Intelligent on-demand inverse design of high-degree-of-freedom meta-atoms: multi-constraint-latent diffusion model and contrastive learning

The inverse design of metamaterials is a mapping process from low dimension (spectral response) to high dimension (structure). In this design process, a "one-to-many" mapping problem may arise (similar spectra may correspond to multiple structures). Traditional methods use generative models (GAN, VAE, etc.) to solve the "one-to-many" problem. However, these methods randomly generate multiple structural images, which may produce complex structures that are difficult to manufacture. To address the above problems, this paper proposes an intelligent on-demand design method for high-degree-of-freedom meta-atoms based on latent diffusion and contrastive learning models. By introducing multiple generation condition constraints in the design space encoded by the latent diffusion model, meta-atoms that are easy to manufacture can be designed on demand. Furthermore, this paper computes and sorts the similarity values between the meta-atom structure library and the target spectral response through a contrastive learning model and then selects the optimal meta-atom that matches the spectrum, overcoming the potential trade-off between optimal design and result diversity, enabling the intelligent design of meta-atoms. Experimental results showed that our method can generate optimal structures that meet conditional constraints, are easy to manufacture according to requirements, and are superior to the generation accuracy and quality of traditional models.