Enhanced posterior sampling via diffusion models for efficient metasurfaces inverse design

The inverse design of metasurfaces is made difficult by the highly nonlinear relationship between geometry and electromagnetic response, and conventional optimization is both computationally demanding and prone to local minima. We present a diffusion-based generative framework with two additions aimed at enforcing the electromagnetic specification. The first is a posterior sampling guidance constrained by a spherical Gaussian formulation, which, unlike unconstrained posterior sampling, does not lose accuracy as the array grows, and which allows models trained only on 23 × 23-pillar arrays to design metasurfaces of up to 98 × 98 pillars in a few minutes. The second is a consistency training loss that evaluates the electromagnetic Figure of Merit during training, which reduces the sensitivity of the model to its diffusion hyperparameters and thus the tuning effort required. All reported accuracies are verified by full-wave FDTD simulation of the generated designs. Across the evaluated range, the proposed method reaches a mean R 2 of 0.962 against 0.913 for a phase-retrieval and look-up-table baseline and 0.699 for surrogate-driven gradient descent.