Generative Inverse Design of Metamaterials Enhanced by Physics-Informed Neural Network

Metamaterial design traditionally depends on computationally expensive physics-based simulations, while deep learning approaches require extensive, high-quality training data. This work introduces a data-free deep learning framework that combines a Conditional Deep Convolutional Generative Adversarial Network (cDCGAN) with a Physics-Informed Neural Network (PINN) for inverse design of transmission-type metasurfaces. The cDCGAN generates flexible meta-atom shapes, while the PINN acts as a fast, physics-based simulator, enforcing Maxwell's equations and generating adaptive training data. Our approach enables diverse, pixel-level metasurface designs that align with target spectra, outperforming traditional methods in flexibility and data efficiency.