MetaDiff

This repository presents a novel approach for the inverse design of metasurfaces using diffusion models, enhanced with posterior sampling techniques to achieve results unattainable through standard ancestral sampling. The method addresses the limitations of traditional sampling by introducing three advanced posterior sampling strategies: Posterior Sampling (PS) PS Monte Carlo PS with Spherical Gaussian constraint These techniques significantly improve the quality and diversity of generated designs, enabling the exploration of complex and high-performance metasurface configurations. Additionally, the proposed framework supports scaling the inverse design process to larger metasurfaces than those encountered during training, making it highly adaptable for real-world applications. This work also introduces a consistency loss-based technique during training, which reduces the sensitivity of diffusion models to hyperparameter choices. By doing so, it enhances both the robustness and convergence of the training process.