Inverse Design of Metasurfaces Using Physics-Informed Diffusion Models with Spectral Constraints

Abstract

The inverse design of metasurfaces constitutes a formidable challenge in computational electromagnetics and nanophotonics, primarily due to the non-uniqueness of the scattering problem and the high dimensionality of the design parameter space. Conventional optimization techniques, such as topology optimization and evolutionary algorithms, often succumb to high computational costs and convergence to local minima. Deep learning approaches, while promising in accelerating the design process, frequently struggle to strictly adhere to the governing Maxwell’s equations, leading to physically unrealizable or suboptimal structures. This paper introduces a novel framework: Physics-Informed Diffusion Models with Spectral Constraints (PIDM-SC). By integrating a pre-trained forward surrogate solver into the reverse diffusion process, we establish a generative mechanism that is explicitly guided by physical laws. The model is conditioned on desired spectral responses, ensuring that the generated meta-atoms not only exhibit high structural diversity but also strictly satisfy the target optical properties. Our approach utilizes a modified U-Net architecture capable of handling multi-modal data input, merging geometric features with spectral embeddings. Experimental validation on a dataset of silicon-on-insulator dielectric metasurfaces demonstrates that PIDM-SC outperforms state-of-the-art Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) in terms of spectral accuracy and fabrication feasibility. The results indicate a significant step forward in the reliable, data-driven design of complex nanophotonic devices.