Physics-informed latent diffusion model for high-efficiency metasurface optimization
Yuheng Chen, Michael Bezick, Blake A. Wilson, Alexander V. Kildishev, Vladimir M. Shalaev, Alexandra E. Boltasseva · 2025
Integrating adjoint topology optimization with a physics-informed latent diffusion model enables the efficient inverse design of thermophotovoltaic (TPV) metasurfaces, achieving an exceptional efficiency of ~97%. This approach addresses the limitations of traditional trial-and-error design methods and overcomes the challenges of complex, multi-objective optimization in nanophotonics. The framework incorporates photonic insights directly into the denoising process by leveraging a physics-informed U-Net architecture within a latent diffusion model, ensuring high-fidelity, physically viable solutions. Our method delivers superior performance compared to benchmark conventional generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). This physics-informed machine learning framework paves the way for advancements in metasurface optimization, offering broad applicability in sustainable photonics and beyond.