Aug 2026· Nano letters (Print)· 0 citations· 44 references
PhysicsComputer Science
TL;DR
These results show that a shared sequence-based LLM interface can provide a practical route to cross-family metasurface design while reducing the need for task-specific surrogate architectures.
Abstract
Metasurface design increasingly requires fast models that can operate across structurally distinct device families rather than retraining a separate surrogate for every geometry class. Conventional neural network surrogates often depend on fixed-dimensional descriptors, family-specific output formats, and repeated architecture tuning, which limit their scalability across heterogeneous meta-atoms. Here, we present a unified large language model (LLM) workflow for multifamily metasurface modeling and inverse design. Geometries, design parameters, and optical response channels were converted into a shared instruction-following text format and used to fine-tune Gemma-2-9B across 8 metasurface families. Compared with single-family baselines, the joint model simultaneously predicted the optical responses of all metasurface families while reducing the MSE for each family by an average of 56.5%. The same representation was also used for the inverse design. These results show that a shared sequence-based LLM interface can provide a practical route to cross-family metasurface design while reducing the need for task-specific surrogate architectures.
Stacked intelligent metasurfaces (SIMs) enable near-field wavefront shaping via multiple programmable layers. However, widely used wave-domain models often neglect power scaling, mutual coupling, and geometric flexibility, while multiport-network formulations are physically consistent but computationally heavy and typically assume fixed layer spacing. This paper develops a hybrid near-field SIM framework that remains in the wave domain yet is anchored to a physically meaningful power scale and supports geometry-aware optimization. From a Rayleigh-Sommerfeld propagation model, we construct coupling-aware hop matrices and enforce hop-wise power scaling via a Friis-based anchoring rule. Meta-atom responses obey amplitude-phase coupling with a transmission-reflection trade-off, and the inter-layer distances are treated as continuous design variables under total-thickness and minimum-spacing constraints. The resulting transmissive-reflective cascade retains only the dominant single-bounce inter-layer reflections and admits efficient forward evaluation with stable gradients. An alternating optimization (AO) algorithm based geometry-aware architecture was proposed to optimize the metasurface coefficients and the spacings. Simulation results show that, under realistic near-field layer coupling and hardware losses, the achievable sum-rate is non-monotonic in both the number of layers and the total SIM thickness. Geometry-aware spacing significantly improves the conditioning of the effective downlink channel and consistently outperforms uniform spacing, providing a realistic and optimization-ready basis for SIM-assisted near-field communication system design.
Yi-Nuo Dong, S. X. Ng, M. El-Hajjar· IEEE Transactions on Communi...· 0 citations
Chromatic dispersion fundamentally limits metasurface performance by coupling beam steering to frequency, thereby constraining bandwidth and undermining practical deployment. Here, we break this limitation by introducing a fundamentally different design paradigm for metasurfaces that decouples wavefront control from frequency. Our approach is enabled by a physics-grounded parallel equivalent circuit model that directly governs dispersion at the meta-atom level, transforming achromatic metasurface synthesis from an empirical, geometry-driven process into a deterministic and scalable design problem. Using this framework, we generate a large-scale meta-atom library exceeding 100,000 candidates with systematically engineered dispersion responses. This large and diverse library allows the direct selection of meta-atoms with near-unity reflection amplitude and independently tailored phase slopes, enabling precise control of both phase and its frequency derivative. As a result, we realize metasurfaces that preserve prescribed beam deflection angles (0° and 25°) across a wide bandwidth with negligible angular dispersion. A fabricated prototype experimentally demonstrates achromatic reflection from 8.0 to 16.0 GHz, representing a broad bandwidth for dispersion-free beamforming metasurfaces. The measured results closely match full-wave simulations and theoretical predictions, validating the robustness and accuracy of the proposed framework. By establishing dispersion as a directly engineerable degree of freedom, this work redefines the design methodology of metasurfaces and unlocks a universal route toward broadband, multifunctional wavefront control, with far-reaching implications for next-generation satellite and mobile communication systems.
Xin-Bo Chen, Ming-Yong Zhuang, Si-Ning Li et al.· Journal of Physics D: Applie...· 0 citations
Accurate nano-photonics simulations of large scale devices like optical metasurfaces require high accuracy reduced models for the device constituents. We present an automated framework for the optimization of Global Polarizability Matrix (GPM) models, which represent a complex scatterer as a small set of non-local effective dipoles. Our goal is to find the most frugal model that reproduces a particle's scattering response within a user-defined accuracy. The method iteratively removes redundant dipoles while re-adapting the positions of the remaining ones via gradient based optimization, stopping at the smallest model that still meets the target. Automatic differentiation, combined with an untrained neural network that reparametrizes the dipole positions, helps to place the dipoles at physically intuitive locations. We demonstrate the versatility of this approach across diverse geometries, from two dimensional ridges over simple spheres to complex three-dimensional particles, achieving compression factors of typically two orders of magnitude compared to full-wave simulations, for target accuracies in the order of few percent. We finally demonstrate how accurate, frugal effective models enable large-scale meta-deflector optimization without periodic approximations. This robust recipe for constructing frugal effective models paves the way for the rapid simulation of large-scale photonic assemblies, required for example for metasurface design.
The emergence of deep neural networks (DNNs) has greatly alleviated the time-consuming and phase-discretization problems in conventional metasurface design processes. However, most DNN-assisted design methods are constrained by predefined target electromagnetic (EM) parameter formats, making retraining unavoidable when the design objective changes. To address this, we develop an objective-configurable inverse design framework paired with thermally tunable metasurfaces for multi-channel terahertz (THz) wavefront manipulation. The designed metasurface combines anisotropic structural responses with thermally tunable VO 2, providing four independent linear polarization (LP) channels through polarization and state multiplexing. The inverse design framework couples a residual convolutional neural network forward surrogate model with an estimation-of-distribution algorithm implemented via the cross-entropy method. By reusing the same surrogate model and reconfiguring only the design objective, different channel combinations and wavefront functions can be selectively activated, which allows on-demand multi-channel wavefront manipulation. As proof-of-concept demonstrations, four addressable LP channels are realized, and two additional circular polarization (CP) channels are further introduced through adaptive phase allocation. With all four LP channels activated, four-channel letter hologram multiplexing is achieved on a single metasurface, with an average imaging efficiency of 70.8%. After extension to CP channels, six-channel wavefront manipulation is achieved with inter-channel crosstalk below 30%. By integrating a thermally tunable metasurface with surrogate modeling and probabilistic optimization, this work establishes a robust and scalable paradigm for next-generation reconfigurable multi-functional THz photonic devices.
Jue Xin, Yan Wang, Shide Zhang et al.· Journal of Physics D: Applie...· 0 citations
Metasurfaces, two-dimensional arrangements of subwavelength nanopillars, provide local control over the phase of light, enabling flat optical functions beyond conventional refractive components. Their inverse design, finding the pillar distribution producing a target response, faces two obstacles: the immense dimensionality of the design space and the prohibitive cost of rigorous electromagnetic simulations, precluding exhaustive exploration at device scale. This thesis addresses the challenge in three stages. The first assesses the reliability of rigorous Maxwell solvers: comparing three implementations of Li's factorization rules for RCWA shows that spectral convergence does not guarantee physical fidelity, as these rules implicitly distort the simulated permittivity, most severely in the plasmonic regime. FDTD, immune to such artifacts, is retained as ground truth throughout. The second stage removes the computational bottleneck: a local phase-approximation model, then fully convolutional surrogates trained on FDTD simulations of large pillar metasurfaces, predict the near field almost instantaneously. Exploiting problem symmetries quadruples the training database, and the surrogates generalize to much larger apertures while remaining differentiable. The third stage benchmarks three strategies under a common FDTD protocol ($R^2$ between realized and target far fields): Gerchberg-Saxton retrieval and Local Model ($R^2\approx0.925$), surrogate-based and heuristic-initialized gradient descent ($R^2\approx0.975$), and a generative framework based on diffusion models and Schr\"odinger bridges. Hybrid posterior sampling and amplitude constraints restore scale-invariant fidelity on surfaces over 230 times larger than training, with diverse, fabrication-tolerant designs. Database enhancement lifts every approach to $R^2\approx0.97$.
Mathys Le Grand Institut des Nanotechnologies de Lyon, Stmicroelectronics· 0 citations
Full-wave electromagnetic simulations provide accurate field distributions for nanophotonic structures, but their high computational cost limits their direct use in large-scale inverse design. Here, we introduce a deep-learning-assisted inverse-design framework in which a neural network is trained with finite-difference time-domain (FDTD)-calculated field maps and used to estimate the transmitted field distribution of candidate structures during genetic optimization. This approach allows the genetic algorithm to evaluate a large number of structures while using FDTD-derived spatial field information for the design objective. As a model system, we apply this framework to a metasurface color router, where red, green, and blue light must be directed to prescribed sub-pixel regions at the photodiode plane. The router is implemented as a single-layer Si3N4 metasurface within a conventional 2 µm × 2 µm Bayer unit cell, discretized into a 16 × 16 grid with a 125 nm pitch for experimental validation. For this design, the corresponding simulations predict a total transmission efficiency of 95.4% and RGB routing efficiencies of 72.9%, 68.6%, and 45.1%, with crosstalk values of 37.1%, 53.4%, and 39.5%, respectively. We further show that the same trained model can be reused for different photodiode-aperture layouts without generating new FDTD training data or retraining the network. These results demonstrate a reusable inverse-design framework for optimization of complex nanophotonic devices.
Hyoseok Park, Samuel Kim, Duk-Yong Choi et al.· Reports on progress in physi...· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
