducing the computational cost of full-wave simulations required during iterative refinement. To better preserve sensitivity peaks, we introduce a stage-wise FNO that progressively refines residual errors with increasing emphasis on higher-frequency components. We curate a meta-optics dataset from paired forward/adjoint FDTD simulations and evaluate it across three tasks: spectral sorting (color routers), achromatic focusing (metalenses), and waveguide mode conversion. Our method reduces design time from hours to seconds. These results suggest a practical route toward fast, large-scale volumetric meta-optical design enabled by AI-accelerated scientific computing. 1. Introduction Figure 1. Breaking the simulation bottleneck in 3D meta-optics. (Top) Traditional optimization relies on expensive iterative full-wave high-fidelity simulations (Maxwell solvers) to compute gradients, taking hours or days. (Bottom) The proposed Neural Adjoint Method replaces the solver with a Fourier Neural Operator. By inferring volumetric gradients in seconds. Optical-system miniaturization has accelerated the commercialization of meta-optics, indicating a major transition from academic research to industrial-scale manufacturing ( 19 ; 15 ; 31 ) . In particular, the display and semiconductor industries are increasingly adopting 3D volumetric meta-optics, with color routers for Complementary Metal-Oxide-Semiconductor (CMOS) image sensors being a prominent example, to address the fundamental efficiency constraints of conventional dye-based color filters ( 20 ; 29 ; 40 ) . Unlike thin diffractive components, volumetric meta-optical devices leverage complex multiple scattering within micrometer-scale volumes to enable precise control of wavefronts and spectra. For deployment in consumer electronics, these devices must meet stringent performance requirements over broad wavelength ranges, necessitating an unprecedented degree of precision in the design and optimization process ( 10 ) . However, industrial deployment of these devices remains limited by substantial computational overhead. The inverse design of 3D meta-optics is typically a high-dimensional, non-convex optimization problem constrained by Maxwell’s equations ( 4 ) . The adjoint method is the standard approach for computing gradients, requiring two full-wave simulations per iteration (one forward and one adjoint) ( 26 ) . Although adjoint optimization is computationally efficient in terms of gradient evaluation, the overall cost of rigorous 3D solvers such as finite-difference time-domain (FDTD) grows rapidly with problem size ( 16 ) . In addition, commercial designs must be optimized over the full visible band, which typically requires separate simulations for each wavelength or color channel ( 38 ) . As a result, a single optimization run can involve tens of thousands of full-wave simulations and take days to weeks to complete. This turnaround time poses a major obstacle for high-throughput industrial R&D, where rapid prototyping and iterative refinement are critical. To mitigate this cost, Deep Learning has emerged as a promising alternative. Generative models (GANs ( 13 ; 35 ; 1 ) , Diffusion ( 39 ; 22 ; 36 ; 24 ) ) offer rapid "one-shot" designs, but they suffer from a fundamental precision-speed trade-off ( 17 ) . While fast, they often lack the physical rigor required for engineering-grade optics, producing plausible-looking structures that fail to meet the strict efficiency and fabrication tolerances of mass production. Conversely, numerical optimization guarantees physical optimality but is constrained by the extreme slowness of the solver ( 5 ) . To overcome this bottleneck and narrow the gap between research and manufacturing, we propose the Neural Adjoint Method , a real-time surrogate framework that combines the speed of generative AI with the precision of iterative optimization. Our key insight is to decouple the “Optimizer” from the “Solver.” Instead of replacing the entire optimization process with an end-to-end generator, we replace only the expensive Maxwell solver with a Fourier Neural Operator (FNO) ( 21 ; 7 ) based surrogate. This allows us to predict the 3D volumetric gradient field in a second, enabling a solver-free optimization loop that drastically shortens the design turnaround time. Our approach addresses the challenges of volumetric inverse design through three key contributions: • Volumetric neural-adjoint gradient learning for Meta-optics. We formulate inverse design around dense 3D adjoint sensitivities and train a neural operator to predict full volumetric adjoint gradient fields from voxelized devices. • Stage-wise Fourier Neural Operator (SW-FNO) for multi-scale sensitivities. To address the low-frequency bias of single-pass operator networks, we introduce a coarse-to-fine residual refinement scheme with progressively higher-frequency emphasis, improving recovery of localized, high-magnitude sensitivity peaks that are critical for topology refinement. • Reduced-solver optimization for fast design iteration. We integrate the inferred gradients into a gradient-based optimization loop, enabling rapid design iteration, and validate the approach on meta-optics problems. We validate the proposed framework on three representative 3D inverse-design tasks: RGB color routing, achromatic metalens design, and waveguide mode conversion. Across all tasks, the Neural Adjoint Method achieves performance comparable to rigorous FDTD-based adjoint optimization, while reducing end-to-end design time by several orders of magnitude, with speedups of up to \sim 10^{4} – 10^{5}\times . By mitigating the bottleneck of repeated full-wave simulations, these results suggest a practical pathway toward rapid, large-scale 3D meta-optical design enabled by AI-accelerated scientific computing. 2. Problem Formulation In this section, we formulate the volumetric inverse design problem as a constrained optimization task governed by Partial Differential Equations (PDEs). We outline the computational complexities inherent in this process and introduce our Neural Adjoint formulation. 2.1. Inverse Design for Diverse Physical Objectives Figure 2. 3D meta-optical inverse-design tasks considered in this work: (a) spectral sorting (color router), (b) light focusing (metalens), and (c) guided-wave mode conversion (waveguide). Inverse design in meta-optics and nanophotonics seeks a device parameterization (e.g., a 2D/3D permittivity distribution) that maximizes or minimizes a physics-based objective function ( 16 ) . A common approach is gradient-based adjoint optimization: given an objective functional \mathcal{J}(\mathbf{x}) evaluated by solving Maxwell’s equations, the adjoint method computes \nabla_{\mathbf{x}}\mathcal{J} using only two full-wave simulations, a forward solve to obtain the electromagnetic fields and an adjoint solve to backpropagate sensitivity, enabling scalable updates even when \mathbf{x} is high-dimensional ( 26 ) . In this work, we use this conventional formulation as the reference pipeline and focus on accelerating the expensive field/gradient evaluations. As illustrated in Fig. 1 , we consider volumetric inverse design within a three-dimensional physical domain \Omega\subset\mathbb{R}^{3} . The design variable is the spatial permittivity distribution \epsilon(\mathbf{r}) , which is discretized as a voxel grid \mathbf{x}\in\mathbb{R}^{N} with N=N_{x}N_{y}N_{z} degrees of freedom. For a given design \mathbf{x} and wavelength \lambda , a full-wave Maxwell solver computes the complex electric field \mathbf{E}(\mathbf{r};\mathbf{x},\lambda) , from which task-specific FoMs are defined. We study three representative classes of meta-optics/nanophotonics objectives, which are described in Fig. 2 : • Intensity maximization (metalens). We maximize

Read the complete text