Inverse Design of Diffractive Metasurfaces Using Diffusion Models
Liav Hen, Erez Yosef, Dan Raviv, Raja Giryes, Jacob Scheuer · 2025
Read Inverse Design of Diffractive Metasurfaces Using Diffusion Models by Liav Hen, Erez Yosef, Dan Raviv, Raja Giryes, Jacob Scheuer (2025) 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-2506-21748.txt.
Introduction
Liav Hen and co-authors use diffusion models to design metasurfaces, flat optical parts built from tiny structures that steer light. Their 2025 paper attacks inverse design: given the light pattern you want, find the geometry that makes it. That search is hard because structure and optical response relate in a tangled, nonlinear way, and hand tuning gets stuck easily.
A physics simulator first generated training pairs of geometries and their scattering patterns. A conditional diffusion model then learned to propose a structure and height from a target pattern at a chosen wavelength. Once trained, the model designs new parts directly, replacing slow expert search with learned generation for these light-bending surfaces.
Catalog record
| CREATOR | Liav Hen, Erez Yosef, Dan Raviv, Raja Giryes, Jacob Scheuer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-21748
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Inverse Design of Diffractive Metasurfaces Using Diffusion Models Liav Hen, Erez Yosef, Dan Raviv, Raja Giryes, Jacob Scheuer Abstract Metasurfaces are ultra-thin optical elements composed of engineered sub-wavelength structures that enable precise control of light. Their inverse design - determining a geometry that yields a desired optical response - is challenging due to the complex, nonlinear relationship between structure and optical properties. This often requires expert tuning, is prone to local minima, and involves significant computational overhead. In this work, we address these challenges by integrating the generative capabilities of diffusion models into computational design workflows. Using an RCWA simulator, we generate training data consisting of metasurface geometries and their corresponding far-field scattering patterns. We then train a conditional diffusion model to predict meta-atom geometry and height from a target spatial power distribution at a specified wavelength, sampled from a continuous supported band. Once trained, the model can generate metasurfaces with low error, either directly using RCWA-guided posterior sampling or by serving as an initializer for traditional optimization methods. We demonstrate our approach on the design of a spatially uniform intensity splitter and a polarization beam splitter, both produced with low error in under 30 minutes. To support further research in data-driven metasurface design, we publicly release our code and datasets. 1 1 1 https://github.com/liavhen/metagen Keywords metasurfaces, diffusion models, inverse design, nanophotonics 1 Introduction (a) A meta-atom represented as a binary image. (b) A schematic of a meta-atom. (c) A schematic of the scattering scenario: A periodic metasurface, excited by a monochromatic plane-wave. Diffractive metasurfaces are optical devices composed of sub-wavelength meta-atoms that enable precise control over light propagation. Designing these structures to achieve specific optical properties remains a significant challenge due to the complex and non-linear relationship between the geometry and the optical response. Early research on metasurface inverse design focused on classical optimization techniques. In particular, gradient-based methods relying on numerical light-scattering simulations, such as Rigorous Coupled-Wave Analysis (RCWA) [ 1 ] , were extensively explored [ 2 , 3 , 4 ] . However, these approaches often converge to local minima. To address this, alternative strategies, such as topology optimization, evolutionary algorithms, and particle swarm optimization, have been proposed [ 5 , 6 ] . These are frequently combined with global phase initialization schemes, such as the Iterative Fourier Transform Algorithm (IFTA), to improve convergence [ 7 , 8 , 9 ] . While effective, these methods are computationally intensive, often requiring hours to design a single metasurface, and depend on expert-tuned regularization strategies. With the rise of deep learning, neural networks (NNs) have also been explored for metasurface design [ 3 , 10 , 11 , 12 , 13 ] . However, these models often underperform analytical methods due to their reliance on large, task-specific datasets which are rarely available in practice. It is important to note that most prior work on metasurface inverse design using NNs has focused on the spectral response as the primary design objective. In contrast, controlling the far-field spatial power distribution, namely the directionality and intensity of scattered Bragg orders, has received limited attention, despite its importance in beam splitting, shaping, and multiplexing applications. This paper addresses the inverse design of metasurfaces that produce a desired spatial far-field power distribution. We consider periodic, binary-thickness dielectric metasurfaces under normal incidence. Since the wavelength and geometric features are of comparable scale, the periodicity induces scattering into discrete diffraction orders. Our goal is to recover a meta-atom configuration that yields a prescribed angular power distribution at specified wavelengths. To this end, we have developed a generative framework based on diffusion models [ 14 , 15 ] - a class of neural generative algorithms increasingly used to generate high-fidelity samples and solve inverse problems. We train the model on a large dataset of simulated structure–response pairs, constructed using a novel, scalable data generation pipeline. It should be emphasized that in contrast to the more common tasks tackled by NNs, such as image classification, generation or text prediction, a large training dataset does not exist and must be constructed ad hoc. Our data generation method addresses the lack of high-quality task-specific data and can be adapted to various nanophotonic design problems. Once trained, diffusion models can be sampled to generate solutions that match the learned distribution. We explore two operational modes: • RCWA guidance (posterior sampling): full-wave simulations (e.g., RCWA) are integrated into the diffusion generative process, trading off accuracy vs. runtime. • Integrative mode: The diffusion model provides initial guesses for downstream optimization routines. It should be noted that the training process (especially the dataset generation method - see Section 2 ) creates a distribution of topology and diffraction pattern pairs. It can be expected that some of the metasurface topologies (and, hence, the attainable diffraction patterns) are not well-represented within this distribution. Consequently, the design of metasurfaces targeting such diffraction patterns is a challenging task for our diffusion model. We tackle this gap through incorporation of careful considerations to the data generation process and state-of-the-art methods for querying diffusion models in light or desired measurements consistency. We demonstrate the effectiveness of our approach on several challenging design tasks, particularly targeting patterns which are not well-represented within the distribution of the learned dataset. These include the generation of SiO 2 metasurfaces exhibiting an X-shaped scattering pattern across multiple wavelengths, the design of a silicon dual-polarization uniform beam splitter, and a polarization splitter refined via gradient descent, all within minutes. This paper is organized as follows: Section 2 details the data generation process, the inverse design process, and evaluation methodology. Section 3 presents experimental results and demonstrates the usage modes of our diffusion model for metasurface inverse design. 2 Method In this section, we present our proposed method for the inverse design of diffractive metasurfaces using diffusion models. We begin by introducing a novel data generation procedure. We then describe the training of our diffusion model, which we designate as MetaGen , and how it is used to design new metasurfaces. Finally, we outline our evaluation methodology, which includes comparisons with alternative generative frameworks and the design of metasurfaces corresponding to target scattering patterns. Additional details are provided in Sections B - D of the appendix. 2.1 Data Generation Dataset B2 C2 Polarizations TE Only TE & TM Material SiO 2 ( n\approx 1.45 ) Si ( n\approx 3.6 ) Wavelengths \lambda [\mu m] \{0.8,0.85,0.9,0.95,1.0\} \{1.4,1.45,1.5,1.55,1.6\} Heights h [\mu m] 0.75-1.25 0.1-0.6 Periodicity p [\mu m] 2.86 4.6 Table 1: Physical settings of the main datasets. The B2 and C2 datasets were generated using our method and are used for training. Both are designed to ensure the presence of scattering patterns with at least 5\times 5 propagating Bragg orders. Additional datasets and configurations are described in the appendix. The datasets consist of pairs of binary images that represent the topology of the metasurface and the transmitted and reflected scattering efficiencies T and R , which are represented as matrices. Each pixel in these matrices contains the total power projected into one of the Bragg diffraction orders, where the central pixel corresponds to the zero-order diffraction lobe. Binary Structures Randomized Generation. To construct a diverse dataset of binary metasurface structures, we developed a custom generation pipeline that