Inverse Design using Physics-Informed Quantum GANs for Tailored Absorption in Dielectric Metasurfaces

TL;DR

This paper introduces a hybrid quantum-classical GAN combined with physics-informed neural networks for the inverse design of dielectric metasurfaces, enabling precise control of narrow-band absorption with fewer training samples and higher Q-factors.

Contribution

It presents a novel quantum-enhanced inverse design framework that integrates QGANs with PINNs, improving efficiency and physical accuracy in designing high-Q dielectric metasurfaces.

Findings

Faster convergence with 99.5% fewer samples using QGAN+PINN

Achieved Q-factors exceeding 10^5 from training data with Q~10^3

Generated highly asymmetric metasurfaces with superior spectral properties

Abstract

High Q-factor narrow-band absorption exhibits high spectral selectivity enabling high-sensitive photodetectors, sensors and thermal emitters. All-dielectric metasurfaces are widely regarded as excellent candidates for giving rise to such narrow-band absorption. However, designing metasurfaces with specific functionalities remains a challenging task both experimentally and computationally, which is why inverse design methods are increasingly being explored. Inverse design process is highly complex due to its non-unique solutions and the higher dimensionality of the design space, making it challenging to precisely control the resonance wavelength, linewidth, and absorption intensity. In this paper, we present a novel hybrid methodology that integrates generative adversarial networks (GANs) (both classical and quantum) with physics-informed neural networks (PINNs) for the inverse design of narrow-band absorbing metasurfaces. By introducing a Fano-shaped absorption spectrum equation into the PINN loss function, we enforce physical constraints on the resonance behavior, ensuring outputs that are both spectrally accurate and physically consistent. The study presents a comparison between a conventional GAN + PINN framework and a PINN augmented by a hybrid quantum-classical GAN (QGAN). The findings indicate that the integrated PINN + QGAN model achieves faster convergence, requires 99.5\% fewer training samples, and yields an order of magnitude lower MSE compared to conventional GANs. Remarkably, even though the training dataset only contains metasurfaces with Q-factors on the order of $10^3$, the model is able to generate highly asymmetric metasurface structures with Q-factors exceeding $10^5$. This study presents a novel framework that integrates quantum machine learning with physics-based modeling, providing a promising method for quantum-enhanced inverse design in nanophotonic systems.