Inverse Design of Tunable Infrared Metasurface Absorbers via a Conditional Wasserstein Generative Adversarial Network

TL;DR

The paper tackles the challenge of designing narrowband infrared metasurface absorbers by reframing inverse design as a probabilistic generation problem using a conditional Wasserstein GAN. It introduces a dual-channel image encoding for geometry and $Si_3N_4$ thickness, enabling multiple viable designs per target spectrum and robust performance under oblique illumination via transfer learning. Results show spectral peak errors below $5$ nm and MSE around $2.16\times 10^{-3}$ across 500 test cases, along with demonstrable one-to-many design generation and a hybrid plasmonic–dielectric absorption mechanism confirmed by FDTD simulations. The approach eliminates dependence on laborious parameter sweeps, offers a physics-informed design space, and provides a practical pathway for on-demand metasurface devices in sensing, spectroscopy, and imaging.

Abstract

Narrowband perfect absorbers are interesting for spectrum sensing, molecular detection, and infrared imaging. However, their design remains constrained by intuitive, iterative methods that lack flexibility, while also facing challenges in multi-objective optimization. Here, we introduce a deep learning-enabled inverse-design framework that overcomes these limitations through a conditional Wasserstein Generative Adversarial Network (WGAN). The main contribution of this work is a dual-channel image encoding scheme that jointly represents the geometry and thickness of a Si$_3$N$_4$ meta-layer, facilitating the network to learn the distribution of viable structures for a target optical response. This approach naturally solves the inherent ``one-to-many'' design issue, giving a diverse portfolio of functional candidates from a single input spectrum. The designed absorbers achieve exceptional spectral fidelity, with resonance peak errors below 5 nm, a mean squared error (MSE) on the order of $10^{-3}$, and the capacity to produce over 10 distinct, high-performance designs per target. Furthermore, we demonstrate the model's robustness under oblique illumination, showing that it can be efficiently fine-tuned to maintain spectral accuracy across incidence angles from $10^\circ$ to $40^\circ$ by transfer learning, thus extending its practical utility to non-normal operating conditions. Full-wave simulations confirm that the generated geometries support a hybrid plasmonic-dielectric resonance, leading to near-perfect absorption and strong near-field enhancement. Our study provides a robust, physics-aware design paradigm that moves beyond conventional parametric optimization. The introduced framework establishes a versatile platform for the on-demand inverse design of advanced photonic devices for sensing, spectroscopy, and optical signal processing.

Figures (8)

• Figure 1: Schematic diagram of the proposed structure (a) and sampling algorithms and encoding methods for geometric shapes (b).
• Figure 2: Architecture and training schematic of the conditional Wasserstein Generative Adversarial Network (WGAN). (a) The WGAN framework consists of a generator $G$ and a discriminator (critic) $D$. (b) The training algorithm illustrates the adversarial loop: the generator produces a candidate structure $G(z, y)$; the discriminator computes scores for both the generated output $D(G(z, y), y)$ and a real sample $D(x, y)$; gradients are calculated via backpropagation from the combined critic loss; and the weights of both networks are updated alternately to minimize the Wasserstein distance between the real and generated distributions. This process enables stable and diverse inverse design of metasurface absorbers.
• Figure 3: Inverse design results and spectral validation for target absorption peaks at (a) 1440, (b) 1480, (c) 1520, and (d) 1560 nm. For each target resonance (red dashed curve), two independently generated metasurface designs are shown alongside their corresponding numerically simulated absorption spectra (orange and green dash curves).
• Figure 4: Statistical distribution of spectral accuracy for the inverse-designed metasurface absorbers. Histogram of the mean squared error (MSE) between target and simulated absorption spectra across a test set of 500 independently generated designs.
• Figure 5: Exemplary "one-to-many" design capability for a 1440 nm target absorption peak. Ten distinct metasurface absorber designs generated by the WGAN for the same target spectrum.