Physics-conditioned deep generative model for high-efficiency metasurface thermal emitter design

Metal- and dielectric-based nanophotonics have long been seen as one promising route to significantly enhance the efficiency of photovoltaics (PV), thermophotovoltaic (TPV), and photocatalysis systems, contributing to greener energy sources. Conventional design methods for these nanophotonic systems utilize manual trial-and-error and intuition-based models and show limitations in providing the global solution to complex, multi-objective problems. Multiple constraints on their optical performance, materials, and scalability increase the size of the optimization space thus making the problem too recourse-heavy. Machine learning techniques, such as deep generative models—such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs)—have shown promise in addressing this challenge by enabling global optimization and generating highly efficient designs. However, the generation of high-fidelity designs with realistic constraints remains an important issue. In this work, we combine topology adjoint optimization with a physics-conditioned latent diffusion model to provide the robust inverse design of a refractory plasmonic emitting metasurface for TPV systems. By embedding the figure of merit prediction into a conditioning U-Net architecture, the model learns photonics information within the diffusion model denoising process and demonstrates substantial efficiency (~97%), exceeding the state-of-the-art generative model benchmarks. The proposed approach could enable a much broader scope of optimal designs and materials synthesis; it could be applied to other sustainable photonic applications, including energy-saving photocatalytic processes, more efficient coherent and incoherent light sources, and photonic accelerators for nascent information technologies.