GiBS: Generative Input-side Basis-driven Structures

TL;DR

GiBS introduces a Generative Input-side Basis-driven Structures framework that parameterizes metasurface geometry with a compact set of coefficients in smooth bases such as Fourier or Chebyshev, thereby compressing the design space and enabling efficient optimization of complex, nonlocal devices. The approach couples this low-dimensional geometry with autoencoder-based manifold learning to map structure to electromagnetic response, producing interpretable latent spaces and facilitating inverse design across wavelengths and material states. Four specialized autoencoders capture absorption and scattering responses in insulating and metallic PEDOT:PSS phases, trained on 201-point spectra to yield high-fidelity reconstructions within a two-dimensional latent space. Experimental validation on a large-area PEDOT:PSS metasurface (20 μm × 20 μm, 80–900 nm pillars) demonstrates broadband scattering from 500 to 1100 nm with strong agreement between measured and simulated spectra, confirming GiBS as a scalable, fabrication-aware platform for data-efficient inverse design of multifunctional metasurfaces. The framework bridges AI-guided representation learning with practical photonic architectures, offering a path toward active and multilayer metasurfaces through structured basis parameterization and manifold-enabled optimization.

Abstract

Designing large-scale metasurfaces with nonlocal optical effects remains challenging due to the immense dimensionality and fabrication constraints of conventional optimization methods. We introduce GiBS (Generative Input-side Basis-driven Structures), an inverse-design framework that represents the entire device using a compact set of coefficients from smooth parametric bases such as Fourier or Chebyshev functions. This formulation compresses the design space by more than an order of magnitude, enabling efficient optimization of complex, broadband, and aperiodic geometries. GiBS integrates this low-dimensional representation with an autoencoder-based manifold-learning workflow to map the relationship between geometry and optical response, facilitating rapid exploration, discovery of high-performance designs, and systematic analysis of fabrication sensitivity. The inherent smoothness of the basis functions ensures manufacturability while capturing the asymmetry required for nonlocal optical interactions. We experimentally validated the framework through the realization of a PEDOT:PSS broadband scattering metasurface, whose measured response closely matched full-wave simulations across 500-1100 nm. These results establish GiBS as a scalable, data-efficient, and fabrication-aware platform for the inverse design of multifunctional metasurfaces, bridging AI-guided representation learning with experimentally realizable photonic architectures.

Figures (6)

• Figure 1: Parametric basis modeling in the GiBS framework. (a) Representative values of the basis coefficients $A_k$ and frequency parameters $(\omega_x,\omega_y)$ defining the nanopillar distribution. (b) A single supercell geometry generated using these coefficients, showing a smooth and asymmetric variation of pillar radii. (c) Extension of the same pattern to a larger metasurface array, demonstrating how continuous control of the basis terms produces globally varying yet fabrication-friendly layouts.
• Figure 2: Workflow of the autoencoder-assisted design pipeline. (a) Design parameters (basis coefficients) define (b) a supercell geometry, which is simulated to yield (c) electromagnetic spectra, including both scattering and absorption cross-sections for different material states. (d) One of the trained autoencoders compresses these spectra into a low-dimensional latent space, and (e) reconstructs them with high accuracy, validating the fidelity of the learned manifold. Four separate autoencoders were trained in total, one for each combination of response type (absorption, scattering) and material phase (insulating, metallic). Four independent autoencoders are trained, one per (response, phase) pair.
• Figure 3: Latent-space analysis of the GiBS framework applied to PEDOT:PSS metasurfaces in the insulating phase. (a) Two-dimensional latent embedding of the scattering cross-sections $\sigma_{\text{sca}}$ showing the distribution of random (red), Fourier cosine (blue), and Chebyshev (green) basis-generated geometries. The Fourier and Chebyshev parameterizations span a broader and more continuous manifold compared to random designs. Insets $\alpha$ and $\beta$ mark representative 400 nm-thick PEDOT:PSS metasurfaces on SiO$_2$ substrates with supercell periods of 20 $\mu$m and 10.3 $\mu$m, respectively. (b) Optical response of the Fourier (cosine) basis design $\alpha$, showing uniform scattering and transmission balance across the visible and near-infrared spectrum. (c) Optical response of the Chebyshev-based design $\beta$, exhibiting enhanced scattering intensity in the visible range due to strong boundary modulation. These examples illustrate how different basis functions can generate distinct yet physically consistent electromagnetic responses within the unified GiBS design manifold.
• Figure 4: Active-optic design space exploration enabled by the GiBS framework. (a) Schematic illustrating the implementation of the Fourier basis parameterization for 400 nm-thick PEDOT:PSS metasurfaces on SiO$_2$, modeled in both insulating and metallic states. (b) Latent embeddings of absorption ($\sigma_{\text{abs}}$) and scattering ($\sigma_{\text{sca}}$) spectra for randomly generated geometries, showing limited and disjoint coverage of the response space in both material phases. (c) Corresponding embeddings for GiBS-generated (Fourier basis) designs, where blue points represent basis-driven geometries and red points denote random ones. The structured parameterization provided by GiBS significantly broadens and smooths the latent coverage, revealing continuous and connected manifolds across both insulating and metallic states.
• Figure 5: Scanning electron microscope (SEM) images of the fabricated PEDOT:PSS metasurface optimized via the GiBS framework. The device demonstrates large-area periodicity and excellent pattern transfer fidelity, closely matching the designed basis-driven geometry. The structure consists of nanopillars with diameters ranging from 80 nm to 900 nm within a 20 $\mu$m $\times$ 20 $\mu$m supercell. The coexistence of these sub- and superwavelength features supports both local and nonlocal resonances, enabling broadband scattering from 500 to 1100 nm. The insets highlight the smooth transitions and uniform feature definition achieved across the 400 nm-thick PEDOT:PSS film.